Exploiting Holistic Temporal Patterns For Sequential Recommendation

Recommender systems have played a significant role in addressing the problem of information overload,and they have been widely applied across various domains,achieving considerable success.Among the diverse range of recommender systems,the sequential recommendation has become a prominent research focus due to its ability to leverage the sequential information of user-historical interactions.Due to the issue of data sparsity in recommendation systems,relying solely on interaction data is limited in achieving optimal recommendation performance.To accurately capture users’ dynamic preferences,researchers have begun exploring time information to improve recommendation performance.This dissertation systematically investigates the utilization of temporal information in sequential recommendation and identifies three essential temporal patterns:(1)absolute time,(2)relative item time intervals,and(3)relative recommendation time intervals.Despite the significant progress made by existing work,it still has several problems.Existing approaches primarily focused on a single time pattern and overlooked the importance of integrating multiple aspects of temporal information.Moreover,existing methods neglect the inherent connection between items and absolute time.And the relationship between relative item time intervals and relative recommendation time intervals has been disregarded.A Holistic Temporal Pattern based sequential recommendation model is proposed to address the above challenge.HTP model innovatively integrates a variety of temporal information that is useful for recommendation and proposes an item-driven absolute time model while addressing the subtle correlations between the relative item time interval and the relative recommendation time interval.This model consists of three main components,each of which is designed to handle a specific type of temporal pattern at multiple time granularities(e.g.,month,week,and day).Firstly,the Absolute Time module(ATM)learns the absolute time patterns for each item through global data-driven approaches.Secondly,the Item Time Interval module(ITIM)utilizes the time interval information between items to capture their relationships and establish item-specific time interval patterns.Lastly,the alignment between recommendation time intervals and item time intervals is performed to handle the relationship between relative item time intervals and relative recommendation time intervals.To address the issues of data sparsity and time encoding in the HTP model,a Global Item Interval(GTI)sequential recommendation model is proposed.The GTI model tackles the problem of data sparsity by mining global item interval patterns and resolves the shortcomings of time encoding by incorporating continuous time encoding techniques to represent interval information.As a result,the GTI model not only overcomes the limitations of the HTP model but also addresses the shortcomings of existing methods in handling time information in recommendation systems.Compared to the baseline methods,the proposed model leverages the advantages of integrating multiple time patterns,resulting in a significant improvement in recommendation performance.Experimental results on three publicly available datasets demonstrate that the HR@10,NDCG@10,and AUC performance metrics outperform those of the baseline models,validating the superiority of the proposed model.