A picture is worth a thousand words: Introducing visual similarity into recommendation

Recent recommender systems work well in terms of prediction accuracy, making use of a variety of features, such as users' personal information, purchasing history, browsing history and comments. However, traditional recommendation models have not made full use of item information and met difficulties with cold-start problems. On the other hands, visual information on item images is one of the most basic and informative features of the item, which has not been well-studied and applied in recommendation yet. In this paper, we introduce “visual similarity” between different items into recommendation, which measures the probability between items that are similar in terms of visual effect or “styles”. Observations on real e-commercial site data show that users tend to buy similar items, or items with similar “style”, indicating that visual information can be considered as a reliable feature in recommending process. Furthermore, a new matrix supplement approach is proposed to integrate item-item similarity matrix and traditional user-item matrix for collaborative filtering. Finally, a novel recommendation model is proposed which leverages visual similarity to collaborative filtering. Experiments on e-commercial website data shows that the proposed approaches result in superior performance compared with traditional recommendation algorithms, including Baseline Predictor, KNN (k-nearest-neighbors) and SVD (Singular Value Decomposition). Results also verifies that visual information does help relieve the “cold-start” problem in recommendation.