Succinct and Informative Cluster Descriptions for Document Repositories

Large document repositories need to be organized, summarized and labeled in order to be used effectively. Previous clustering studies focused on organizing, and paid little attention to producing cluster labels. Without informative labels, users need to browse many documents to get a sense of what the clusters contain. Human labeling of clusters is not viable when clustering is performed on demand or for very few users. It is desirable to automatically generate informative cluster descriptions (CDs), in order to give users a high-level sense about the clusters, and to help repository managers to produce the final cluster labels. This paper studies CDs in the form of small term sets for document clusters, and investigates how to measure the quality or fidelity of CDs and how to construct high quality CDs. We propose to use a CD-based classification for simulating how to interpret CDs, and to use the F-score of the classification to measure CD quality. Since directly searching good CDs using F-score is too expensive, we consider a surrogate quality measure, the CDD measure, which combines three factors: coverage, disjointness, and diversity. We give a search strategy for constructing CDs, namely a layer-based replacement method called PagodaCD. Experimental results show that the algorithm is efficient and can produce high quality CDs. CDs produced by PagodaCD also exhibit a monotone quality behavior.