User Preference Quantity versus Recommendation Performance: A Preliminary Study

摘要

Recommender system has become one of the most promising techniques in the era of big data. It aims to help users to quickly find the valuable information from the massive data. Many recommendation approaches have been proposed in recent years. Currently, a majority of researchers still pay attention on designing more effective and efficient methods, and they usually put all the user data into model training without considering the quantity of individual preferences. However, we argue that not all user preferences contribute to the adopted models, especially for active users who generate plentiful preferences. We claim that some representative preferences contain enough information to profile users and thus are enough to get sound recommendations. Particularly, we attempt to explore the relationship between the quantity of user preferences and recommendation performance, and focus on the representative preference selection. In order to achieve this, we first elaborate the recommendation performance tendency on different sub datasets splitted by the quantity of user preferences. We consider both the rating prediction and the top-N item recommendation tasks. Furthermore, we propose several preference selection strategies to choose the most representative preferences. Finally, we conduct several series of experiments on a large public data set and experimentally conclude that part of user preferences are able to generate desirable recommendations at a rather lower computational cost.