A co-evolutionary framework for adaptive multidimensional data clustering

Abstract

Data clustering refers to constructing groups of objects that are highly correlated, based on some similarity measure. It is a very popular technique for intelligent knowledge discovery. A challenge that arises in automatic data clustering, though, is the high dimensionality of data, since each object can be described by several relevant features. Thus, we often need to assign a relative weight for each feature to indicate its importance during the clustering process. With the absence of domain knowledge about the nature of data, assigning such weights becomes a challenging task. Dynamic adjustment of feature weights in an unsupervised manner is an attractive solution for such problem. In this paper, we propose a co-evolutionary algorithm for the dynamic adjustment of feature weights during data clustering. Two populations are simultaneously evolved for the optimization of both the clusters and their associated feature weights. In addition, the number of clusters are also learnt and optimized in the evolutionary process. Extensive experimental results on several datasets from UCI machine learning repository indicate the efficacy of the proposed approach. The algorithm outperforms both a non-adaptive version, where feature weights are not considered, as well as K-means clustering for a fixed number of clusters.