Multiviewpoint-Based Agglomerative Hierarchical Clustering

The cosine similarity is a similarity measure useful for document clustering. The cosine similarity between two points is determined by the angle between their corresponding vectors observed from the single reference viewpoint, the origin. Recently, Nguyen et al. [6] proposed a new similarity measure called MVS (MultiViewpoint-based Similarity) in which the vectors are observed from multiple viewpoints. They incorporated MVS into some non-hierarchical clustering algorithm and showed that MVS outperforms the original cosine similarity. This paper proposes an agglomerative hierarchical clustering which couples the average-link method with MVS. Despite MVS is more complex than the cosine similarity, our clustering algorithm achieves the same time complexity as the average-link method with the cosine similarity by computing the inter-cluster similarity smartly. Interestingly, our algorithm can be expanded to control the size fairness among clusters. Experimentally in document clustering, our algorithm outputs more accurate clustering results than the average-link method with the cosine similarity almost without lengthening the running time.