Learning Similarity Of Symbolic Feature Values And Feature Importance Measure

When studying cased-based reasoning classifiers (i.e. CBR classifiers), similarity between features occupies a decisive role for the results of classification and decision-making. Similarity measure between features depends on calculation of similarity between each feature value. This paper learns similarity between values of symbolic features. The symbolic features considered here have completely unordered values, such as for the feature"color", values are"red","yellow"and"blue". Most researchers considers similarities between these feature values can only be either 0 or 1, this approach will lead to loss of information. Existing work has improved these values from {0, 1} to [0, 1] and has presented A GA-based approach for learning similarity measure of symbolic feature values.However, when the number of feature values and features become larger, the convergence speed based GA obviously slows down, and the accuracy of classification may be also affected. Considered this reason, this paper proposed a PSO-based method to get similarity measure of symbolic features. The results of the experiments show that, the convergence speed based on PSO algorithm is much faster than based on GA algorithm and the accuracy is also improved.In addition, this paper has futher indicated that similarities of feature values we learnt can roughly reflect feature importance and proposed feature importance measure. Experiments proved the feasibility of this measure,Lastly, based on Rough Set theory, this paper proposed a new way to judge whether interaction exits among features in a decision table.