Cold Start Recommendations: A Non-negative Matrix Factorization Approach

Recommender systems suggest to users items that they might like (e.g., news articles, songs, movies) and, in doing so, they help users deal with information overload and enjoy a personalized experience. One of the main problems of these systems is the cold-start, i.e., when a new item or user is introduced in the system and no past information is available, no effective recommendation can be produced. The cold-start is a very common problem in practice: modern online platforms have hundreds of new items and millions of visits from logged-out users every day. Despite the importance of this problem not many solutions have been proposed in the literature. We contribute to closing this gap by studying whether it could be overcome without sacrificing performance. We do so by exploiting two aspects: the combination of the content and collaborative information, and the users’ location. First, we combine the properties of the items and past user preferences by introducing Joint NMF, a novel recommender system based on Non-negative Matrix Factorization (NMF). We also propose a variation, Joint NMF with Graph Regularization, which accounts for the local geometric structure of the data. We present two training strategies, based on multiplicative update rules and alternating least squares. The experimental results show that the two proposed methods outperform the existing content-based recommenders, which are often used for recommending newly introduced items (item cold-start). Second, to exploit user location, we test whether specific interests are linked to specific locations. To this end, we study the geography of user engagement in online news platforms on a random sample of 200K news articles and the corresponding 41M comments posted on the Yahoo! News website for the USA. We find that time zones play an important role on the user engagement: users from the same time zone preferentially engage with each other on the same articles about the same topics. Thus, we propose a time-zone-aware recommender that suggests the most popular articles not in the whole USA but in the users’ time zones, when no past information of the user preferences is available (user cold-start). We find that suggesting recommendations that are specific to time zones improves the recommendation accuracy by a factor of one point five.