Semantically-Guided Clustering of Text Documents via Frequent Subgraphs Discovery

In this paper we introduce and analyze two improvements to GDClust [1], a system for document clustering based on the co-occurrence of frequent subgraphs. GDClust (Graph-Based Document Clustering) works with frequent senses derived from the constraints provided by the natural language rather than working with the co-occurrences of frequent keywords commonly used in the vector space model of document clustering. Text documents are transformed to hierarchical document-graphs, and an efficient graph-mining technique is used to find frequent subgraphs. Discovered frequent subgraphs are then utilized to generate accurate sense-based document clusters. In this paper, we introduce two novel mechanisms called the Subgraph-Extension Generator (SEG) and the Maximum Subgraph-Extension Generator (MaxSEG) which directly utilize constraints from the natural language to reduce the number of candidates and the overhead imposed by our first implementation of GDClust.