Using Trust Prediction Mechanism to Support Trust-Based Recommendations

Abstract

As the emergence and rapid growth of Internet, it is more convenient for users to search for items (including information, products or services) via Internet. However, information explosion makes people difficultly assimilate large amounts of information. A solution to overcome this well-know information overload problem is to adopt a recommendation approach. The recommendation approach makes users searching items faster by suggesting the products they may like or be interested in and therefore could result in more consumption or usage of e-commerce websites for the enterprises. Recently, with the development of Web 2.0, the trust-based recommendation approach is the emerging recommendation approach. Based on the analysis of trust relationships, the trust-based recommendation approach finds the neighbors with highest trust values for an active user and then makes rating predictions based on their opinions. However, the trust-based recommendation approach suffers from the problem of trust network sparsity. The popular way is to adopt the trust propagation method to expand the trust network. Even so, the trust propagation method still has the limitation that two users have no chance to form a propagated trust relationship if there is no path in the trust network between them. In response, our proposed technique aims to develop a trust prediction mechanism to predict the trust value of each non-appeared trust relationship in the trust network. Specifically, we extract various structural predictors from a training dataset as the input variables for the machine learning algorithm to build the trust prediction model. We collect the evaluation dataset from Epinions.com and implement the benchmarks, i.e., the traditional collaborative filtering approach and the traditional trust-based recommendation approach for comparison. The experimental results demonstrate that our proposed technique can achieve higher coverage than that of the benchmarks. Moreover, our proposed technique outperforms the benchmarks on prediction accuracy for those items that can also be predicted by the benchmark techniques.