Data Mining with Oracle 10g using Clustering and
Classification Algorithms
Nhamo Mdzingwa
Computer Science Department
Supervisor: John Ebden
7th November 2005
Submitted in partial fulfilment of the requirements of
the Bachelor of Science (Honours) degree of Rhodes University
Deciding on
which algorithm to use, in terms of which is the most effective and accurate
algorithm in data mining, has always been a challenge for most data miners.
This has been made even more complex by the increasing number of data mining
tools that are available.
The objective of
this research is fundamentally focused at investigating the effectiveness of
two Clustering algorithms available in Oracle10g for data mining. These are the
K-Means and the O-Cluster algorithms. I intend to provide data miners with
adequate information regarding to Oracle data mining algorithm accuracy and
effectiveness in building and applying models. Although, the evaluation of
unsupervised data mining algorithms is a generally difficult task as the goals
of an unsupervised data mining session are frequently not as clear as the goals
of supervised learning, I have adopted evaluation techniques proposed by
[Roiger et al, 2003]. The second objective is to gather information from the
dataset used in the evaluation. Information gathering involves finding
predictors of HIV AIDS prevention behaviour attributes. The data was obtained
from the Centre for AIDS Development, Research and Evaluation Institute for
Social and Economic Research, Rhodes University and is HIV AIDS related
The results
obtained are as follows; the first set is concerned with the evaluation of the K-Means
and O-Cluster algorithms. Here it was observed that the O-Cluster algorithm builds
more accurate models than the K-Means algorithm and also that the models by the
O-Cluster algorithm find more accurate clusters when applied to new data. From
these results it is evident that the choice of modelling algorithm has
significant difference in its efficiency and accuracy. The second set of
results involves information gathering from the dataset. Here the attributes HIV
Test and Know AIDS were identified as predictors of prevention
behaviour of condom use and abstinence. These were found by distinguishing the
clusters found in the dataset.
The success of this project can be attributed to the
help of a number of people. Firstly, I would like to thank my project
supervisor, John Ebden for guiding me through the development process, for his support, technical
advice during the course of the project and
offering much constructive criticism along the way.
I would also like to thank Kevin Kelly, from the Centre for AIDS
Development, Research and Evaluation Institute for Social and Economic Research,
Rhodes University, for
providing me with the dataset that made this project possible as well as more
interesting.
Many thanks go to
the Andrew Mellon Scholarship Foundation for providing me with the funds to
complete my Honours degree. I also acknowledge the financial and technical
support of this project by Telkom SA, Business Connexion, Comverse SA, and
Verso Technologies through the Telkom Centre of Excellence at Rhodes University.
Warm thanks also go to my family and friends who have
provided support during the challenges, and understanding of my many hours
spent at University.
Lastly, recognition should go to my fellow project
students. The people I have spent 90% of my life with this year, your help and
support has far outweighed the distractions.
Thanks guys!
TABLE OF CONTENTS
Abstract 2
Acknowledgements. 3
CHAPTER 1. 8
1 Background. 8
1.1 Introduction to Data Mining. 8
1.2 Supervised and Unsupervised Learning. 10
1.3 Chapter Summary. 11
CHAPTER 2. 13
2 Research Direction Indicators. 13
2.1 Choice of Mining Tool 13
2.1.1 Oracle Data Miner 15
2.2 Choice of Algorithms. 16
2.3 Chapter Summary. 17
CHAPTER 3. 18
3 Research Approach. 18
3.1 Mining Process. 18
3.2 Mining Tool Used. 19
3.3 Clustering Algorithms. 20
3.4 Algorithm Settings in Model Building. 21
3.4.1 Building Models by using the K-Means Algorithm.. 21
3.4.2 Building Models by using the O-Cluster Algorithm.. 22
3.5 Dataset Used. 23
3.6 Techniques Adapted. 24
3.6.1 Determining Model Accuracy. 25
3.6.2 Determining Algorithm Accuracy. 27
3.7 Chapter Summary. 29
CHAPTER 4. 30
4 Method Implementations. 30
4.1 Data Preparation. 30
4.2 Cluster Models. 33
4.2.1 Building of Models. 33
4.2.2 Interpretation of Initial Model Results. 34
4.3 Applying the Best Models. 37
4.4 Testing of Model Results. 39
4.4.1 A Brief Recap of Technique to Determine Cluster
Quality. 39
4.4.2 Building Classification Models. 40
4.4.3 Applying the Adaptive Bayes Network (ABN) Models. 41
4.4.4 Comparison of ClusterIDs. 42
4.5 Chapter Summary. 43
CHAPTER 5. 45
5 Interpreting Evaluation Results. 45
5.1 Comparing the 1st Ten Cluster Models. 45
5.2 Comparing the 2nd Ten Cluster Models. 46
5.3 Accuracy of Algorithms. 47
5.4 Chapter Summary. 48
CHAPTER 6. 49
6 The Gathering of Information. 49
6.1 Determining Predictors by Distinguishing Clusters. 50
6.2 Determining Predictors by using Association Rules. 61
6.3 Chapter Summary. 63
CHAPTER 7. 64
7 Summary and Conclusions. 64
7.1 Key Results of the Work. 64
7.1.1 Conclusions Regarding Algorithm and Model Accuracy. 64
7.1.2 Conclusions Regarding Information Gathered from
Dataset 65
7.2 Oracle Data Miner and Available Algorithms. 66
7.3 Conclusion. 67
REFERENCES: 68
Appendix A: 70
Installation Problems Encountered in ODM... 70
Appendix B: 73
Configuring ODM... 73
Appendix C: 80
ODM Tutorials. 80
Appendix D: 84
Datasets. 84
Spreadsheets. 84
List of
Figures
Figure 1:
Evaluation of data mining tools from META Group. 14
Figure 3:
Mining process adopted. 18
Figure 4:
K-Means algorithm settings. 22
Figure 5:
O-Cluster algorithm settings. 23
Figure 6:
Example of model Confidence and Support values. 25
Figure 7:
Example showing option Only show Rules for leaf clusters active. 26
Figure 8: A
clearer view of the ClusterID structure in ODM.. 26
Figure 9:
Data preparation. 31
Figure 10:
Sample of Mining Data Table Structure. 32
Figure 11:
Algorithms, model names and their settings with distinct number of clusters. 34
Figure12:
Average confidence values for the 1st 10 models (biased results) 35
Figure.13.
Computed confidence and support averages for the models built 36
Figure 14: Facility for selecting attributes. 38
Figure 15:
Adaptive Bayes Network default algorithm settings. 41
Figure 16:
Comparison of models with default and adjusted settings. 46
Figure 17:
Sample of output table ANALYSE_CLUSTER_1. 51
Figure 18:
Sample of output table ANALYSE_CLUSTER_2. 53
Figure 19:
Sample of output table ANALYSE_CLUSTER_3. 54
Figure 20:
Sample of output table ANALYSE_CLUSTER_4. 55
Figure 21:
Sample of output table ANALYSE_CLUSTER_5. 56
Figure 22:
Sample of output table ANALYSE_CLUSTER_6. 57
Figure 23:
Sample of output table ANALYSE_CLUSTER_7. 58
Figure 24:
Sample of output table ANALYSE_CLUSTER_8. 59
Figure 24:
Sample of output table ANALYSE_CLUSTER_9. 60
Figure 25:
Error obtained after installations. 70
Figure26:
Connection settings for odm_use. 77
Figure 27:
Select orc2 connection name. 78
Figure 28:
connecting to the ODM server. 78
Figure 29:
ODM server interface when connected. 79
Figure 30:
login screen to Oracle10 Enterprise Manager. 81
Figure 31:
Example of how to load data. 82
List of
Tables
Table 1: Summary of dataset and model naming_ 42
Table 2:
Database tables compared_ 42
Table 3:
Results from the comparison of cluster and classification ClusterIDs 43
Table 4:
Summary of clusters in APPLY_OC3_TSHATSHA_ 50
Problem Statement:
The main objective of this project is to evaluate two algorithms
available in Oracle 10g data mining with respect to algorithm accuracy and
effectiveness. This research project pays attention to the evaluation of
clustering algorithms, with the adoption and implementation of techniques
proposed by [Roiger et al, 2003]. However, due to the nature of the field of
data mining my second objective involves gathering information from the dataset
used for the evaluation of the algorithms. This involves finding predictors of
HIV AIDS prevention behaviour attributes. Describing and distinguishing the clusters
found was necessary in order to do achieve this. Further more, other types of mining algorithms such as association rules (the Apriori)
and classification algorithm (Adaptive Bayes Networks) were employed to facilitate
further analysis on the data.
1 Background
1.1
Introduction to Data Mining
Data mining is a
powerful and new technology that has been steered by the revolutionary progress
in digital data acquisition and storage which has resulted in the creation of
huge databases. In fact, the production and accumulation of these digital
databases is occurring at a faster rate than our ability to comprehend and use
them.
The idea behind
Data mining is to find patterns of information in these databases such that an
organisation such as a business can make use of. This new technology has been
defined in almost as many ways as there
are authors involved in the field. It has been defined by [Han et al,
2001] and [Mannila et al, 2001] as a process of extracting or mining knowledge
from large amounts of data, or simply knowledge discovery in databases.
However, because data mining sits as an intermediate between statistics,
computer science, artificial intelligence, machine learning, database
management and data visualization, only to mention a few of the fields, the definition
rapidly changes depending on the user’s perspective.
The main parts
of data mining are concerned with the analysis of data and the use of software
techniques for finding patterns and regularities in datasets. Fundamentally it
is the computer which is responsible for finding the patterns by identifying
the underlying rules and features in the data. However, the data mining
software used by the computer makes use of algorithms that facilitate sifting
through the dataset in turn building models. The choice of a particular
combination of techniques to apply in a particular situation depends on both
the nature of the task to be accomplished and the nature of the available data.
As a summary, data mining is a
technology that is ready for application in the business community because it
is supported by three technologies that are now sufficiently mature:
Ø
Massive data collection stored in databases
Ø
Powerful multiprocessor computers
Ø
Data mining algorithms
Techniques in data
mining can yield the benefits of automation on existing software and hardware
platforms, and can be implemented on new systems as existing platforms are
upgraded and new products developed [Mannila et al, 2001]. When data mining tools
are implemented on high performance parallel processing systems, they can
analyze massive databases in minutes. Faster processing means that users can
automatically experiment with more models to understand complex data. High
speed makes it practical for users to analyze huge quantities of data. Larger
databases, in turn, yield improved predictions [Han et al, 2001].
1.2 Supervised and
Unsupervised Learning
There is a wide
range of sources available on data mining and most of these have various ways of
implementing data mining processes. Most authors have found it more convenient
to categorise data mining corresponding to different objectives to make it
easier for the person analysing the data. Due to this the learning from data
has been split into mainly two flavours: supervised learning and unsupervised
learning.
According to [Roiger et al, 2003], data mining is classified into
supervised and unsupervised concept learning methods. Supervised learning
builds classification models by forming concept definitions from sets of data
containing predefined classes while unsupervised clustering builds models from
data without the aid of predefined classes were data instances are grouped
together based on a similarity scheme defined by the clustering system.
The supervised learning builds models by using input attributes to
predict output attribute values while in unsupervised learning no target
attributes are produced but rather gives a descriptive relationship by using an
objective function to extract clusters in the input data or particular features
which are useful for describing the data. The number of
variables in unsupervised learning is usually much higher than in supervised
learning. The properties of interest are usually more involved than simple locations
(e.g. means, medians) and dispersions (e.g. standard deviations and median
absolute deviations) [Roiger et al, 2003]. Typical examples of unsupervised learning tasks are association
rules, cluster analysis and principal component analysis, independent component
analysis and multidimensional scaling.
[Berry et al,
2000] also categorises data mining into directed data mining and undirected
data mining as the two main styles of data mining. According to the authors, the
goal in directed data mining is to use the available data set to build a model
that describes one particular variable of interest in terms of the rest of the
available data. The authors also point out that directed data mining often
takes the form of predictive modelling, where one knows what he wants to
predict. Classification, prediction and estimation are the techniques used in
directed data mining. In undirected data mining, no variable is singled out as
the target. The goal is to establish some relationship among all the variables.
From the above,
it is clear that the classified categories described by these authors all
involve similar techniques. We can therefore say that directed data mining,
supervised learning and predictive modelling of data mining describe similar
techniques that can be collectively referred to as supervised learning.
Unsupervised learning, undirected data mining and descriptive modelling are
techniques in the same category and can therefore be referred to as
unsupervised learning.
However, the
problem in data mining at the moment is that, although data mining is
categorised, there are now too many data mining algorithms available. This
makes it difficult for miners to know which of the categorised algorithms
builds the most accurate model and whether the model built finds accurate
patterns when applied to new data.
1.3 Chapter Summary
This chapter
provides a brief introduction to the field of data mining. The field is steered
by the ever increasing size of digital data and storage in databases. Data
mining is classified into mainly 2 categories, supervised and unsupervised
learning. However little knowledge on algorithm accuracy has lead to miner
having difficulties when selecting algorithms to mine for data.
The structure of the rest of the paper is as follows: Chapter 2 highlights
on research direction indicators elaborating on various work related to the
evaluation of data mining algorithms relating to algorithm accuracy and
effectiveness. Chapter 3 details the methods adopted for the investigation
detailing on the mining tool used, the algorithms that will be investigated,
the dataset and techniques adapted. Chapter 4 is a brief breakdown of the implementation
process of the methods suggested while chapter 5 gives an interpretation of
results achieved. Chapter 6 details the gathering of information from the
dataset by finding predictors of HIV AIDS prevention behaviour attributes. The
document concludes with chapter 6 which gives a summary of all findings.
2 Research
Direction Indicators
This chapter
provides a brief overview of research direction indicators. Basically, it
explains why Oracle Data Miner (ODM) was chosen for use from a variety of data
mining tools available in the market in order to perform the evaluation of the
algorithms. It also explains why clustering algorithms were chosen for
evaluation rather than evaluating other algorithms from other categories such
as classification or even regression.
There is currently a variety of data mining tools available in the
market. Some of these tools are from big software house companies such as IBM,
Microsoft, and Oracle. The various tools employ various mining techniques as
well as algorithms depending on the functionalities available in these tools.
Currently the most popular tools available in the market include the following
(only to mention a few):
-
Oracle Data Miner (ODM) from Oracle
- Gnome
Data Mine Tools provided as free open source software
- Insightful
Miner
- Clementine
- Data
mind
- Enterprise miner
- Intelligent
miner for data
For this
research project, Oracle Data Miner (ODM) version 10.1 from Oracle was selected
for use to provide a means for the evaluation of clustering algorithms and also
in gathering information from the datasets used by describing the clusters
found.
The use of ODM mining tool was
mainly motivated by research findings from the META Group, a
leading provider of Information Technology research, advisory services and
strategic consulting, which published its METAspectrum report for Data Mining.
The report ranked Oracle Data Mining a leader amongst
other data mining tools [Berger, 2004]. The data mining tool rankings are
clearly displayed in Figure 1.
Figure
1: Evaluation of data mining tools from META Group
From this figure we can clearly see that there is tough competition
amongst the mining tools in the market. This is mainly influenced by the
functionalities that the mining tools possess. These may include the ease of
using the tool (whether it provides wizards to help user follow the mining
process) and how the tool accesses data from the database.
Oracle10g Data
Miner (ODM) is the mining tool that I selected for use in the evaluation of the
algorithms. ODM version 10.1 is a data mining software embedded in the Oracle
10g Database Enterprise Edition (EE) that enables one to discover new insights
hidden in data [Berger, 2004]. The Oracle Data Mining suite is made up of two
components, the data mining Java API and the Data Mining Server (DMS). The DMS
is a server-side, in-database component that performs data mining that is
easily available and scalable. The DMS also provides a repository of metadata
of the input and result objects of data mining. The Oracle Data Miner is a
component that resides within the DMS and is responsible for the actual data
mining.
All data mining
components have been added to the Oracle Data Miner, this also includes the ODM
Browser interface which provides a full set of mining wizards. A data analyst
can perform data mining tasks without the need to generate code or to use
Oracle Java Developer as it was the case with older versions of Oracle data
mining [Oracle 2005].
[Berger, 2004] states
that Oracle Data Miner supports supervised learning techniques (classification,
regression, and prediction problems), unsupervised learning techniques
(clustering, associations, and feature selection problems), attribute
importance techniques (for finding key variables), text mining, and has a
special algorithm for life sciences sequence searching and alignment problems.
The availability of these algorithms provides all the necessary tools required
in gathering information from the dataset.
The main
advantage about Oracle Data Miner is that all data mining processing occurs
within the Oracle database. I took this into consideration when selecting the
mining tool to use. Other mining tools force you to extract the data out of the
database before the actual mining process which may result in a number of flaws
in these mining tools. The aspect of ODM were everything occurs in the
database results in a more secure and stable data management which enhances
productivity as the data does not have to be extracted from the database before
mining it.
Another fact I
took into account about this mining tool is that since the ODM model building
and model scoring functions are accessible through the Java API and the Oracle
Data Miner’s graphical user interface (GUI), the combination of the two enables
Oracle to provide an infrastructure for one to integrate data mining seamlessly
with database applications.
All the
functionalities and advantages mentioned here have resulted in Oracle Data Miner
being the most preferred mining tool to use when performing the algorithm
evaluation and gathering of information from the dataset.
2.2 Choice of Algorithms
The two
algorithms I selected for the evaluation are the K-Means and O-Cluster. Both
algorithms are clustering algorithms available in ODM. The decision to evaluate
these algorithms is influenced by a number of factors which I will discuss in
detail.
Firstly, Oracle
data mining supports classification, clustering, association and regression
algorithms and from these only the classification algorithms available in Oracle
Data Miner. These include the Adaptive Bayes Network and the Naïve Bayes which
have been evaluated by [Davis, 2004]. This leaves other algorithm classes
pending investigation. I therefore found it valuable and thus necessary to
investigate the clustering algorithms in ODM.
[Davis, 2004] describes an investigation of the commercial data mining suite, which is available
with Oracle9i database software which primarily involves determining classification
algorithm effectiveness and efficiency. As a brief conclusion drawn by the
author regarding the investigation, Oracle Data Mining provides all the
functionality necessary to easily build an effective data mining model.
Furthermore she concludes that the Adaptive Bayes Network algorithm produced
the most effective data mining model as well as producing the most accurate
results when the model is applied to new data.
The work by [Davis, 2004] plays a significant role in the choice of algorithms to investigate. Here I am
left with the choice of investigating the Clustering algorithms. However this
will provide Oracle data miners with a clear idea concerning the effectiveness
of algorithms available in Oracle Data Miner.
It is also worth
mentioning that the results by this author are valuable as they give me an
indication of which algorithm to use when employing a supervised learning
algorithm (classification) when determining cluster quality. This will be
discussed in the following chapters.
This chapter has
given a brief highlight on some research projects that are closely related to
the evaluation of data mining algorithms and how these evaluations have benefited
and influenced my research project. The reasons for selecting Oracle Data
Mining for this research have been discussed as well as why clustering
algorithms will be investigated.
The next chapter will describe the
methods adopted for this investigation, as well as how the techniques are to be
implemented.
3 Research
Approach
This chapter provides
a detailed description of the approach that I took in order to perform an
effective evaluation. It explains the methods adapted for the evaluation of
both the algorithms and in determining the quality of clusters found.
In the study of data
mining literature, many data mining authors felt it is of high importance that data
mining should be done in a procedural manner. Many of these authors presented
their own mining processes which they felt will develop non-experts to conduct
data mining. With the knowledge attained from the various authors, I adapted
their mining processes to a process that I then adopted. The resultant process
is depicted in Figure.3 in a nutshell.
Figure 3: Mining process
adopted
Although
a more detailed discussion on the mining process is going to be given later in
the chapter, Figure 3 simply shows that by using the Oracle 10g database, I
select the dataset that I wish to mine. However at this stage proper identification of goals to be accomplished by the data mining
project will help with domain understanding and determining what is to be
accomplished. At this point it is necessary to mention that my research project
goals involve an effective evaluation of clustering algorithms, for this it was
important to find and use understandable data which in turn will be relatively
easier to interpret after the mining event.
The next step involves data preprocessing or cleansing of the
dataset which primarily deals with noisy data. This
stage involves locating duplicate records in the data set, locating incorrect
attributes, smoothing the data and dealing with outliers in the data set. It
includes data transformation which involves the addition or removal of
attributes and instances, normalizing of data and type conversions.
The actual data mining stage then follows;
here the models are built by the selected algorithm that sifts through the data
set. The resulting model is then applied to new data which is interpreted to
determine if the results it presents are useful or interesting. The acquired
knowledge is then applied to the problem.
Although not depicted in Figure 3, some algorithm evaluation
techniques were adopted from [Roiger et al, 2003] in
building models and determining cluster quality, and these will be explained
later in the chapter.
For the purpose
of the algorithm evaluation Oracle data miner version 10.1 will be used mainly
due to the reasons discussed in chapter 2. Oracle10g database version
10.1.0.2.0 was installed and configured. The data mining tools and software
(Oracle data miner 10g version 10.1.0.2.0) was also installed and configured
for use with the database.
The Oracle Data Miner 10g is a user
interface to Oracle Data Mining (ODM 10.1.) which replaces the Data mining for
Java (DM4J) an ODM Component and ODM Browser user interface that was initially
used in the older versions of ODM, for instance, Oracle9i data mining. In
addition to Oracle Data Miner, there is the ODM Java Code Generator, which is
an Oracle JDeveloper extension [Oracle, 2005].
Although the
Oracle data mining tool was installed successfully, there were a number of
installation and configuration flaws encountered. These are discussed in
Appendix A in detail resulting in a working configuration manual (Appendix B).
I also provided an introductory manual before the user can engage to the data
mining tutorial for this particular mining tool (Oracle data miner) which is
Appendix C. Appendix C is very useful because if the user is new to the tool,
he normally finds himself stuck not knowing what to do next.
Clustering is used to identify distinct segments of a
population and to explain the common characteristics of members of a cluster,
and to also determine what distinguishes members of one cluster from members of
another cluster. Clustering algorithms make use of the Euclidean
distance formula to determine the location of data instances and their position
in clusters and so requires numerical values that have been properly scaled
[Han et al, 2001].
Oracle Data Miner supports two clustering algorithms and these I selected
for investigation in this research project. The reasons why I selected these algorithms
for investigation have been discussed in the preceding chapter. The algorithms
are, namely:
Ø The
Enhanced version of K-Means and,
Ø Proprietary
Orthogonal Partitioning Clustering (O-Cluster)
algorithm
Both the Enhanced k-Means (EKM) and Proprietary O-Cluster algorithm
support identifying naturally occurring groupings within the data population
[Berger, 2004].
The Enhanced version of K-Means algorithm supports hierarchical clusters,
handles numeric attributes and will cut the population into the user specified
number of clusters. The proprietary O-cluster algorithm handles both numeric
and categorical attributes and will automatically select the best cluster
definitions [Berger, 2004]. Both algorithms divide the data set into k number
of clusters according to the location of all members of a particular cluster in
the data. When choosing the number of clusters to create, it is possible to
choose a number that doesn’t match the natural structure of the data which
leads to poor results. For this reason [Berry et al, 2000] suggests that it is
often necessary to experiment with the number of clusters to be used.
Each of the two
Clustering algorithms in ODM has a setting that is used to tune the algorithm
when building models. Both algorithms also have a parameter; the maximum number
of clusters (k), this is available so that the user can pre-define the number
of clusters that he wishes to find from the dataset.
3.4.1 Building Models by using the K-Means Algorithm
There are two
settings for this algorithm when building models, Minimum Error Tolerance and
Maximum Iterations. These determine how the parent-child hierarchy of clusters
is formed and can be modified experimentally to observe the changes in cluster
definitions. Increasing the tolerance or lowering the iteration maximum will
cause the model to be built faster, but possibly with more poorly-defined clusters
[ODM tutorial, 2004].
Figure 4: K-Means
algorithm settings
3.4.2 Building Models by using the O-Cluster Algorithm
O-Cluster finds
natural clusters up to the maximum number entered as a parameter. That is, the
algorithm is not forced into defining a user-specified number of clusters, so
the cluster membership is more clearly defined. O-cluster has only one setting
the Sensitivity; it determines how sensitive the algorithm is to differences in
the characteristics of the population. Thus, a higher sensitivity value usually
leads to a higher number of clusters [ODM tutorial, 2004].
Figure 5: O-Cluster algorithm
settings
It must be noted that the main purpose of these algorithm settings is to fine
tune the algorithm so as to achieve finer and more accurate clusters. These
settings are also useful in this evaluation as they will be used to distinguish
the models built as well as to see if the accuracy of an algorithm is affected
by the change in these settings. The algorithm settings will be based on trial
and error followed by a critical analysis.
The dataset used
in the evaluation for this research was obtained from the Centre for AIDS
Development, Research and Evaluation Institute for Social and Economic
Research, Rhodes University. The data is a result of a
questionnaire survey relating to HIV AIDS as well as a South African television
drama, Tsha Tsha, which is a HIV AIDS awareness program. A copy of the complete
questionnaire can be found in the Appendix D on the CD-ROM
that accompanies this project.
The survey was conducted
by the above research institute around three provinces in South Africa namely, Gautang, Kwazulu Natal and the Eastern Cape. The research institutes’ goal is
to gather information from the three provinces as a sample of the South African
population by using a questionnaire survey, then use statistical tools to
determine predictors of HIV AIDS prevention behaviour.
[Berger, 2004] states
that in the data mining process the problem definition is the most
important step. Basically, this is where the domain expert decides the
specifics of translating an abstract business objective. In my case, the problem
definition is to use the Oracle data mining tool to determine the predictors of
HIV AIDS prevention behaviour instead of using statistical tools.
According to the
researchers at the Centre for AIDS Development, Research and Evaluation
Institute for Social and Economic Research, three rounds of the survey were
conducted at different times with the same questionnaire resulting in three
sets of data. The final datasets were then compiled by Kevin Kelly, from the
above research institution, and loaded into 3 separate excel files namely,
Tsha_tsha_round1, Tsha_tsha_round2 and Tsha_tsha_round3. Each of the three
datasets consists of 131 attributes and more that 800 rows (cases); this is
large enough for the evaluation. These dataset files are also available on the
Compact Disc that accompanies this project.
A couple of
techniques were adopted so as to carryout the evaluation. These help in
determining the accuracy of models built and determining the quality of
clusters found after applying the models to new data. The techniques are
detailed below;
In order to be
able to carry out an effective evaluation of the algorithms, it is valuable to
use the best models built by either algorithm. Ideally, the best models should
possess the highest probability of finding accurate clusters and must have
minimum error .To determine the accuracy of each model built I make use of two
parameters that the mining tool provides as output after the building of the
model. These parameters are the Confidence and Support. A sample of how these
parameters are displayed by the mining tool is shown in Figure 6.
Figure 6: Example of model Confidence and Support values
The Confidence
is a measure of the homogeneity of the cluster; that is, how close together are
the cluster members [ODM Tutorial, 2004]. I therefore made it a measure of
accuracy such that a cluster with the highest confidence value is more accurate
and effective than that with a lower value. Thus, computing an average
confidence value of all the clusters in a model would determine the accuracy of
a model.
The support is a
measure of the relative size of a cluster (the total need not be 1.00), such
that the higher the value the larger the cluster [ODM Tutorial, 2004]. In this
paper it is used as an alterative measure of accuracy to the confidence.
The ClusterID (see
Figure 6 and Figure 8) is a value that differentiates the clusters found. The
order of numbering used for the ClusterID is as follows; the mining tool
generally looks for all clusters in the dataset depending on the algorithm
settings. The maximum number of clusters (k) that one sets during model
building determines the number of clusters that the mining tool displays as the
leaf clusters. When the option Only show Rules for leaf clusters is active
(see Figure 7), the maximum number of clusters that you selected during model
building is displayed with each cluster having a distinct ClusterID. In this
case the maximum number of clusters (k) was set to 4.
Figure
7: Example showing option Only show
Rules for leaf clusters active
ClusterID = 1
will always represent the entire dataset, while ClusterID 2 and 3 are two
separate clusters in ClusterID 1 (there can be more), this continues until the
maximum number of clusters that you will have set is reached (see Figure 8).
Figure
8: A clearer view of the ClusterID structure in ODM
The evaluation
of unsupervised data mining algorithms is a generally difficult task, since the
goals of an unsupervised data mining session are frequently not as clear as the
goals of supervised learning. However, [Roiger et al, 2003] suggests and
presents a number of techniques for the evaluation of unsupervised algorithms and
models. To increase the robustness of this evaluation, I plan to adapt and
implement the evaluation techniques by these authors.
The techniques by
[Roiger et al, 2003] are explained fully on pages 58 and 232. Here the four main methods that I
adopt are as follows:
1)
Employing supervised learning to evaluate unsupervised learning.
2)
Apply alternative technique’s measure of cluster quality.
3)
Create own measure of cluster quality (in this case making use of
Confidence).
4)
Perform a between cluster attribute value comparison.
Although all of
the above techniques will be employed at some stage of the evaluation, it is
important to mention that the technique of employing supervised learning to
evaluate the unsupervised learning is the most crucial technique. I will
use it to determine cluster quality as I will evaluate cluster accuracy.
[Roiger et al, 2003] details the technique as follows:
i.
Perform an unsupervised clustering. This step involves the building of
models from Cluster algorithms then apply the best model (most accurate) to new
dataset. Then designate each cluster found in the dataset as a class and assign
each cluster an attribute name. For example, if the clustering technique
outputs three clusters, the clusters could be given the class names C1, C2 and
C3. The Oracle data mining tool when applying the cluster models to new data
includes another attribute the ClusterID, this attribute signifies a cluster
digit that an instance belongs to.
ii.
Choose a random sample of instances from each of the classes formed as a
result of the instance clustering. Each class should be represented in the
random sample in the same ratio as it is represented in the entire dataset. The
percentage of the total instances to sample can vary, but a good initial choice
is two-thirds of all instances.
iii.
Build a supervised learner model with class name as the output attribute
using the randomly sampled instances as training data. Employ the remaining
instances to test the supervised model for classification correctness.
Using this
technique I will make use of a classification algorithm (a supervised learning
algorithm). I plan to use the Adaptive Bayes Networks (ABN) algorithm for
building the classification models. Making use of the ABN is motivated by the
results obtained by [Davis, 2004] which concluded that the algorithm is more
accurate in predicting attributes for the classification algorithms in Oracle.
As a brief
summary of the technique by [Roiger et al, 2003] but applying it to my
situation is as follows: I will use Oracle data miner as the mining tool. The
evaluation technique will involve taking the resultant table (for instance
table APPLY_RESULTS) obtained after applying a cluster model, then pick a
random sample of instances (roughly two thirds) from each cluster found from
this table and place them in another table that will be used to build a
classification model using the Adaptive Bayes Network (ABN) algorithm. The
resultant model is then applied to the remaining instances from the table APPLY_RESULTS.
The attribute being predicted by the classification model (ABN model) is the ClusterID.
The results after applying the ABN model which predicts the ClusterIDs are compared
to the ClusterIDs in the table APPLY RESULTS. The ClusterIDs for a particular
instance is uniquely identified by the instance’s REFNUM (reference number). The
ClusterIDs that occur in both tables are counted and a percentage is found.
This percentage is used as a measure of cluster accuracy.
This chapter has
discussed in detail the approach that will be taken in evaluating the two
clustering algorithms available in Oracle data mining 10.1. The next chapter is
solely dedicated to the implementation of the methods presented in this
chapter.
4 Method
Implementations
This chapter
provides an insight on how Oracle Data Miner provides data mining functionality
as I implement the methods discussed in the preceding chapter. It describes the
pre-processing of the data as well as the actual model building. The best
models by each algorithm will be identified and applied to new data. The actual
evaluation is performed at this stage and the process is explained in this
chapter.
For
this evaluation, the datasets Tsha_tsha_round1 and Tsha_tsha_round2 were
used. Tsha_tsha_round1 was used to perform the main evaluation while
Tsha_tsha_round2 was used to repeat the whole evaluation process implemented on
Tsha_tsha_round1 mainly for verification and confirmation of the results.
Before
the datasets were loaded into the database tables initial, pre-processing of
the datasets was necessary. This stage involves
locating duplicate records in the data set, locating incorrect attributes,
smoothing the data, dealing with null values, outliers, inconsistencies and
table properties. It also involves the removal of rows with no values at all.
Another good factor about Oracle Data Miner (ODM) is that it also provides an option that helps in the data
preparation step. Here on can specify automatic binning for the input data,
were the format of the data is automatically converted to one that is understood
by the mining tool.
Figure
9: Data preparation
Since the major evaluation process was to be performed on the data
Tsha_tsha_round1, the dataset was then partitioned into three datasets (with
a random selection of instances). I then created three identical database
tables, one that will be used to build models, the other to apply the built
models and the third for testing the accuracy of models. The tables are
TSHA_TSHA_BUILD1, TSHA_TSHA_APPLY1 and TSHA_TSHA_APPLY2 with each table having
131 attributes. A sample of the resulting structure of the database tables is
depicted in Figure 10.
Figure 10: Sample of Mining Data
Table Structure
The technique
and tool for loading data into database tables is explained in Appendix C.
The technique involves the use of an application called sql*loader. Using this application
I loaded the three datasets into the tables that I created.
The building of
cluster models was done in two stages. The first stage involved building models
with random algorithm settings as well as distinct cluster size such that each model
has distinct parameters. This is to determine if cluster size has any
significant effect on algorithm accuracy. The building of models in the second
stage was dependent on model accuracy results obtained from the first stage;
the second stage of model building involved setting the maximum number of
clusters (k) for both algorithms to the same value for all models built. This
basically helps in evaluating the resultant model accuracy.
4.2.1 Building of Models
In the first stage
ten models were built in total, with five from each algorithm. Since the model
settings are based on trial and error, it was necessary to build a large number
of models to provide a wide range of models to select the best from. This also
helps when analysing model accuracy to determine if the algorithm settings do affect
the algorithm’s performance.
The ten models
all have distinct values of the maximum number of clusters (k). Here, each
model built from the K-Means algorithm was named in the form BUILD1_KM_TSHATSHA
while models built by the O-cluster algorithm were named in the form
BUILD1_OC_TSHATSHA. The digit 1 in these model names denotes the first model, digit
2 second model, and so on. Figure 11 shows the models and their settings in
detail,
Figure 11: Algorithms, model names and their settings with
distinct number of clusters
After building
the ten models from the algorithm settings in Figure 11, it was observed that
each model did actually discover clusters. This is evident from the ClusterIDs,
Confidence and Support values that the mining tool depicts on displaying the model
results. A sample of the model output was shown in both Figure 6 and 7. For all
the models that I built in this stage, I computed an average confidence value
and average support which both determine model accuracy; these are depicted in
Figure 12.
On analysing these
average confidence values computed from the model clusters, I observed a high
degree of bias in the results. Here the bias is mainly due to the variation in
the value of the maximum number of clusters (k) that was set for each algorithm
during model building as shown in Figure 11 (algorithm settings). This makes it
difficult to determine the best model built from the two algorithms.
Figure12: Average confidence values for the 1st
10 models (biased results)
On analysing Figure
12, I observed a general trend in the average confidence values. The trend
shows that increasing the maximum number of clusters for each algorithm results
in a higher confidence value. Hence increasing the value of k has a significant
effect on the accuracy of the algorithm in model building.
However,
observations show that these model results are biased therefore are not
suitable for this evaluation. The biasness comes from varying the value of k when
building models. When I look at the models built using the O-Cluster algorithm
in Figure 12, I observe that increasing k results in more accurate models as
the average confidence values increases. This is also the same case with the
K-Means algorithm but it is apparent that the O-cluster models are more
accurate. Thus it would be false to conclude from these biased results that the
O-Cluster algorithm builds better models.
To overcome the
problem of bias, I then decided to set the value for the maximum number of clusters
(k) to a fixed value. I set the value k to 7 for both algorithms because, the
default number of clusters for the K-Means algorithm in ODM is 4 and
that for O-Cluster is 10, therefore setting the value of k for the two
algorithms to an average of the two default values was reasonable.
I then re-built
10 more models with the same settings as in Figure 10 but with the maximum
number of clusters (k) fixed at 7 for all models. The new model names are in
the form BUILD1_OC_TSHATSHA2 for O-cluster and BUILD1_KM_TSHATSHA2 for the K-means,
with the digit 2 at the end indicating the second set of models built. These
models built at this stage are unbiased. I then made a new computation of
average confidence and average support figures as shown in Figure 13.
Figure.13. Computed confidence and support averages for
the models built
From Figure 13,
it is evident by analysis that the model BUILD3_OC_TSHATSHA2 from the O-Cluster
and the model BUILD5_KM_TSHATSHA2 from the K-Means posses the highest values
for both the average confidence and support values.
Critical
observations of these results shows that changing the settings for the K-Means
algorithm has little effect on the clusters found. This results in very small
differences in the computed average values for models built by this algorithm.
On the other hand, the
O-cluster
algorithm models were affected greatly by the changes in the settings where the
average confidence increases greatly with k.
In conclusion it
is evident from Figure 13 that the O-cluster algorithm builds more accurate
models than the K-Means algorithm.
The two models
identified in the preceding section as the best and most accurate were
BUILD3_OC_TSHATSHA2 from the O-Cluster algorithm and BUILD5_KM_TSHATSHA2 from
the K-Means. These two models were applied to the new data TSHA_TSHA_APPLY1.
At this stage the mining tool allows you to select attributes that
will be displayed in the output table after applying the models. This facility
becomes valuable in this case since the dataset that we are using has a large
number of attributes (131). Since the primary goal with this dataset is to
determine predictors of HIV AIDS prevention, I then selected as many potential
predictors. This filters out those attributes that I felt had little
significance to HIV AIDS. Although, this may result in the loss of information
it is necessary as it makes the analysis and interpretation of the clusters
simpler.
Figure 14: Facility for selecting
attributes
The mining tool also provides a wizard which supplies a
name for the output table on applying the models; this name may be modified at
this stage. On applying the models clusters were discovered in the new data. These
cluster results were loaded into the tables APPLY_OC3_TSHATSHA for the
O-cluster and APPLY_KM5_TSHATSHA for the K-Means. The results were then exported
to spreadsheets to allow further inspection for algorithm evaluation. This will
be discussed in the following sections. The interpretation of the clusters will
also be discussed in the next chapter after I have given reasons and proved why
I selected a particular algorithm as one that provides more accurate results
than the other.
This section
primarily deals with determining cluster quality from the cluster results
obtained after applying the cluster algorithm models. Ideally, this involves
finding out which algorithm model finds more accurate clusters. Although,
[Roiger et al, 2003] indicates that the evaluation of clustering algorithms is
difficult, I intend to use the technique that these authors proposed as
discussed in Chapter 3, section 3.5.2
The technique is
by [Roiger et al, 2003] and it uses supervised learning evaluation to evaluate
unsupervised clustering. Here I make use of a classification algorithm
(supervised learning), the Adaptive Bayes Networks (ABN) algorithm with the technique.
Making use of the ABN was motivated by the results obtained by [Davis, 2004] which concluded that the algorithm is more accurate in predicting attributes
for the classification algorithms in Oracle Data Miner.
Basically, the
evaluation technique involves taking the resultant table obtained after
applying a cluster model (APPLY RESULTS), pick a random sample of instances
(roughly two thirds) from each cluster found, place them in a new database
table that will be used to build a classification model, in this case using
ABN. The attribute being predicted is identified; in this case it will be the ClusterID.
The resultant model is then applied to the remaining instances from the APPLY
RESULTS (the one third of instances) with the ClusterIDs removed (stored in
excel for comparison later). The results after applying the ABN model which
predict the ClusterIDs are then compared to the initial APPLY RESULTS ClusterIDs
(the cluster ids in the excel file).
Two database
tables, build1_abn_FROM_OC and build1_abn_FROM_KM were created
and these are used for building the Classification models with the ABN
algorithm. The table build1_abn_FROM_OC was loaded with two thirds of instances
from the O-cluster model results table (APPLY_OC3_TSHATSHA) created as was
explained in section 4.3. Two thirds of instances from the table APPLY_KM5_TSHATSHA
(section 4.3) were also loaded into the table build1_abn_FROM_KM to
cater for the K-Means model results evaluation.
The steps taken
in building ABN models are clearly explained in the research by [Davis, 2004] and I did is simply adopt these steps. Since [Davis, 2004] concluded that the
ABN algorithm provides more accurate results than the Naïve Bayes algorithm in
Oracle for classification algorithms, I then decided to use the default ABN algorithm
settings in building these models. These settings included a SingleFeatureBuild
model type, a maximum number of predictors of 25, a maximum network feature
depth of 10 and no time limit for the running of the algorithm.
Figure 15: Adaptive Bayes Network default algorithm
settings
The two models
created were named OC_abn_Build from the dataset build1_abn_FROM_OC
and KM_ abn_Build from the dataset build1_abn_FROM_KM.
Investigating
the accuracy of the models built here is unnecessary for this evaluation. This
is because any errors or abnormalities found in the algorithm would exert the
same effect on both models built since one algorithm with the same conditions
(i.e. algorithm settings) is used.
The resultant ABN
models were applied to the remaining one third of instances from the respective
cluster models. Table.1. shows a summary of the datasets used in applying a
particular model during the evaluation of the clusters with the ABN algorithm.
|
Dataset used to
build model
|
Resultant model
name
|
Dataset used to
apply model
|
|
build1_abn_FROM_OC
|
OC_ abn_Build
|
apply1_abn_FROM_OC
|
|
build1_abn_FROM_KM
|
KM_ abn_Build
|
apply1_abn_FROM_KM
|
Table 1:
Summary of dataset and model naming
The target
attribute in both instances when applying the ABN models is the ClusterID. The
results were exported to spreadsheets to allow for inspection and comparisons. All
the datasets used are provided in the CD ROM that accompanies this research
paper.
The comparison of
ClusterIDs is between the classification model results and the cluster model
results. Here I wish to find ClusterIDs that appear in both the two distinct
model results. In this case I made a comparison of the K-Means model results
with the ABN model results and a comparison of the O-Cluster model results with
the ABN model results. The comparison and counting was done by importing the
resultant tables into a Microsoft Access database followed by the construction
of SQL queries to perform the actual comparison and counting. Table.2. gives a
summary of the tables that were used in the comparison (the tables contain the ClusterIDs)
while Table.3 depicts the outcome of the comparison.
|
Classification
table
|
|
Cluster table
|
|
OC_APPLY_ABN
|
Vs
|
APPLY_OC3_TSHATSHA
|
|
KM_APPLY_ABN
|
Vs
|
APPLY_KM5_TSHATSHA
|
Table 2:
Database tables compared
Table.2. shows
that the classification tables from the results of the ABN model are compared
with the clustering table results obtained from the cluster algorithm models.
The clustering algorithms basically find clusters in the data; ODM then assigns
each data instance to a particular cluster by assigning it a ClusterID. These ClusterIDs
are removed from the clustering results and are predicted by the classification
algorithm, after which a comparison is made.
|
Data Source
|
ClusterIDs in
both tables
|
Percentage of
ClusterIDs in both models
|
|
From O-Cluster results
|
42 out of 107
|
39%
|
|
From K-Means results
|
18 out of 107
|
17%
|
Table 3:
Results from the comparison of cluster and classification ClusterIDs
Table.3 shows the
percentage of the ClusterIDs that appear in both model results. 39% of ClusterIDs
appeared in both the cluster model results and classification model results for
the O-Cluster algorithm while only 17% for the K-means algorithm. According to
this evaluation technique by [Roiger et al, 2003], the percentage outcome is
treated as a measure of accuracy for the algorithms, were the greater
percentage indicates that that algorithm has more accuracy in finding clusters
of high quality. According to Table 3 the O-cluster algorithm has a larger
percentage hence is more accurate than the K-Means algorithm.
Initially, ten
cluster models were built. However, I then discovered that these initial
algorithm settings were biased in order to perform the evaluation. This bias
made it difficult to point out the best two models. I then decided to build a
further ten models with a slight change in algorithm settings. For the new ten
models, I set the number of maximum clusters (k) to seven for all models. This
removed the bias encountered in the first ten models making it easier to select
the most accurate models. Here the O-cluster algorithm built the most accurate
model.
It is also apparent
from results that the most accurate models selected had found cluster patterns
in the data on applying them to new data. The algorithm evaluation technique
proposed by [Roiger et al, 2003] was then implemented. Here it was determined
that the model built by the O-Cluster algorithm found more accurate clusters
when applied to new data as compare to the K-Means. A further evaluation
follows in the next chapter with a conclusive interpretation of the all the
results.
5 Interpreting Evaluation
Results
This chapter
provides an interpretation of the results obtained so far during the evaluation
process. The interpretation is from all the data mining models built using
similar techniques but different algorithm settings. In this chapter I extract
important facts that will help and support in describing the conclusions
reached regarding the algorithm that has been determined as the most accurate
algorithm. Each model comparison is made then interpreted detailing the facts
for the results obtained.
It has already been
explained in preceding chapters that the results in the 1st ten
models are biased therefore they provide little information for the evaluation
of the two algorithms. In general the average confidence values displayed in Figure
12 (chapter 4) show that increasing k the maximum number of clusters results in
the average confidence increasing for an algorithm model. The bias is that the
O-Cluster seems to be producing more accurate that the K-Means when the value
of k is increased or rather varied.
Although the
confidence for each cluster increases with k, from a logical point of view, the
clusters found tend to become less complex resulting in less interesting
clusters. In simple terms the clusters loss their value. As the number of
clusters increase, the data is divided into many groupings and this result in loss
of attribute relationships. Therefore this should be kept in mind when
increasing the value of k. However it is also difficult to indicate the right
k value to accommodate both accuracy and cluster meaning.
The default
maximum number of clusters for the O-Cluster algorithm is 10 while that for the
K-means is 4. In Figure 12 the models BUILD1_OC_TSHATSHA from the O-Cluster and
BUILD1_KM_TSHATSHA from the K-Means algorithm were built from the algorithm
default settings. On analysing their average confidence values one can see that
the O-cluster algorithm produced a more accurate cluster model. Further more;
it is also clear that the O-cluster algorithm seems to be producing models with
a much higher average confidence value as compared to the K-Means models with
increased k.
Figure 13 in
chapter 4 shows the average confidence and the support values computed for each
model. The major difference in the 2nd model building phase in the
algorithm settings for these ten models as compared to the settings for the 1st
ten models is in the maximum number of clusters that I set to seven for all the
models. Making this parameter, the maximum number of clusters, to be the same
for all the models made the evaluation process unbiased.
Figure 16: Comparison of models with default and adjusted
settings
When the comparison
is inspected from Figure 16 (shortened version of Figure 13), it is clear that
the average confidence for the O-Cluster models is greater than that for the
K-Means in both cases. Since the models were all built with the maximum number
of clusters set to seven it is evident that the O-Cluster algorithm builds more
accurate models as compared to the K-Means algorithm. This conclusion is
primarily based on the results obtained after the building of models with a
variation in algorithm settings for both algorithms.
In order to
determine algorithm accuracy a technique proposed by [Roiger et al, 2003] was adapted.
The technique basically employs supervised learning algorithms to evaluate
unsupervised algorithms. Although the technique determines the final outcome
regarding algorithm accuracy by indicating which algorithm produces more
accurate cluster results, another factor was considered. This deals with how
accurate an algorithm is in building models. This makes it possible to determine
whether model accuracy gives a good indication of model performance when
applied to new data. It is evident as discussed in sections 5.1 and 5.2, that
from the models built and the average confidence figures computed, the O-Cluster
algorithm resulted in the building of more accurate models.
Also to be
emphasised is the effect that the maximum number of clusters (k) has on the
performance of the algorithm during model building. It is clear that for the
K-Means algorithm, the variation of k, the maximum number of clusters, has very
little effect on the accuracy of the resultant model while the opposite is true
for the models built by the O-cluster.
Once the models
had been applied to new data, the results of this step had been used in the building
of Adaptive Bayes Network (ABN) algorithm models. This was followed by applying
the models to the remaining one third of instances (basic implementation of unsupervised
evaluation using technique proposed by [Roiger et al, 2003]). From these
results it is significant that the best model built by the O-Cluster algorithm produces
more accurate cluster results when applied to new data. The implementation step
is discussed in section 4.3.4.
The conclusions
drawn here are that the O-Cluster algorithm produces both more accurate models
and more accurate clusters when applied to new data as compared to the K-Means algorithm
in Oracle data mining.
It is apparent in this chapter that from the results obtained from the
different models that the most effective model was built using the O-cluster
algorithm. The results also show that the model built using the O-cluster finds
more accurate cluster when applied to new data. This was determined by
employing supervised learning evaluation for unsupervised learning, a technique
by (Roiger et al, 2003).
The next chapter involves gathering information from the dataset. This
will need distinguishing the clusters found by the O-Cluster model which
provides more accurate cluster results.
This chapter is
mainly concerned with gathering information from the dataset. Here I need to find
predictors of HIV AIDS prevention behaviour. In order to this it is necessary
to define what we mean by predictors of prevention behaviour. This makes it easier
and feasible to know what we are actually looking for exactly. Therefore my
definition is as follows:
HIV AIDS
predictors of prevention behaviour are attributes within our dataset that influence
an individual to:
a)
use a condom when he/she decides to be sexually active,
b)
lead to abstain from having sexual intercourse for at least a
year or more and
c)
attributes that lead one to having fewer sexual partners. These
are the attributes that I want to find from the dataset.
Now that I have a definition of what the predictors of prevention are,
how do I intend to find these attributes? I firstly intend to use a trial and
error mechanism that involves making use of the O-Cluster algorithm to find
clusters that contain either one of the attributes use_ condom (which
means that the person either makes use of a condom or not during sexual
intercourse), abstnyr (meaning person has abstained for a year or more)
or lespart (meaning has decided to have less or fewer partners).
The other approach that I used in order to determine the predictors was
to employ the association rule algorithm (the Apriori). Association rule mining
searches for interesting relationships among items in a given data set.
The approach I took was to build a cluster model with the O-cluster
algorithm because from my algorithm evaluation, I determined that the O-Cluster
algorithm builds more accurate results than the K-Means algorithm and that this
model by the O-Cluster algorithm also finds more accurate clusters when applied
to new data. Here I simply selected for use the model that I determined as the
most accurate model by the O-Cluster algorithm during the evaluation process
and this was the model Build3_OC_tsha_tsha2, this model was built from
the dataset Tsha_tsha_Build2. I then applied the model to new data to
find the clusters and the new data was in table Tsha_tsha_Apply1.
After I applied
the model to new data it is apparent that cluster patterns were found and this
is shown in the resultant table APPLY_OC3_TSHATSHA, this is evident from
the allocation of ClusterIDs to the dataset instances. The results are exported
to spreadsheets once the cluster model had been applied to new data. The
spreadsheet file name is APPLY_OC3_TSHATSHA. Table.4 shows the
distribution of instances in the clusters found.
|
CLUSTER_ID
|
NUMBER OF
INSTANCES
|
PERCENTAGE OF
INSTANCES (%)
|
|
4
|
55
|
18
|
|
6
|
54
|
18
|
|
8
|
9
|
3
|
|
10
|
61
|
20
|
|
11
|
47
|
16
|
|
12
|
50
|
17
|
|
13
|
23
|
8
|
|
Totals
|
299
|
100
|
Table 4:
Summary of clusters in APPLY_OC3_TSHATSHA
To meet my goal of finding predictors I first have to distinguish the
clusters found. However from this resultant table APPLY_OC3_TSHATSHA it
proves to be difficult due to the large number of attributes in the table (number
of attribute is 21). To overcome this I simply re-apply the same O-Cluster
model to the same dataset, but in the output table I removed any attribute that
I felt had little contribution to solving this problem. This step was repeated
nine times with each predictor that was mentioned in my definition of
prevention predictors being in each table. The outcome is detailed below
corresponding to each table result.
- Table ANALYSE_CLUSTER_1
This is the first output table after applying the model to new data and
was named table ANALYSE_CLUSTER_1 (simply indicating 1st set of
cluster analysis). This table includes the predictor use_condom from the
definition above. The other attributes that were included here were EDUC1
(education level), LESPART (less/fewer partners) and SEX1 (which is
gender). This was to see if these attributes have any relation or influence
to why one would result in using a condom to prevent from being affected by the
HIV virus.
Figure
17: Sample of output table ANALYSE_CLUSTER_1
On analyzing this table after sorting the table according to the
attribute use_condom, I observe that the people who used condoms
(signified by use_comdom=1) are fewer than those who do not. Therefore it
becomes difficult to draw conclusions on what attributes influence condom use.
This is mainly due to the small dataset or sample size.
However from these results it is observed that education level has very
little influence to condom use. The distribution shows that even though people have
at least been to high school some do to make use of condoms while the others do
not. Therefore this table was discarded, had
little information thus resulting in the mining of a new table.
- Table ANALYSE_CLUSTER_2
This table also has the attribute use_condom but now includes
attributes EDUC1 (education level), EMPLOY (if employed or not), HIV_TST
(if person would have an HIV TEST) and HIVTEST (if person has
had an HIV TEST). In this table semi-clear relations were established, it was
apparent that all those people who had an HIV TEST did make use of a condom.
This indicates that if one had an HIV Test (although we don’t know the outcome
of test results) it is apparent that that person eventually makes use of a
condom when having sexual intercourse.
Figure
18: Sample of output table ANALYSE_CLUSTER_2
Therefore, the attribute HIVTEST in this case influences condom
use, hence is a predictor of prevention behaviour.
- Table ANALYSE_CLUSTER_3
In this table I am now looking for attributes that influence one to
abstain from having sexual intercourse for a year or more. I therefore included
the following attributes in the resultant table: ABSTNMTH (abstain for a
month), ABSTNYR (abstain for a year or more), FAITHFL (faithful
to partner), and HIVTEST (have an HIV test).
Figure
19: Sample of output table ANALYSE_CLUSTER_3
On analysis it is observed that ClusterID = 6 is the one cluster that
contains the majority of people who managed to abstain for a year or more.
However, it was also clear that of those who managed to abstain none of them
could be faithful and none had been tested for HIV. This clearly shows that
testing for HIV has no influence in resulting one to abstaining (although it
has an influence on condom use as explained in ANALYSE_CLUSTER_2).
- Table ANALYSE_CLUSTER_4
EMPLOY (employment status), HIV_TST (if person would go for
HIV test), KNOWAIDS (if person knows AIDS), KNOWHIV (if person knows
HIV), LESPART (if person has decided to have fewer partners) and TALK_OPN
(if person talks openly about the virus) were attributes included in this
table.
Figure
20: Sample of output table ANALYSE_CLUSTER_4
Here focus is on the attribute LESPART from my definition of
prevention predictors: Observations in this table show that these results are
not really useful. This is because 277 out of 299 people in this table did not
decide in having fewer partners thus leaving only 20 people out of 299 who
decided to have fewer partners. However from the 20 only 2 seem to know about
AIDS. Thus no conclusions can be draw due to small sample size.
- Table ANALYSE_CLUSTER_5
Attributes included in this table are ABSTNYR (abstain for a year
or more), KNOWAIDS (person knows AIDS), KNOWHIV (person knows HIV) and MARRIED
(marital status).
Figure
21: Sample of output table ANALYSE_CLUSTER_5
Attribute in to focus on is ABSTNYR: Observations here draw an
interesting pattern from the results. It is evident that ClusterID = 6 is the
only cluster that contained people who have managed to abstain for a year or
more. And from this same cluster it is evident that the majority of the people
know about AIDS. This alone brings me to the conclusion that one abstains
because he knows AIDS (the dangers and effects of it). Therefore I
conclude that Knowing about AIDS somehow leads to abstinence.
- Table ANALYSE_CLUSTER_6
From the results obtained in previous table I then decided to continue
investigating the attribute ABSTNYR (abstain for a year or more) hence included
different attributes to see what outcome I get if I keep in mind that ClusterID
= 6 contains only people who managed to abstain for a year or more. These other
attributes are AGE1 (persons age), HIVTEST (if been tested for
HIV), MARRIED (marital status), SEX1 (gender) and YTHHELP
(if person has volunteered to help care for people who have been infected with
the virus).
Figure
22: Sample of output table ANALYSE_CLUSTER_6
Since the attribute in focus here is ABSTNYR: Observations show
that the people who abstained (mainly in ClusterID = 6) were not influenced by
having an HIV test, by their marital status or gender. This therefore brings a
conclusion that other attributes not included in this table may have influenced
these people to abstain.
- Table ANALYSE_CLUSTER_7
For this table I decided to focus on the use of condoms again by
including the following attributes: USECOND, EDUC1 and SEX_YET
(if person has had sex before or not). From this table I was wishing to
establish a relationship between condom use and education level, but it has
proven impossible to determine whether literacy (education level) plays a role
in why one may use a condom. These results are thus not really clear to draw
any conclusions.
Figure
23: Sample of output table ANALYSE_CLUSTER_7
- Table ANALYSE_CLUSTER_8
This table has the following attributes LESPART (fewer partners), KNOWAIDS
and KNOWHIV. Here I turn my attention to the attribute fewer partners,
where I wish to determine if knowing both HIV and AIDS leads to one resulting
in having fewer partners. The problem here though is that the sample size of
those people who decided (for whatever reason) to have fewer partners is very
small (23 out of 299). Therefore it is difficult to draw conclusions here.
Figure
24: Sample of output table ANALYSE_CLUSTER_8
- Table ANALYSE_CLUSTER_9
From table ANALYSE_CLUSTER_5, I observed that the ClusterID = 6 is
different from all other clusters in that it only contains people who have
managed to abstain. So in this table (ANALYSE_CLUSTER_9) I intend to
further find other attributes that have closely influenced an individual to be
able to abstain for a year or more. To do this I included the following
attributes together with ABSTNYR (abstain for a year or more), and these
are EDUC1 (education level), EMPLOY (employment status), FAITHFL
(faithful to partner), HOUS_TYP (your house or home type), KNOWAIDS,
KNOWHIV, MARRIED, SEX1 (gender) and YTHHELP.
Figure
24: Sample of output table ANALYSE_CLUSTER_9
On analyzing this table, I observed that the majority of the people who
abstained (see ClusterID = 6) have at least been to high school. However this
cannot be conclusive as it is evident that for other clusters (containing
people who failed to abstain) had a similar distribution. It is also clear that
gender and the other attributes included here did not make significance to why
people abstain.
6.1.1 Conclusions drawn from the Analyse Cluster tables
Observations in table ANALYSE_CLUSTER_2 make it evident that the
attribute HIVTEST influences condom use. However, ANALYSE_CLUSTER_5 clearly
concludes that knowing about AIDS leads to abstinence. Therefore from these
table observations I have concluded that the attributes HIVTEST and KNOWAIDS
have been clearly been identified as predictors of prevention behaviour.
According to
[Al-Attar, 2004], association rules are similar to decision trees and
association rule induction is the most established and effective of the current
data mining technologies. This technique involves the definition of a business
goal and the use of rule induction to generate patterns relating this goal to
other data fields. The patterns are generated as trees with splits on data
fields. This technique allows the user to add their domain knowledge to the
process and decide on attributes for generating splits [Han et al, 2001].
In order to
employ the association rules, I simply applied the Apriori algorithm which is
the only association rule available in ODM to the datasets TSHA_TSHA_APPLY1 and
APPLY_OC3_TSHATSHA. The resulting tables from the algorithm sifting through
these datasets were named ASSOCIATION_MODEL for dataset APPLY_OC3_TSHATSHA and
ASSOCIATION_MODEL2 for dataset TSHA_TSHA_APPLY1 and these were exported to
spreadsheet files for analysis. The spreadsheets were also named as
ASSOCIATION_MODEL and ASSOCIATION_MODEL2 and these are available on the CD ROM
that accompanies the project.
The dataset
APPLY_OC3_TSHATSHA was used to find rules because it the resultant data that
contains the ClusterIDs found using the O-Cluster model. Analysing the rules in
ASSOCIATION_MODEL I observed the following.
The majority of
relationships here lead to people not having less partners, not using condoms,
being unfaithful to partners in an event to prevent HIV infection. Most of the
relationships are however meaningless.
From the table it
is however evident that ASSOCIATION_MODEL is giving relationships of attribute
behaviour from a negative perspective (for example, if one is married and talks
openly about HIV AIDS then they will not be able to abstain for a year). This
is making the identification of predictors difficult. By interpreting these
relationships, I observe that even though individuals may talk openly about HIV
AIDS and married, this will not result them in having less partners or abstain
for a year. Therefore, in this association table it is difficult to identify
predictors.
The dataset
TSHA_TSHA_APPLY1 is the dataset that was used when applying the cluster models.
Analysing the outcome of this dataset (outcome is in ASSOCIATION_MODEL2) I
observe the following; an individual who talks openly about HIV AIDS and hasn’t
had sex before is most likely to use a condom when he/she decides to have sex.
Table results also show that if one had an HIV Test and had sex before, he/she
is most likely to abstain for a year. The majority of the association rules here
show that one will have fewer partners, if one has at least been to high
school, has had an HIV Test and talks openly about HIV.
Here
ASSOCIATION_MODEL2 is giving relationships from a positive perspective, for
example, an individual who talks openly about HIV AIDS and hasn’t had sex
before is most likely to use a condom when he/she decides to have sex.
.Therefore for this table it is evident that the most likely predictors are
having an HIV test and talking openly about HIV AIDS.
It must be
mentioned that the identification of HIV AIDS prevention predictors is a very
difficult process which needs reasoning and domain knowledge. From the
associations the attributes HIV test and talk openly have been
identified as predictors for condom use as prevention. Although, the
justification of these predictors is mainly based on the clusters and
association rules it must be mentioned that the interpretation of the
relationships in the associations is solely dependent on domain knowledge of
the field.
Analysis of the
Clusters found was made in a trial and error approach. However it was evident
that there was a significant distinction between some clusters, namely
ClusterID=4 and ClusterID = 6. These two clusters made it possible to identify
predictors. In the cluster analysis the attributes HIVTEST and KNOWAIDS have
been clearly been identified as predictors of prevention behaviour of condom
use and abstinence.
By employing an association rule algorithm, the Apriori, the attributes HIV
test and talk openly were identified as HIV AIDS predictors of
prevention of condom use.
The following chapter will summarise the conclusions that have be drawn
from all the results obtained. It also gives a conclusion regarding Oracle Data
Miner as a mining tool.
7 Summary
and Conclusions
In this chapter,
the main results of this work are summarised. This project has been a very
successful one, with the major project aims and goals met. There were two main
objectives in this research. Firstly, it was aimed at analysing and evaluating
the effectiveness of two Clustering algorithms, the O-Cluster and K-Means in
terms of accuracy in building models and finding clusters when applied to new
data. The second objective was to find predictors of HIV AIDS prevention
behaviour attributes from the dataset obtained from the Centre for AIDS
Development, Research and Evaluation Institute for Social and Economic
Research, Rhodes University.
7.1 Key Results of the Work
[Roiger et al, 2003] discusses aspects
of evaluating the performance of models built during data mining. According to
these authors, when evaluating performance it is necessary to consider whether
the results of the data mining can be interpreted and whether the results can
be used with confidence. In regard to this it was observed that the O-Cluster
algorithm builds more accurate models as compared to the K-Means algorithm. The
outcome of these accuracies is depicted by average confidence and support
values in Figure 12 and Figure 13.
The most
effective model built was the model built by using the O-cluster algorithm as it
possesses an accuracy figure of 95,5% (Figure 13). On applying this model to
new data and using the evaluation technique by [Roiger et al, 2003] to analyse
the accuracy the algorithm model has in finding clusters, it was evident that
42 out of 107 (39%) ClusterIDs were correctly predicted in the outcome of the
Adaptive Bayes Network model. This shows that the model built using the
O-Cluster algorithm performs better than the model built by the K-Means
algorithm which only managed to correctly predict 18 out of 107 (17%) ClusterIDs.
In order for an
algorithm to perform well, it was necessary to tune the algorithm by adjusting
its settings. According to the model results obtained from the algorithm
evaluation phase, it is evident that tuning in some cases results in the
algorithm performing well and sometimes performing badly in terms of model
accuracy.
Observations
indicated that tuning the K-Means algorithm had very little effect on
increasing the accuracy of the models built by the algorithm. On the other hand,
tuning the O-Cluster algorithm proved to make the algorithm perform better thus
building the most accurate model as depicted in Figure 12 and 13. The setting to
enable tuning for the O-Cluster is the Sensitivity while those for the K-Means are
the Minimum error tolerance and Maximum Iterations.
A few attribute
predictors of HIV AIDS prevention were found from the dataset. The process of
finding these involved critical analysis of the results as well as good
reasoning. These were the attributes HIV test and Know Aids have
been clearly been identified as predictors of prevention behaviour of condom
use and abstinence. By employing the Apriori algorithm, the attributes HIV
test and talk openly were identified as HIV AIDS predictors of
prevention of condom use.
Ø HIV
test – if one has had an HIV test
Ø Know
Aids – if one knows about AIDS
Ø Talk
openly – if one talks openly about HIV AIDS or not
The fact that Oracle Data Mining’s data mining functionality is
embedded inside the Oracle Database makes the data gathering and preparation
process simpler. It also makes model building and model applying very simple as
this is evident from the step by step wizards provided by the mining tool. ODM
provides easy and reliable access to the database and the tables stored in the
database. This makes it possible to search for a specific data set during the
data mining process. This also allows the user to view summaries of the data
including distributions of attributes in the data set, which is of use during
the data preparation phase. Any results obtained can also be exported to
spreadsheets allowing increased accessibility and ensuring they can be easily
worked with.
ODM
supports the following algorithms as stated by [Berger,
2004] and the evaluation results of some of these algorithms investigated seems
to be interesting.
·
Enhanced k-Means Clustering (clustering)
·
Orthogonal Partitioning (clustering)
·
Adaptive Bayes Network supporting decision trees
(classification)
·
Naive Bayes (classification)
·
Model Seeker (classification)
·
Apriori (association rules)
The results
obtained in this research when making an evaluation of the two Clustering
algorithms available in ODM have shown that the O-Cluster algorithm builds more
accurate models as well as in finding more accurate clusters when the model is
applied to new data as compared to the K-Means. On the other hand, [Davis, 2004] also proves in her research that for the evaluation of Classification
algorithms between the Adaptive Bayes Network and the Naive Bayes, the ABN algorithm gives more accurate results when used
to build models as well as in applying the models to new data.
It must be
emphasised that the ease and speed of building a model using the wizards allows
for a number of models to be built at the same time. This approach is recommended
in order to ensure the most effective model possible is produced.
Following on from the conclusions and recommendations of the theory
research, I have managed to conclude that the O-Cluster algorithm has
been identified as the most accurate clustering algorithm in Oracle data mining
10g. The results achieved by the evaluation fulfil the aims of this project. However
considerable time was spent laying the groundwork for the
practical stages were different proposals were considered as to how to
implement theory suggestions in a practical way. Eventually, a novel solution
was adapted and hence leads to these results.
REFERENCES:
[Berger,
2004] Oracle Data Mining (Know More, Do
More, Spend Less) - An Oracle White Paper, October 2004 by Berger,C.URL:http://www.oracle.com/technology/products/bi/odm/pdf/bwp_db_odm_10gr1_1004.pdf,
Accessed: 14 April 2005
[Berry et al, 2000] Mastering Data
Mining: The Art and Science of Customer Relationship Management by Michael
J.A. Berry and Gordon S. Linoff, USA, Wiley Computer Publishing, 2000
[Bradley et al, 1998] Refining
Initial Points for K-Means Clustering by P. Bradley and U. Fayyad: ICML 1998
[Davis, 2004] An Evaluation
of Commercial Data Mining by Emily Davis. Oracle Data Mining Honours
Research Project 2004. URL: http://www.cs.ru.ac.za/research/previous/g01d1801/
Accessed: 14 September 2005
[Elder et
al, 1998] Fourth International Conference on
Knowledge Discovery & Data Mining Research - Friday, August 28, 1998 by John F. Elder IV & Dean W. Abbott Elder, http://www.datamininglab.com
Accessed:
11 September 2005
[Han et al, 2001] Data mining:
concepts and techniques by Jiawei Han and Micheline Kamber, San Francisco,
California, Morgan Kauffmann, 2001.
[Insightful
Miner, 2003] Insightful Corporation Seattle, Washington, Insightful Miner 3.0 User Guide:
http://www.insightful.com/support/iminer30/getstart.pdf,
June 2003- Accessed: 11 October 2005
[Mannila et al,
2001] Principles of data mining, by David
Hand, Heikki Mannila and Padhraic Smyth, , Cambridge Massachusetts, MIT Press,
2001.
[Oracle,
2005] The Oracle Home Page. Revised
February 2005. Oracle 10g Data Mining FAQ
http://www.oracle.com/technology/products/bi/odm/odm_10g_faq.html#api
Accessed:
17 May 2005
[Oracle
Installation, 2004] Oracle Database Installation Guide 10g
Release 1 (10.1.0.2.0) for 64-bit Windows- http://download-east.oracle.com/docs/html/B13805_02/toc.htm Accessed:
2 July 2005
[ODM Release Notes, 2004] Oracle Data Miner 10.1Release
Notes and Installation Instructions November 2004. http://www.oracle.com/technology/products/bi/odm/odminer_install.htm Accessed: 2 August 2005
[ODM tutorial, 2004] Oracle Data Miner
Tutorial Part 1 & Part 2, 1 September, 2004. http://www.oracle.com/technology/products/bi/odm
Accessed: 17 August 2005
[Roiger et
al, 2003] Data mining: a tutorial- based primer
by Richard J. Roiger and Michael W. Geatz, Boston, Massachusetts, Addison
Wesley, 2003, pg 58 and 232.
Oracle10g
version 10.1.0.2.0 and Oracle Data miner 10g version 10.1.0.2.0 were installed on
the server machine – Athena.ict.ru.ac.za using the installation guide from [ODM
Release Notes, 2004]
Two errors have
been obtained as a result of this installation guide with the one discussed
first being the major problem.
Since the Oracle
database 10g and ODM are both version 10.1.0.2, this installation guide
strongly recommends that you upgrade to the following patch sets:
Ø Oracle
database 10.1.0.3 and
Ø Oracle
Data Mining 10.1.0.3.1
On installing and configuring all the required software as described in
this installation guide the following errors appear when logging into the data
mining server:
Ø
odm api is incompatible with odm api for the schema and the
required is odm api 10.1.0.2.0
Figure 25: Error obtained after installations
Discussion of Problem and Solution:
Installing the
above patches was hoped to upgrade the Oracle database from version 10.1.0.2.0
to version 10.1.0.3 but that was not the case. The patches would not upgrade
the database to the higher version as required while the ODM does upgraded to
version 10.1.0.3.1. When upgrading, the odmapi.jar used by Oracle Data Miner, must be the
same version as that installed in the database. So the patching of ODM 10.1.0.3
results in the loading of a new odmapi.jar into the database. The result of the patching
makes the database and ODM versions to be incompatible with each other.
The database
version can be obtained by logging into the database using the Enterprise
manager then run the following sql query from an application iSQL plus provided
within the enterprise manager (iSQL plus is similar to SQL plus, difference is
that iSQLplus runs on a web browser while SQLplus is an application ). Refer to
Appendix C for instruction on how to logon to the enterprise manager on the
machine Athena.
->SQL>
select COMP_ID, COMP_NAME, VERSION, STATUS from dba_registry where COMP_ID =
'ODM'
returns:
COMP_ID
------------------------------
COMP_NAME
------------------------------------------------------------------
VERSION
STATUS
------------------------------
-----------
ODM
Oracle
Data Mining
10.1.0.2.0
VALID
The solution is
to de-install the patches and replace the original ODM API in both the ODM and
Oracle database. The original odmapi.jar for this database is for the
version 10.1.0.2.0. This api can be obtained from the Oracle website or by the
reinstallation of the Oracle 10g database version 10.1.0.2.0 (and can be found
in the database lib directory: C:\oracle\product\10.1.0\em_2\dm\lib in this
database I installed on athena). The same version of the odmapi.jar has to be
placed in the Oracle data miner lib directory (C:\odminer\lib).
I therefore
recommend that when installing and configuring the database and the data miner
not to install or patch with the Oracle database 10.1.0.3 and Oracle Data
Mining 10.1.0.3.1 patches because this results in the error discussed above.
Another error that I encounter was the same as that
depicted by Figure.15. The error will be obtained if the Oracle database is not
properly configured for data mining. Here a data mining user account in the database
is configured so that it can be linked to and used by the Oracle data miner.
The proper way to configure the database for data mining is discussed in detail
in Appendix B, were I provide all the necessary steps to installing and
configuring the database and Oracle data miner for data mining.
This
configuration is the Oracle 10g database version 10.1.0.2 and Oracle data miner
version 10.1.0.2 on a Windows Platform (Microsoft Windows Server 2003, Standard
Edition). Basically the Oracle Data Mining server runs against an Oracle 10g
Database. Both the database and ODM can be downloaded from the Oracle
website.
STEP 1-Install Database
For the
installation of the Oracle 10g database refer to the Oracle
Database 10g Release 1 (10.1) Documentation [Oracle Installation,
2004].
It is recommended to install the Enterprise version of the
database as it contains all the necessary all the data mining tools needed by
the Oracle data miner software.
After a successful
installation, all the ODM software is located in the directory:
C:\oracle\product\10.1.0\em_2\dm
During installation of the database the passwords for all default users
was set and in this case it was set to oracle10g for all users. The
username for super user is sys and the corresponding password is
oracle10g. To log into the database you use the Enterprise manager (EM).
The EM for oracle10g is accessed through a web browser and the URL for this
particular database is URL:http://athena.ict.ru.ac.za:5500/em/
Enter the credentials, username-sys, and password-oracle10g
then Connect as-SYSDBA
When I logged onto the database I then create my own
data mining user account:
Username-odm_use AND password-datamine then
Connect As normal user.
STEP 2-Install Data miner
After
downloading the data miner version 10.1.0.2.0 (odminer.zip file) unzip the
entire contents of odminer.zip to the directory C:\odminer.
To run the Data
miner, you run (double click) the odminerw.exe executable file located in the
directory where Oracle Data Miner was installed. In this case it can be found
in the directory; C:\odminer\bin\odminerw.exe
STEP 3-Configure Data miner user
To configure the
database for a data miner user the script odmuser.sql is executed in SQL*Plus which is
located in located in the directory containing all the Data mining software (C:\oracle\product\10.1.0\em_2\dm\admin).
Therefore to be able to run the script you need to logon to the database as a
super user (in this case as sys).
Start SQL*Plus and login as follows:
>SQLPLUS sys/oracle10g as
sysdba
Run the script odmuser.sql with the following
command:
SQL> @
C:\oracle\product\10.1.0\em_2\dm\admin\odmuser.sql
The output is as follows:
=============================================================
Enter value for 1: odm_use // data mining user account we are giving mining permissions in the
database
Enter value for 2: datamine //corresponding password
Enter value for 3: odm // default tablespace for this user
old 1: drop user &USERNAME cascade
new 1: drop user odm_use cascade
drop user odm_use cascade
*
old 1: create user &USERNAME identified by
&PASSWORD default tablespace &TBSNAME temporary tablesp
new 1: create user odm_use identified by datamine default
tablespace odm temporary tablespace temp
create user odm_use identified by datamine default
tablespace odm temporary tablespace temp quota un
*
old 9: to &USERNAME
new 9: to odm_use
Grant succeeded.
old 1: grant execute on ctxsys.ctx_ddl to &USERNAME
new 1: grant execute on ctxsys.ctx_ddl to odm_use
Grant succeeded.
old 1: grant DMUSER_ROLE to &USERNAME
new 1: grant DMUSER_ROLE to odm_use
STEP 4-load ODM sample datasets
This step is
option if you have your own datasets to perform the data mining, but it becomes
necessary if one needs to follows the data mining tutorial which will be
discussed in Appendix C. In is also important to note that all data mining is
performed by the data mining user granted permission as shown above and all the
data that is intended to be used during mining is loaded into that user’s table
space. The database only grants one data mining user.
To load the sample
datasets execute the following scripts:
First run the odmtbs.sql
script to create a tablespace and specify the location of the tablespace file
to be used by the data mining users.
Ø
SQL>C:\oracle\product\10.1.0\em_2\dm\admin\odmtbs.sql
TEMP01.DBF C:\oracle\product\10.1.0\oradata\orc2\ TEMP01.DBF
Run the dmuserld.sql
script to load the sample data mining datasets into the specified Data Mining
User schema as follows:
Ø
SQL> C:\oracle\product\10.1.0\em_2\dm\admin\dmuserld.sql
odm_use datamine
C:\oracle\product\10.1.0\oradata\orc2\
TEMP01.DBF
The ODM installation
guide specifies that for Oracle10, ODM Java and PL/SQL sample programs use
datasets shipped with the common schema SH. Therefore it is necessary to
execute the following script to grant necessary SH access privileges. As
mentioned all default database user password was set to oracle10g. To grant
these privileges run the script dmshgrants.sql as follows:
Ø
SQL> @ C:\oracle\product\10.1.0\em_2\dm\admin\dmshgrants.sql
oracle10g odm_user
Then logout of SQL*Plus as superuser
Ø
SQL> exit
Re-Login as the Data mining user (odm_use)
Ø
sqlplus odm_user/datamine
Execute the
script dmsh.sql to create related views and tables in the SH Schema.
Ø
SQL> C:\oracle\product\10.1.0\em_2\dm\admin\dmsh.sql
STEP 5-launching the Oracle data
miner
Double click the executableà 
odminerw.exe
Since
this is the first time you are launching ODM Data Miner (odminerw.exe from the
directory we installed the data miner: C:\odminer\bin\odminerw.exe),
you will be prompted to create a new database connection. At this point simply
enter the information for the user odm_use as follows:
Figure26: Connection settings for odm_use
Note that the Connection
name can be anything but in this case I set it to orc2 (the database name)
NB: if all the above installation and configuration steps
are implemented successfully, the
following screens should be seen when logging on to the ODM server.
Figure 27: Select orc2 connection name
Figure 28: connecting to the ODM server
Figure 29: ODM server interface when connected
This appendix aims to provide the reader with information on how to use
the Oracle data miner used in the evaluation of the algorithms in this research
project. I assume the reader has access to the Oracle10g Data Mining
Tutorials which are available on the CD-ROM that accompanies this project.
It is also assumed the user has access to the server machine
Athena.ict.ru.ac.za which has the Oracle10g Release 1 version 10.1.0.2 and
Oracle data miner version 10.1.0.2 installed and configured for use.
C.1 Preparing for Data Mining
The server machine was installed with the mining tools such that any user
can make use of the mining tools. However if the reader has no credentials to
the machine, it is possible to login to the machine with the username – g05m5125
and the password – final22.
Ø Launch
the Oracle Data mining server, this is located in the following directory:
C:\odminer\bin\odminerw.exe
Double click the executable odminerw.exe (look at
Appendix C step 5)
Ø If
it is required to load any datasets into the database it will be necessary to
create database tables first. This is achievable using the Enterprise manager
(EM). The EM for oracle10g is accessed through a web browser and the URL for
this particular database is URL:http://athena.ict.ru.ac.za:5500/em/
Figure 30: login screen to Oracle10 Enterprise Manager
Here enter the credentials, username-sys, and password-oracle10g
then Connect as-SYSDBA (the system database administrator account) or the Data
miner user account with Username-odm_use AND password-datamine
then Connect As normal user.
When logged into the database, through the Enterprise
Manager, you can view the tables available or even create tables to load new
data (go to Administration à Tables).Any
data sets that will be mined should be loaded into the odm_use schema in
the Oracle database. The loading of data can be accomplished by using sql*loader
(sqlldr.exe).
Starting sql*loader
Click on Start > programs > command prompt then cd (change
directory) to the directory with the data you want to load into the database. Then
type the command sqlldr to start the application. Note that sql*loader is
an application for loading data into an Oracle database hence it can only be
found on machine with the Oracle database installed.
To make a batch load of data sets
into a table
Firstly create a control file and save it as a .ctl file in the same
directory as the dataset file. The control file contains information on where
the dataset file can be found and which database table the dataset is going to
be loaded to. Most success in loading datasets was achieved by the use of .txt
files with items in records separated by commas and each record on a new line.
Open command
prompt, then change directory to the location where the .ctl file is located
then type sqlldr (sql loader application) and odm_use/datamine
(ODM account name & password) as follows to load the dataset into
the database table:
Ø C:\DATA\tsha_tsha>sqlldr
odm_use/datamine
This will
prompt for the control file name after which it will load the data.
Figure 31: Example of how to load data
C.2 The Actual Data Mining
At this stage the reader should refer to the Oracle10g Data Mining
Tutorials in the CD ROM that accompanies this research project. The
tutorials provide a step by step mining methodology. The tutorials are in
different stages, the building by a particular algorithm and applying of the
model built. For this research the major algorithms used are the Clustering
algorithms (O-Cluster and K-Means), however tutorials for all algorithms used
will be provided in the CD ROM.
The dataset files used in this evaluation are available in
the CD ROM that accompanies this project.
The spreadsheet files used for the evaluation and in the
identification of predictors are also available in the CD ROM that accompanies
the project.