Experiments on Trust Prediction Based on Reputation Features

Abstract

In Social Network services, Trust Prediction is the problem of guessing if one user, the truster, will trust another, the trustee. The information used for this prediction can be extracted from various sources, such as the user attributes or the Web of Trust (WoT). The WoT contains the user explicit trust declarations, allowing to know who already has expressed trust or distrust on the trustee. Collecting these trust statements we obtain the reputation features of the target trustee. Trust Prediction is then a classification problem that can be solved by machine learning approaches. However, most learning algorithms work on fixed diamension feature spaces, while reputation features are of arbitrary size for each social agent. Using probabilistic descriptors of the reputation feature vector we obtain a fixed dimension representation that is amenable to machine learning aproaches. We demonstrate the Trust Prediction approach on the WoT obtained from the Wikipedia vote network database.