PSO-based method for learning similarity measure of nominal features

摘要

This paper presents a PSO-based method for learning similarity measure of nominal features for case based reasoning classifiers (i.e. CBR classifiers). The symbolic features considered here takes completely unordered values. It has been indicated in [3] that in specific classification task, the similarities between these nominal feature values can not be simply considered as either 0 or 1. A GA-based approach has been developed for learning similarity measure of such feature values. However, when the number of features and feature values become larger, the GA-based algorithm's convergence speed obviously slows down, and the accuracy of classification may be also affected. To address this problem, we propose a PSO-based algorithm for learning similarity measure of nominal features and further describe feature importance through the learned similarity measure. The experimental results show that, using the proposed PSO-based algorithm, the convergence speed is much faster than that of GA-based algorithm and the accuracy is also improved. In addition, we also explain that the feature importance defined through the learned similarities is essentially consistent with that in rough sets, and an illustrative example is finally provided.