Sur l'utilisation de LDA en RI pair-à-pair

We revisit the problem of defining a peer-to-peer system for Information Retrieval when each peer’s topic-based profile is obtained using Latent Dirichlet Allocation. This method, defined for a centralized collection, provides a rich representation of the topics and of the documents. We describe two ways of using it in a distributed system and analyze their advantages and drawbacks. Then, we illustrate the use of the obtained topic-based profiles within two systems. The first one is unstructured and uses a gossip-based algorithm to obtain dynamic overlays of topically related peers. This requires defining a similarity between profiles. The second one uses super-peers and maintains a topic-based index of the peers, which is recorded in a distributed Hash table. The keys are derived from the topic-based profiles. MOTS-CLÉS : Recherche d’information, systèmes pair-à-pair, Allocation Latente de Dirichlet (LDA).