Paolo Avesani

Today Paolo Massa will defend his PhD Thesis “Trust-aware Decentralized Recommender Systems”.
The event takes place in Room 7, 3:00pm, at the Department of Information and Communication Technologies, University of Trento. The abstract of the PhD dissertation:

Recommender Systems are widely used online to suggest items that the active user may like, such as movies, songs, etc. The most used technique, Collaborative Filtering, works by finding users similar to the active user, based on their provided ratings on items, and recommending to her items appreciated by these similar users. Current Recommender Systems suffer some important weaknesses, mainly data sparsity, cold start and vulnerability to attacks. In order to overcome these weaknesses, we propose Trust-aware Decentralized Recommender Systems in which user similarity weights are replaced by predicted trust scores. Trust Metrics are used to predict the trust scores of users unknown to active user exploiting trust propagation over the explicit trust network. We analyse the differences between global and local trust metrics and we propose a time-efficient local trust metric. The use of local trust metrics makes our proposal particularly suitable for decentralized environments such as the Semantic Web and peer-to-peer networks. We validate our proposal against a large, real world dataset derived from Epinions.com Web community. We have found that this dataset (as many other real-world datasets) exhibits different characteristics with respect to the dataset most used in Recommender Systems research. For example, a trivial algorithm returning an average rating as a prediction for any item achieves a smaller error than the state of the art Collaborative Filtering technique. Nevertheless, our Trust-aware Decentralized Recommender System powered by the local trust metric we have defined outperforms all the other tested techniques. This is particularly true when the analysis concentrates on challenging and significant portions of the data such as controversial items and cold start users.