Evaluating Session-Based Recommendation Approaches on Datasets from Different Domains

摘要

Recommending relevant items of interest for user is the main purpose of recommendation system based on long-term user profiles. However, personal data privacy is becoming a big challenge recently. Thus, recommendation system needs to reduce the dependence on user profiles while still keeping high accuracy on recommendation. Session-based recommendation is a recent proposed approach for recommendation system to overcome the issue of user profiles dependency. The relevance of problem is quite high and has triggered interest among researchers in observing activities of users. It increased a number of proposals for session-based recommendation algorithms that aiming to make prediction of next actions. In this paper, we would like to compare the performance of such algorithms by using various datasets and evaluation metrics. The most recent deep learning approach named GRU4REC and simpler methods based are included in our comparison. Six real-world datasets from three different domains are included in our experiment. Our experiments reveal that in case of numerous unpopular items dataset, GRU4REC's performance is low. However, its performance is significantly increased after applying our proposed sampling method. Therefore, our obtained results suggested that there is still room for improving deep learning session-based recommendation algorithms.