Criterion functions for document clustering
Fast and high-quality document clustering algorithms play an important role in providing intuitive navigation and browsing mechanisms by organizing large amounts of information into a small number of meaningful clusters. In this thesis, we focus on a class of clustering algorithms that treat the clustering problem as an optimization process which seeks to maximize or minimize a particular clustering criterion function defined over the entire clustering solution.
In this thesis, we present a comprehensive study on desirable characteristics and feasibility of various criterion functions under different clustering requirements raised by real world applications. In particular, we focus on seven global criterion functions for clustering large documents datasets, three of which are introduced by us.
The first part of this thesis consists of a detailed experimental evaluation using 15 different datasets and three different partitional clustering approaches, followed by a theoretical analysis of the characteristics of the various criterion functions. Our analysis shows that the criterion functions that are more robust to the difference of cluster tightness and produce more balanced clusters tend to perform well. Our three new criterion functions are among the ones achieving the best overall results.
We further discuss how the various criterion functions perform to produce hierarchical and soft clustering solutions. We present a comprehensive experimental evaluation of six partitional and nine agglomerative hierarchical clustering methods using twelve datasets. A new class of agglomerative algorithms, constrained agglomerative algorithm, is also proposed and achieves the best results. We also focus on four criterion functions, derive their soft-clustering extensions, present a comprehensive experimental evaluation involving twelve different datasets, and analyze their overall characteristics. Finally, we extend the various criterion functions to incorporate prior knowledge on natural topics existing in datasets. Specifically, we define the problem of topic-driven clustering, which organizes a document collection according to a given set of topics. We propose three topic-driven schemes that consider the similarity between documents and topics and the relationship among documents themselves simultaneously. Our experimental results show that the proposed topic-driven schemes are efficient and effective with topic prototypes of different levels of specificity.