Trust-aware Similarity Metrics

Similarity metrics play a key role in case-based reasoning: an effective retrieval step is a premise for a fruitful reuse of past solutions. In recommendation systems based on collaborative filtering the similarity assessment involves the user profiles. User profiles are very sensitive to the quality of elicited preferences while user similarity doesn t address the issue of quality assessment. We argue that the notion of trust can fruitfully improve the step of similarity assessment. Trust can be conceived as rating persons instead of goods. Some properties of trust, like propagation, allows to overcome known drawbacks of collaborative filtering. We discuss how sparseness of data very often affects the computability of user similarity and we show how trust metrics can overcome this restriction. An empirical evaluation on a real world dataset allow us to argue that trust metrics don t suffer the problem of sparseness and can provide a meaningful improvement in theassessment of user similarity. Moreover we prove that the acquisition of trust values, in contrast with rating values, enables a better trade-off between elicitation effort and similarity accuracy