Offline evaluation of the recommender system initial user interview using estimation of the coverage of content characteristics

Abstract

Recommender systems are used everywhere in everyday life and have been of interest to researchers for many years. In recent years, many new papers have been published in this area, but the problem of cold start remains one of the most significant problems of recommender systems which use historical user data. The simplest solution is to do an initial user interview, but existing solutions for selecting items for it do not use a recommender model. However, they work with it, and they can improve the efficiency by using it for generating interview. This paper proposes adaptive approaches for selecting items for an initial interview based on the model, predictability of users and usefulness of items for a model. We also show how to build an initial interview for a new user by historical information about interactions of other users with items and the corresponding ratings for those interactions only. This is important because the information about content attributes is often missing or unreliable, and the developer is often not aware of specific characteristics that initial interview need to cover. This article and the described research are part of a large project related to the study of the problem of cold start and other problems of recommender systems with a large number of objects and a small number of interactions. As part of this project, we are developing application with which we will continue our research and estimate initial interview with real users. In this paper, we describe the content coverage metrics and demonstrate that our method requires less number of initial user interview screens to cover it than most others. The proposed method requires only three screens to cover 92.9% of the coverage metrics, and only four to cover all, on our dataset which is very sparsed and has 99.99% of missing values.