Based on Multiple Attention and User Preferences for Recommendation

Abstract

Recently, deep learning has been widely used in the field of recommendation systems. Sequence recommendation is an important direction of the current recommendation system, which uses the user's behavior sequence information to calculate the possibility of the user interacting with the target item. Recent work embeds the original features into low-dimensional vectors, then uses RNN or Transformer to extract behavior sequence information, and finally inputs to the MLP layer to get the recommended results. These methods do not consider that the items of different positions have different effects on the next item, and the recent item in the real world often have a greater impact on the target item than the previous ones. So we propose MAUPRec models to solve such problems. Each user's historical behavior learns its corresponding weight through attention mechanism, which better reflects the interest of users at different times. In addition, we find that feedback information means user's preference to some extent, so we also introduce feedback information as user's preference representation in the model. We conducted detailed comparison experiments with the very popular models in the industry on different public data sets, and the results showed that our model MAUPRec has good performance.