A Social Recommendation Based on Metric Learning and Network Embedding

Traditional recommender systems usually face data sparseness and cold-start problems. To alleviate these problems, social recommendations use social information between users to improve the quality of recommendations. At present, most social recommendation is based on matrix factorization, but with the combination of recommendation and metric learning, the user's preference is reflected by distance, which has proven to be an effective recommendation method. However, these social recommendation methods usually use simple similarity information as users' social regularization, which cannot analyze users' social relationship characteristics well. In order to overcome the shortcoming of social recommendation, we propose a new model for which combines network embedding and metric learning (SRMN) in this paper. SRMN believes that users and items have interactive behaviors in social space. The interaction can be regarded as the user's preferences for the item in the social space, and then combines with the use of metrics to reflect the user's preferences. Users' preferences for items are determined by both ratings and social information. Finally, the effects of different network embedding methods on the model are compared. Experimental results on three real data sets show that our method is effective in improving the accuracy of score prediction.