Visual-SLIM: Integrated sparse linear model with visual features for personalized recommendation

Abstract

With the increasingly complexity and dynamically of information, recommendation system has been a key solution to alleviate the problem of information overloaded. Most recommender system models users’ preferences toward items based on users’ historical implicit feedback with item (e.g., product purchase history, browsing logs, etc.). They typically make recommendation for a target user based on her profiles only (e.g., the user’s previous activities), ignoring the existence of other valuable information on items such as the visual features of images corresponding to the items. As a downside, it may limit the performance of recommender systems to some extent. This paper proposes a joint prediction model (visual-SLIM) which extends SLIM method with visual information to predict people’s preference. The proposed approach automatically generates the missing items scores for a target user by aggregating observed user-item interaction matrix and learning linear regression model with items’ visual information. It would not only improve performance of the model, but also do help to better analysis of the effects of visual information on user’s opinions. Extensive experiments conducted on the real-world dataset of the Amazon have demonstrated the effectiveness of our proposed model.