Recommendation as link prediction: a graph kernel-based machine learning approach

Recommender systems have demonstrated commercial success in multiple industries. In digital libraries they have the potential to be used as a support tool for traditional information retrieval functions. Among the major recommendation algorithms, the successful collaborative filtering (CF) methods explore the use of user-item interactions to infer user interests. Based on the finding that transitive user-item associations can alleviate the data sparsity problem in CF, multiple heuristic algorithms were designed to take advantage of the user-item interaction networks with both direct and indirect interactions. However, the use of such graph representation was still limited in learning-based algorithms. In this paper, we propose a graph kernel-based recommendation framework. For each user-item pair, we inspect its associative interaction graph (AIG) that contains the users, items, and interactions n steps away from the pair. We design a novel graph kernel to capture the AIG structures and use them to predict possible user-item interactions. The framework demonstrates improved performance on an online bookstore dataset, especially when a large number of suggestions are needed.