Leveraging side information as adjusting embedding to improve user representation for recommendations

Embedding is the cornerstone of recommendation system, and the embedding of users or items is directly related to the accuracy of recommendation. However, many recommendation methods directly use the ID of the user or item as the source of embedding. The advantage of doing so is simple and direct, and the fatal defect is that the meaning of embedding is single, rigid and lack of connotation. In this paper, we propose leveraging Side Information as Adjusting Embedding to improve user representation for recommendation. Our work is to add the attribute embedding of an item to the users initial embedding to create a high-order embedding when the user evaluates an item. In this way, the potential preferences of users can be mined more deeply. We add the main attribute embedding of the item and the users embedding layer by layer to adjust the users embedding. By constantly adjusting the size and direction of the user embedding vector, the user embedding becomes a customized high-level user embedding for different items. In other words, when a user evaluates different items, the user embedding is not fixed, but adapted to the item after adjustment. We do a lot of experiments on three real datasets, and prove that adjusting embedding can improve the ac curacy of the algorithm. Finally, it should be noted that our proposed adjusting embedding representation method can be applied to a variety of interaction processes or graph structures, including biomedical science, transportation, social networks, etc., in addition to a wide variety of recommendation situations.