Affinity Inference with Application to Recommender Systems
Nan Li, Longin Jan Latecki · 2015
We propose a novel framework for affinity inference and apply it to recommender systems. Given a set of objects and affinities between some pairs of them, we infer the relative value of the unknown affinities based on the transitive property of the affinity relationship. An inference chain is defined as any possible transitive inference process between two objects. In general, there are an infinite number of distinct inference chains between two objects. Each inference chain reveals these two objects are affinitive with a certain confidence, which depends on the individual affinities of the links and the length of the chain. We quantify and aggregate all these confidences as the relative value of the unknown affinity with an efficient method. We formulate collaborative filtering recommendation as an affinity inference problem. The given ratings are transformed into affinities between abstract objects of users and item-rating pairs. The unknown ratings are predicted based on the inferred affinities. The recommendations are made according to both the predicted ratings and the prediction confidences, which are also derived from the inferred affinities. Our approach achieves good prediction accuracy and significantly alleviates the cold-start problem on the standard MovieLens dataset. Moreover, experimental results show that our approach can effectively incorporate extra information to improve the predictions and recommendations.