Improving profiles of weakly-engaged users: With applications to recommender systems

The majority of online users do not engage highly with services that are offered via Web. This is a well-known fact and it is also one of the main issues that personalization algorithms try to overcome. A popular way of personalizing an online service is to record users’ actions into user profiles. Weakly-engaged users lead to sparsely populated user profiles, or weak profiles as we name them. Such weak profiles constitute a source of potential increase in user engagement and as a consequence, windfall profits for Internet companies. In this paper, we define the novel problem of enhancing weak profiles in positive space and propose an effective solution based on learning collective embedding space in order to capture a low-dimensional manifold designed to specifically reconstruct sparse user profiles. Our method consistently outperforms baselines consisting of kNN and collective factorization without constraints on user profile. Experiments on two datasets, news and video, from a popular online portal show improvements of up to more than 100% in terms of MAP for extremely weak profiles, and up to around 10% for moderately weak profiles. In order to evaluate the impact of our method on learned latent space embeddings for users and items, we generate recommendations exploiting our user profile constrained approach. The generated recommendations outperform state-of-the-art techniques based on low-rank collective matrix factorization in particular for users that clicked at most four times (78–82% of the total) on the items published by the online portal we consider.