MetricRec: Metric Learning for Cold-Start Recommendations.

Making recommendations for new users is a challenging task of cold-start recommendations due to the absence of historical ratings. When the attributes of users are available, such as age, occupation and gender, then new users’ preference can be inferred. Inspired by the user based collaborative filtering in warm-start scenario, we propose using the similarity on attributes to conduct recommendations for new users. Two basic similarity metrics, cosine and Jaccard, are evaluated for cold-start. We also propose a novel recommendation model, MetricRec, that learns an interest-derived metric such that the users with similar interests are close to each other in the attribute space. As the MetricRec’s feasible area is conic, we propose an efficient Interior-point Stochastic Gradient Descent (ISGD) method to optimize it. During the optimizing process, the metric is always guaranteed in the feasible area. Owing to the stochastic strategy, ISGD possesses scalability. Finally, the proposed models are assessed on two movie datasets, Movielens-100K and Movielens-1M. Experimental results demonstrate that MetricRec can effectively learn the interest-derived metric that is superior to cosine and Jaccard, and solve the cold-start problem effectively.