Apriori approach to graph-based clustering of text documents

Abstract

This thesis report introduces a new technique of document clustering based on frequent senses. The developed system, named GDClust (Graph-Based Document Clustering) [1], works with frequent senses rather than dealing with frequent keywords used in traditional text mining techniques. GDClust presents text documents as hierarchical document-graphs and uses an Apriori paradigm to find the frequent subgraphs, which reflect frequent senses. Discovered frequent subgraphs are then utilized to generate accurate sense-based document clusters. We propose a novel multilevel Gaussian minimum support strategy for candidate subgraph generation. Additionally, we introduce another novel mechanism called Subgraph-Extension mining that reduces the number of candidates and overhead imposed by the traditional Apriori-based candidate generation mechanism. GDClust utilizes an English language thesaurus (WordNet [2]) to construct document-graphs and exploits graph-based data mining techniques for sense discovery and clustering. It is an automated system and requires minimal human interaction for the clustering purpose.