An Effective Implicit Multi-interest Interaction Network for Recommendation

Data features in real industrial recommendation scenarios are high-dimensional, diverse and sparse. Rich feature interaction can improve the model effect and bring practical benefits. Factorization machines (FMs) can perform explicit second-order feature interactions, while deep neural networks (DNNs) can perform implicit non-linear feature interactions. A series of models integrating FMs and DNNs are used to perform diverse feature interactions. However, most of the previous work performed feature interaction without considering the diverse interests of users. In reality, users often have multiple preferences and interests, which are implicitly included in the features and need to be effectively extracted. In this paper, we propose an implicit multiple interest network (IMIN), taking into account the importance of interest. Specifically, the model constructs the implicit multiple interests of the user and the item through the implicit multi-interest layer, and realizes the interest alignment between the user and the item through the interest alignment layer. We further use the interest interaction and aggregation layer to construct rich interest feature interactions. In addition, we introduce an auxiliary loss in the model optimization part to ensure the difference of interest. We conducted comprehensive and rich experiments on three real-world data sets. Experimental results show that IMIN performs better than other competitive models, which proves the effectiveness of the model.