Mining Item Popularity for Recommender Systems

Abstract

Listen

Translate article

Recommender systems can predict individual user's preference individual rating on items by examining similar items' popularity or similar users' taste. However, these systems cannot tell item's long-term popularity. In this paper, we propose an algorithm for predicting item's long-term popularity through influential users, whose opinions or preferences strongly affect that of the other users. Consequently, choices made by certain influential users have the tendency to steer subsequent choices of other users, hence setting the popularity trend of the product. In our algorithm, specifically, through judicious segmentation of the rating stream of an item, we are able to determine whether it is popular, and whether that is the consequence of certain influential users' ratings. Next, by postulating that similar items share similar influential users, and that users rate similar items consistently, we are able to predict the influential users for a new item, and hence the popularity trend of the new item. Finally, we conduct extensive experiments on large movie rating datasets to show the effectiveness of our algorithm.