Research On Context Enhanced Sequential Recommendation Technology

In recent years,sequential recommendation has gradually become a hot research topic in the area of recommender systems.Most methods focus on modeling the user’s historical behavior sequence,ignoring the contextual information and thus failing to fully exploit the potential of user data.In fact,contextual information is of great importance in sequential recommendation.The contextual information contains the situational information and collaborative information of the interaction,and enhancing contextual modeling can help the model understand users’ real intentions,thus giving more personalized and relevant recommendation results.Meanwhile,items that users have interacted with in the past account for only a very small proportion of the entire candidate set,and the use of contextual information can alleviate the problem of data sparsity.With the emergence of dual learning,decoupled representation learning,contrastive learning,and other techniques,recommender systems that integrate these emerging technologies have greater potential for utilizing context information.This paper studies and explores the incorporation of temporal and global context into sequential recommendation task.The main contributions are as follows:For the temporal context information in sequential recommendation,this work mainly focuses on exploring the duality between sequential recommendation task and time prediction task.Time prediction task can be used as an auxiliary task for sequential recommendation.A few previous works have simultaneously learned these two objectives using the parameter-sharing multi-task learning paradigm,but they all ignored the probabilistic connection between the two tasks.Therefore,this paper proposes a dual learning framework to jointly model sequence recommendation and time prediction tasks and incorporates the probabilistic duality between them during the training stage,using a dual regularization term to constrain the joint training process of the two models.In addition,suitable base models are designed for both tasks.Time information is integrated into sequence recommendation task to model the user’s short-term and long-term interests,and a discrete time slice prediction model which simulates temporal point process is proposed for calculating the dual regularization term.Finally,experiments on two public datasets demonstrate the effectiveness of the proposed dual learning framework in sequential recommendation scenarios.For the global context information in sequential recommendation,this work mainly focuses on mining user fine-grained intent from the global context.The goal is to model the main intent behind user behavior and give recommendation that fit the user’s current interests.A few previous works have considered the role of global information,but they treated user intent and item factors as a whole without fine-grained modeling.Therefore,this paper proposes a decoupled graph neural network with global context enhancement for session recommendation.The item embedding representation is divided into different blocks,each corresponding to a factor,and the global graph and local graph are used to learn factor-aware session representations.Furthermore,intent-aware contrastive learning is used to help decouple different factors,and the contrast between global session representation and local session representation enhances the robustness and generalization ability of the model.Finally,experiments on two public datasets demonstrate the effectiveness of the proposed method in session-based recommendation scenarios.In summary,from the perspective of context enhancement,this paper conducts indepth research on temporal context and global context,and verifies the effectiveness of two proposed methods on multiple public datasets.