An Adaptive Genetic Algorithm Approach for Optimizing Feature Weights in Multimodal Clustering

Abstract

Social Media is a popular channel of communication, where people exchange different types of high volume and multimodal data. Cluster analysis is used to categorize this data to extract useful information. However, the variation of features that can be used in clustering makes the clustering process difficult, since some features may be more important than others, and some may be irrelevant or redundant. An alternative to traditional feature selection techniques, especially with the absence of domain knowledge, is to assign feature weights that depend on their importance in the clustering process. In this paper, we introduce a multimodal adaptive genetic clustering (MAGC) algorithm that clusters data according to multiple features. This is done by adding feature weights as an extension to the clustering solution. In other words, feature weights evolve and improve alongside the original clustering solution. In addition, the number of clusters is not determined a priori, but it is adapted and optimized during the evolutionary process as well. Our approach was tested on a large collection of Flickr images metadata and was found to perform better than a non-adaptive genetic algorithm clustering approach and to produce semantically related clusters.