An experimental comparison of label selection methods for hierarchical document clusters

The focus of this paper is on the evaluation of sixteen labeling methods for hierarchical document clusters over five datasets. All of the methods are independent from clustering algorithms, applied subsequently to the dendrogram construction and based on probabilistic dependence relations among labels and clusters. To reach a fair comparison as well as a standard benchmark, we rewrote and presented the labeling methods in a similar notation. The experimental results were analyzed through a proposed evaluation methodology based on: (i) data standardization before applying the cluster labeling methods and over the labeling results; (ii) a particular information retrieval process, using the obtained labels and their hierarchical relations to construct the search queries; (iii) evaluation of the retrieval process through precision, recall and F measure; (iv) variance analysis of the retrieval results to better understanding the differences among the labeling methods; and, (v) the emulation of a human judgment through the analysis of a topic observed coherence measure - normalized Pointwise Mutual Information (PMI). Applying the methodology, we are able to highlight the advantages of certain methods: to capture specific information; for a better document hierarchy comprehension at different levels of granularity; and, to capture the most coherent labels through the label selections. Finally, the experimental results demonstrated that the label selection methods which hardly consider hierarchical relations had the best results.