Multi-Interest Collection with Mixture of Experts for Recommendation Algorithm

We propose a multi-interest collection recommendation algorithm, based on Mixture of Experts (MoE), to capture and collect users' multi-interest features effectively, also minimize data selection bias in offline training for recall tasks in recommendation systems. This method can learn users' multi-interest attention from the items they interacted with, and provide more appropriate recommendations to improve model performance. Additionally, we introduce rectified negative sampling to differentiate between items that users may dislike and those they explicitly dislike, further optimizing the training process and solving the selection bias problem in recall tasks. Experiments on two real-world datasets demonstrate that the proposed multi-interest collection recommendation algorithm can effectively learn users' multi-interest attention and outperforms other methods in terms of recall rate. Therefore, this method holds potential for various recommendation scenarios, such as educational resource recommendation.