Recommendation with Heterogeneous Information Networks Based on Meta-Graph

Abstract

In order to alleviate data sparsity and cold-start problems of traditional collaborative filtering recommendation algorithm, a meta-based fusion heterogeneous information network recommendation algorithm is adopted in this paper. The algorithm integrates the characteristics of multi-relationship social network and user’s preference degree and adopts a universal representation for different types of data. A meta-graph-based similarity measurement method makes it possible to better capture the semantic relationships between different types of data and a score matrix decomposition method based on multiple meta-graphs is used. Each project and user generates a variety of potential feature matrices based on different meta-graphs. Effectively integrates multiple feature matrices into a unified, final implicit feature matrix. We use each factor of each line of the implicit feature matrix as a neural network. The input node predicts user ratings by optimizing the scoring neural network. Finally, we used the data set provided by the Yelp website to do user rating prediction experiments, which proved the accuracy of this algorithm is 5% higher than the traditional collaborative filtering recommendation algorithm.