A Graph Kernel Based Item Similarity Measure for Top-N Recommendation

The item neighborhood-based graph kernel (INGK) has recently been proposed to compute item similarity on the Linked Open Data (LOD) graph and then produce top-N recommendations (the similarity measure is abbreviated as INGK-LOD). This paper explores how to use the graph kernel to compute item similarity on the basis of user-item ratings. We transform the user-item ratings matrix into an undirected graph called a user-item ratings graph, and define the graph kernel based on the graph, which can be used to compute item similarity (the similarity measure is abbreviated as INGK-UIR). We applied INGK-UIR, INGK-LOD and two baseline similarity measures, Cosine and Pearson correlation coefficient, to top-N recommendation, and conducted experimental evaluation of recommendation accuracy using the MovieLens 1M benchmark dataset. The results show that our INGK-UIR significantly outperforms INGK-LOD and the two baseline measures in terms of precision and recall.