Fuzzy control GA with a novel hybrid semantic similarity strategy for text clustering

摘要

This paper proposes a fuzzy control genetic algorithm (GA) in conjunction with a novel hybrid semantic similarity measure for document clustering. Since the common clustering algorithms use vector space model (VSM) to represent document, the conceptual relationships between related terms being ignored, we use semantic similarity measures to solve this problem. In general, the semantic similarity measures can be extensively categorized into two kinds: thesaurus-based methods and corpus-based methods. However, in practice the corpus-based method is rather complicated to tackle. We propose and demonstrate a semantic space model (SSM) as the corpus-based method, where the appropriately reduced dimensions in SSM can capture the true relationship between documents in terms of concepts, rather than specific terms. Thus, the thesaurus-based method is combined with our SSM as a hybrid strategy to represent the semantic similarity measure. In GA field, the balance between the capability to converge to an optimum and the capacity to explore new solutions affects the success of search for the global optimum. We utilize a fuzzy control GA to adaptively adjust the influence between these two factors. Two textual data sets from Reuter document collection and 20-newsgroup corpus are tested in our experiments, and the results show that our fuzzy control GA combined with the hybrid semantic similarity strategy apparently outperforms the conventional GA, FCM and K-means with the traditional cosine similarity in VSM. Moreover, the superiorities of the fuzzy control GA and our hybrid semantic strategy are demonstrated by their better performance, in comparison with conventional GA with the same similarity measures.