AAU2V: User and Item Embedding Model with Attention Mechanism Considering Auxiliary Information

With the recent popularity of e-commerce websites, analyzing user characteristics from purchase history data has become an important challenge for companies. User2Vec is an effective tool to achieve this task, which is a simple neural network model to describe the characteristics using embeddings. However, User2Vec does not utilize auxiliary information, and the relevance between users and items are not explicitly considered into model learning process. To capture the characteristics more accurately, it is effective to consider auxiliary information, and to evaluate the relevance with items for each user. In this study, we propose a model called Attentive and Auxiliary-Informative User2Vec, which learns both user-specific and item-specific embeddings and auxiliary information embeddings simultaneously and utilizes the attention mechanism to associate users and items. By our method, it has become possible to perform characteristic analysis associating users and items, considering various auxiliary information in addition to purchase history data. Additionally, characteristic analysis that excludes the influence of auxiliary information can be performed. To demonstrate the effectiveness of our method, we perform an evaluation experiment with artificial data assuming purchase history. Furthermore, we apply our method to actual user review rating data for movies and show a case study of user characteristic analysis. Implementations are available at https://github.com/tishii2479/aau2v.