Time Lag Aware And Multi-Type Behaviors Sequential Recommendation Algorithm Research

Sequential recommendation aims to capture the time-sensitive preference(or needs)of users by modeling the sequential dependency between their behaviors based on their historical interaction data(e.g.,click,purchase,and check-in)that are collected sequentially by online platforms such as e-commerce websites and location-based networks,then generate a ranked list of items with which users most likely interact in the future.In recent years,sequential recommendation has become a hot topic in the research of personalized recommendation systems.Although a lot of sequential recommendation works have been put forward,the existing work still faces the following two challenges: the first challenge is that the existing sequential recommendation work treats the user preference as a flat distribution over sessions without distinguishing its global stability and local fluctuation,and uses a scalar based weighted schema to fuse the long-term and short-term preferences of users,and ignores the adjacent interaction.which however is too coarse to learn an expressive embedding of current preference.This is because users have long-term preference that will change slowly over time and short-term preference that are volatile and may be completely different from long-term preference,and the impact of these two preferences on users' current preference is dynamic,personalized and finegrained.The second challenge is that the existing sequential recommendation work does not make full use of the behavior type information to help learn the personalized preference of users and items,and ignores the order dependence of users' multi-type interactions,which leads to the limited performance improvement of the model.This is because in the real process of user interaction,different types of behaviors imply that there are differences in the aspects(such as brand,price,etc.)that users pay attention to(or reflect the goods).Moreover,there is a sequence dependence between different types of interactions,which means that different types of user preference prediction should be sequential and related,rather than independent.To address the first challenge,we propose a novel model called Hierarchical self-Attention network for Time lag aware Sequential Recommendation(HATSRec),which integrates a hierarchical modeling of user preference and a time lag sensitive fine-grained fusion of the long-term and short-term preferences.Particularly,we propose a hierarchical self-attention network to learn users' long-term preference with global stability and short-term preferences with local dynamics,as well as a neural time gate by which HATSRec can adaptively regulate the contributions of the long-term and shot-term preferences to the learning of a user's current preference at the fine-grained aspect.The extensive experiments conducted on real datasets verify the effectiveness of HATSRec on sequential recommendation.To address the second challenge,we propose a Sequential Recommendation based on the Collaboration of Multi-Type Behaviors model(MTBCSRec).This model learns both type invariant and type specific preference embedding from the perspective of users and items.This model uses type semantic information to help learn more robust users and items preference embedding.These preference embeddings have not only the essential characteristics of multi type interaction behavior sharing,but also the unique preference characteristics of type specific and different from other types.Multi-Head self-attention machine is used to models the natural sequence dependence of different types of interactions in a user sequence,and personalized learns the users multi-type dynamic preferences for multi-type sequential recommendation.The extensive experiments conducted on real datasets verify the recommended performance of MTBCSRec on sequential recommendation.