Information retrieval model based on graph comparison

We propose a new method for Information Retrieval (IR) based on graph vertices comparison. The main goal of this method is to enhance the core IR-process of finding relevant documents in a collection of documents according to a user’s needs. The method we propose is based on graph comparison and involves recursive computation of similarity. In the framework this approach, documents, queries and indexing terms are viewed as vertices of a bipartite graph where edges go from a document or a query – first node typeto an indexing term – second node type-. Edges reflect the link that exists between documents or queries on the one hand and indexing terms on the other hand. In our model, graph edge settings reflect the tf-ifd paradigm. The proposed similarity measure instantiates and extends this principle, stipulating that the resemblance of two items or objects can be computed using the similarities of the items to which they are related. Our method also takes into account the concept of similarity propagation over graph edges. Experiments conducted using four small sized IR test collections (TREC 2004 Novelty Track, CISI, Cranfield & Medline) demonstrate the effectiveness of our approach and its feasibility as long as the graph size does not exceed a few thousand nodes. The experiment’s results show that our method outperforms the vector-based model. Our method actually highly outperforms the vector-based cosine model, sometimes by more than doubling the precision, up to the top sixty returned documents. The computational complexity issue is resituated in the context of MAC-FAC approaches – many are called but few are chosen. More precisely, we suggest that our method can be successfully used as a FAC stage combined with a fast and computationally cheap method used as a MAC stage.