Using Association Rules for Classification from Databases Having Class Label Ambiguities: A Belief Theoretic Method - Data Mining: Foundations and Practice - page 540

Table 12. Ambiguous UCI databases - classification accuracy (score)

Database KNN KNN-BF ARM-KNN-BF

0.92 0.82 0.94

Breast cancer

0.89 0.83 0.88

Car

0.78 0.76 0.88

Diabetes

0.91 0.94 0.92

Iris

Monks 0.86 0.87 0.89

Post-operation patient 0.82 0.76 0.85

Scale 0.81 0.79 0.84

Solar flares 0.80 0.80 0.84

Tic-Tac-Toe 0.80 0.82 0.85

0.86 0.81 0.89

Voting

0.87 0.87 0.91

Wine

0.89 0.91 0.93

Zoo

The '% Reduction' column in Table 11 shows the percentage reduction
in the number of rules used by the ARM-KNN-BF classifier when compared
with the others. In calculating the classification accuracy in Table 12, some
strategy needs to be developed to compare the 'correctness' of ambiguous
classifications. For example, suppose the actual class label is ( C 1 ,C 2 ). How
does the accuracy calculation be done if the classified label is C 1 ,( C 1 ,C 2 )
or ( C 2 ,C 3 )? Clearly, the classification ( C 1 ,C 2 ) must be considered 'perfect'
and should be assigned the maximum score. How does one evaluate the clas-
sification accuracies corresponding to the classifications C 1 and ( C 2 ,C 3 )? To
address this issue, the following measure, which we refer to as the score ,is
employed:

score = |

True label ∩ Assigned label |

(24)

| True label ∪ Assigned label |

With this measure, the scores of the class labels C 1, ( C 1 ,C 2 )and( C 2 ,C 3 )
would be 1 / 2, 1 / 1and1 / 3, respectively. Table 12 gives the score values for
the three classifiers.

3.3 Experiments on the Airport Terminal Simulation Platform

We have developed a simple simulation platform to mimic an airport terminal
in our Distributed Decision Environment (DDE) Laboratory at the Depart-
ment of Electrical and Computer Engineering, University of Miami, to test
the performance of the algorithms we have developed. The platform consists
of three areas; each area has two gates at its entrance and exit. The potential
threat carriers carry different combinations of features which we refer to as
property packs . Carriers entering and exiting each area are tracked using an
overhead camera. Each gate of the platform contains a stationary sensor mod-
ule that extracts the intensity of the features in the property pack. Based on