Graph-Based Diffusion Method for Top-N Recommendation

Data that may be used for personalised recommendation purposes can intuitively be modelled as a graph. Users can be linked to item data; item data may be linked to item data. With such a model, the task of recommending new items to users or making new connections between items can be undertaken by algorithms designed to establish the relatedness between vertices in a graph. One such class of algorithm is based on the random walk, whereby a sequence of connected vertices are visited based on an underlying probability distribution and a determination of vertex relatedness established. A diffusion kernel encodes such a process. This paper demonstrates several diffusion kernel approaches on a graph composed of user-item and item-item relationships. The approach presented in this paper, RecWalk*, consists of a user-item bipartite combined with an item-item graph on which several diffusion kernels are applied and evaluated in terms of top-n recommendation. We conduct experiments on several datasets of the RecWalk* model using combinations of different item-item graph models and personalised diffusion kernels. We compare accuracy with some non-item recommender methods. We show that diffusion kernel approaches match or outperform state-of-the-art recommender approaches.