GVC: a graph-based Information Retrieval Mode

GVC is a new information retrieval model that is based on Graph Vertices Comparison (GVC). It implements a new similarity measure to compare documents and users' queries based on graph matching. In this model, graphs are composed of two types of nodes. Documents, queries and indexing terms are viewed as vertices of this bipartite graph where each edge goes from a document or a query –first type of nodesto an indexing term – second type of nodes-. Edges reflect the relationship that exists between documents or queries on the one hand and indexing terms on the other hand; they are set according to the tf.idf principal. Our method implements similarity propagation over graph edges using an iterative process. We evaluate the model using 4 different collections (TREC 2004 Novelty Track, CISI, Cranfield and Medline). We show that considering precision at 5 documents, GVC outperforms Okapi model from 9% to 62%, depending on the collections.