Contrastive Learning Based Disentangled Negative Sampling For Collaborative Filtering

Nowadays,recommender systems have been recognized as one of the most powerful tools to solve the information overload problem.As one of the most widely used techniques in recommender systems,collaborative filtering aims to mine the real preferences of a user from his/her historical behavior in order to recommend the next item with which the user is willing to interact.In view of the difficulty of data collection,more and more collaborative filtering models choose implicit feedback as their input by default.In implicit feedback,items that a user has interacted with are usually regarded as positive samples,while negative samples need to be sampled from items that the user has not interacted with via a certain strategy.The process of generating negative samples is called negative sampling.Negative sampling is essential for the training of collaborative filtering models,which affects the accuracy of the recommendation results directly.In recent years,more and more studies in collaborative filtering have focused on designing efficient negative sampling methods.Although existing negative sampling methods have achieved some promising results,they still suffer from a common drawback: all of them consider an item as an indivisible whole and ignore the fact that a user is normally driven by some specific factors of an item,to interact with it.In order to address this limitation,this paper proposes to consider the fine-grained factors of items to select the best negative samples.However,this is non-trivial due to the following challenges: first,how to disentangle specific factors of an item;second,how to effectively measure the quality of negative samples;third,how to ensure the credibility of the disentanglement.Aiming at the above three challenges,a disentangled negative sampling method is proposed in this paper,which mainly includes the following three parts:(1)a hierarchical gating network,which is designed to disentangle factors of positive samples and candidate negative samples with user-specific gated linear units and factor-specific gated linear units,respectively,to guide the subsequent negative sampling process.(2)an evaluation strategy,which is proposed to measure the quality of negative samples with a hyperparameter to balance the impact of different factors on the quality of negative samples and a linearly increasing mechanism to dynamically select the hyperparameter.In addition,based on theoretical analysis,this paper greatly simplifies the intermediate process of the evaluating strategy to ensure its high efficiency and easy implementation.(3)four contrastive learning tasks designed to supervise the disentanglement from different aspects,which help to ensure the credibility of the disentanglement.Finally,this paper conducts extensive experiments on five public datasets.The experimental results show that the negative sampling method proposed in this paper generally outperforms the latest works on two basic models of collaborative filtering.