A Comparison of Rule Induction Using Feature Selection and the LEM2 Algorithm - Feature Selection for Data and Pattern Recognition - page 166

Table 8.6 Data sets used for experiments
Data set

Number of Consistency (%)

Cases Attributes Concepts

Australian Credit Approval 690 14 2 100

Breast Cancer—Slovenia 286 9 2 95.45

Breast Cancer—Wisconsin 625 9 9 94.08

Bupa Liver Disorders 345 6 2 100

Glass 214 9 6 100

Hepatitis 155 19 2 100

Image segmentation 210 19 7 100

Iris 150 4 3 100

Lymphography 148 18 4 100

Pima 768 8 2 100

Postoperative patients 90 8 3 84.44

Soybean 307 35 19 100

Primary Tumor 339 17 21 76.40

Wine Recognition 178 13 3 100

Let us say that attribute a has missing attribute value for case x from concept C
and that the value of a for x is missing. This missing attribute value is exchanged by
the known attribute value for which the conditional probability of a for case x given
C is the largest.
Some of these data sets had numerical attributes ( Australian Credit Approval ,
Bupa Liver Disorders , Primary Tumor and Wine Recognition ). Numerical attributes
were discretized using cluster analysis methods of discretization [ 5 ].
The data mining system LERS uses for discretization a number of discretization
algorithms [ 13 ]. In our experiments we used two approaches to discretization based
on cluster analysis. First, all numerical attributeswere normalized [ 7 ] (attribute values
were divided by the attribute standard deviation).
In our first discretization technique, based on agglomerative cluster analysis [ 7 ],
initially each case is a single cluster, then clusters are fused together, forming larger
and larger clusters. In remaining four cluster analysis discretization methods, where
we used divisive techniques, initially all cases are grouped in one cluster, then this
cluster is gradually divided into smaller and smaller clusters. In both methods, during
the first step of discretization, cluster formation , cases that exhibit the most similarity
are fused into clusters.