Abstract
In the era of big data, the acquisition and utilization of information becomes difficult with the skyrocketing amount of data. It is often difficult for ordinary users to find the information or items they need, and personalized recommendation systems can solve this problem well. Currently, recommendation systems increasingly adopt models based on deep learning. The most critical issue in using deep learning for recommendation systems is how to use neural networks to accurately learn user representation vectors and item representation vectors. Many deep learning models used a single vector to represent users, but users' interests were often diverse. Therefore, some researchers consider using multiple vectors to represent user interests, and each interest vector corresponds to a category of items. This method sounds more scientific. However, these models still have problems. Their interpretation of user interests stays at the item level, and does not go deep into the item feature level.
In order to solve this problem, we consider user interests from the perspective of item characteristics, and propose 3M (Multi-task, Multi-interest, Multi-feature) model. The 3M model trains multiple interest vectors for each user and extracts multiple characteristic vectors for each item at the same time, then uses a multi-task learning model to connect the characteristic vectors with the interest vectors and train them to obtain multiple interest scores. According to the multiple interest scores, the user click probability can be obtained. Experiments show that our model performs significantly better than the classic CTR(Click-Through Rate) prediction model on the experimental dataset.