Recommendation as link prediction in bipartite graphs: A graph kernel-based machine learning approach

Recommender systems have been widely adopted in online applications to suggest products, services, and contents to potential users. Collaborative filtering (CF) is a successful recommendation paradigm that employs transaction information to enrich user and item features for recommendation. By mapping transactions to a bipartite user-item interaction graph, a recommendation problem is converted into a link prediction problem, where the graph structure captures subtle information on relations between users and items. To take advantage of the structure of this graph, we propose a kernel-based recommendation approach and design a novel graph kernel that inspects customers and items (indirectly) related to the focal user-item pair as its context to predict whether there may be a link. In the graph kernel, we generate random walk paths starting from a focal user-item pair and define similarities between user-item pairs based on the random walk paths. We prove the validity of the kernel and apply it in a one-class classification framework for recommendation. We evaluate the proposed approach with three real-world datasets. Our proposed method outperforms state-of-the-art benchmark algorithms, particularly when recommending a large number of items. The experiments show the necessity of capturing user-item graph structure in recommendation.