Research On Multi-Behavior Sequence Recommendation Using Temporal Information

Recommendation systems help users alleviate the pressure of information overload and meet their personalized needs,and have been widely applied in various industries.At present,recommendation systems still face the following challenges:(1)Most methods ignore the dynamic changes of users and items,and do not fully utilize the time information of interaction;(2)Previous studies have typically modeled single interaction behavior data between users and items,failing to fully utilize multi-behavior data;(3)The dataset of recommendation systems often suffers from issues such as data sparsity,data imbalance,and data noise.In response to the above issues,this article investigates how to fully utilize the diversity and dynamism of user and item interaction types to improve recommendation performance.The following work was mainly carried out.(1)Propose the Dy MBRec model by utilizing the time information and multi-behavior data of user and item interactions.Firstly,the user’s historical interaction sequence is divided into project sequence,time sequence,and behavior sequence.Then,combined with absolute position,the weights are calculated through the time-aware self-attention module and the behavior-aware self-attention module,respectively,to obtain embedding representations of time and behavior dependence.Finally,the two embedding representations are weighted to obtain the final user preference embedding representation,and the next recommendation is made.Experiments on two real multi-behavior datasets with temporal information show that Dy MBRec achieves better recommendation performance than related sequential recommendation models.(2)Propose the CL-MBRec model and introduce contrastive learning into multi-behavior sequence recommendation to solve the problem of data sparsity and data noise in recommendation.Firstly,five data augmentation methods based on time intervals and items correlations are used to enhance multi-behavior sequence data.Next,a multi-behavior sequence encoder is used to generate sequence embedding representations,while using contrastive learning to maintain high similarity between the enhanced sequence and the original sequence.Then,use multi-task training to optimize the next prediction item and contrastive learning objective of multi-behavior sequence recommendation.Finally,the next recommendation is made based on the obtained user interaction sequence embedding representation.The effectiveness of multi-behavior data augmentation in CLMBRec was verified through experiments on two real multi-behavior datasets with temporal information.