Intelligent Recommender System with Online Generation of Model-Based Initial Interviews for Poi Recommendations

Abstract

The problem of cold start remains important. There are many solutions, including methods that use an interview to identify preliminary preferences of a user. However, existing solutions are focused on evaluating the informativeness of items themselves rather than their usefulness for a recommender model. Moreover, approaches are mainly evaluated with offline experiments on historical data but not on actual data of active experiments with users. Such evaluations are not reliable because the compared methods and recommender models can not influence the composition and order of recommendations. Moreover, papers insufficiently illustrate intelligent systems where an interview process is integrated into the main recommender process. In this paper, we consider the user cold start problem and try to solve the problems described above by proposing an intelligent system that integrates both processes, and two model-based methods of options selection. We attracted 160 unique users using an advertising campaign and applied this system to conduct an active experiment. It showed that our method with the VAECF model reached an average NDCG@10 metric of 0.467 versus 0.253 for the entropy-based method. It means that recommender models with our method are better at ranking items for a small number of ratings than others.