Enhancing new user cold-start based on decision trees active learning by using past warm-users predictions

The cold-start is the situation in which the recommender system has no or not enough information about the (new) users/items i.e. their ratings/feedback; hence, the recommendations are not well performed. This issue is commonly encountered in techniques based on collaborative filtering, as they mainly rely on the feedback
of users and items. This paper focuses on the active learning techniques based on collaborative filtering and using decision trees to address the new user cold-start in recommender systems. These techniques propose to interact with new users by asking them to rate sequentially a few items while the system tries to detect their
preferences. Their main goal is to find out the best recognizable items for the new user in order to get very informative user’s feedback. Compared to current state of the art, the presented approach takes into account the users’ ratings predictions in addition to the available users’ ratings. The experimentation shows that our approach achieves better performance in terms of precision and
limits the number of questions asked to the users. This is specially interesting in datasets with a low number of users.