Predictive analytics has hit the mainstream, thanks to the emergence of many day-to-day consumer applications such as Pandora, Netflix, Yelp, and Epinions, all of which contain some variation of a recommendation engine. These engines (or systems) recommend items about which the user is more likely to have a favorable opinion. A special class of recommendation systems is the one in which a user likes (trusts) a reviewer. These recommendation systems then use the network of trust to predict what items the user may like. Such systems attain reasonable rating predictions, albeit at a cost of high computational complexity.

In this paper, the authors seek to reduce the computational complexity of trust-aware recommendation systems (TARS). They hypothesize that we can perhaps reach a reasonable level of prediction accuracy by focusing on reputable recommenders alone. They define reputation by the distribution of the reviewer’s recommendations being liked or trusted. The authors conclude that while the hypothesis leads to a much more efficient algorithm, the accuracy of recommendations is compromised greatly, thereby requiring the summary rejection of the hypothesis. The results are supported by extensive experiments on Epinions datasets consisting of many thousands of users and almost half a million trust relations.

For research such as this, in which the authors try an alternative and report that the alternative method does not work, the value comes from our enriched understanding of the system model and the relationships between various components. Trust-aware recommendation systems are part of the family of recommendation systems that do not use the content of the item in the prediction. Rather, they use the similarity between the users and their mutual trust or liking factor to guide the ratings. That observation by itself can be interpreted in two contrasting ways: (1) a user who is trusted (directly or indirectly) by many others can generate more useful ratings, and (2) the reputation of a recommender plays no role for a user who does not like or trust the recommender. From this paper, it appears that the second interpretation is the more correct one, at least for the scenarios and datasets considered in this work. However, it may be interesting to consider the problem of the characterization of system models in which the trust “carries over,” and those in which it doesn’t.