Research On Deep Sequential Recommendation Technology With Temporal Information

For a long time,sequential recommendation has been a research hot spot in the field of recommender systems.Most sequential recommendation models only focus on sequential modeling of user behavior sequences,ignoring the effective use of temporal information and the deep mining of involved semantic knowledge,so the recommendation performances of these methods are restricted.Temporal information modeling is meaningful for sequential recommendation.On the one hand,the timestamps of user interaction behaviours contain rich semantics,and various temporal information to be used is critical for understanding the true intentions of users and then achieving accurate recommendations.On the other hand,temporal information has the unique advantage of easy access,sequential recommendation can use temporal information to relieve the data sparsity problem effectively.From the perspective of introducing temporal information,this paper studies and explores two manners of temporal information utilization in sequential recommendation including next-time prediction and absolute time periodicity modeling.The major work and contributions of this paper can be stated as below:For the next-time prediction in sequential recommendation,this paper focuses on designing a multi-task combination form which is closer to the real scene to improve the recommendation performance.The idea of this manner is to perform multi-task joint learning on the next-item and-time prediction.A few previous works model the two objectives in a parallel multi-task learning mode.However,these works ignore the critical influence of the next-item prediction result on the next-time prediction.Therefore,this paper proposes a sequential multi-task sequential recommendation model that considers the sequential dependency relationship between the main and auxiliary tasks.This method predicts the next-item and-time in a sequential multi-task learning mode to develop the key influence of the next-item prediction result on the next-time prediction.At the same time,this model uses a bidirectional time interval aware Transformer for sequence modeling to capture the impact of time intervals between any two interactions within the sequence on the representation of user interest.Experiments on three public real datasets show that this model exceeds all baseline methods,and verify the availability of key components of this model.For the absolute time periodicity modeling in sequential recommendation,this paper focuses on learning multi-timescale periodicities and effectively fusing them.The purpose of this manner is to mine the potential periodic regularities in the user behavior sequence and generate the recommendation results that satisfy the user's interests and fit the current recommendation time.A few previous works use multiple absolute time representations as input features for attention computation or sequence modeling,but these works ignore the differences in user preference and periodicity at different timescales.Therefore,this paper proposes a multi-timescale fusion sequential recommendation model.This method uses a multi-timescale enhanced Transformer network to model the interactions and periodicities between items in the sequence at seven different timescales and explicitly extract periodic patterns at different timescales.At the same time,this model uses a user-based personalized attention mechanism to fuse the multiple timescale representations of each position in the sequence to capture the user-oriented periodic regularity preferences.In addition,to enhance the correlation partly between the representations of the same user at different timescales and relieve the problem of user interaction sparsity at some timescales,this paper introduces contrastive learning throughout model training to further improve the representation and modeling capabilities of the network.Experiments on three public real datasets show the advantage of the proposed model,and verify the contributions of key components of this model.In summary,from the perspective of introducing temporal information,this paper conducts deep research on two manners of temporal information utilization in sequential recommendation,and proposes two novel sequential recommendation models.Experiments on some public real datasets prove the superiority of the proposed methods.