Research On Sequential Recommendation Enhanced By Item Relations And Temporal Point Processes

With the advent of the information explosion era,people are faced with an overwhelming amount of content and choices.As a solution to information overload,recommendation systems have emerged and are widely applied in various web services.Traditional recommendation systems focus on modeling inherent and longterm user preferences,while neglecting the fact that user preferences and item popularity may dynamically change over time.In order to address this issue,timeaware recommendation systems attempt to model dynamic user intent by incorporating time information.Previous research has often treated time information as contextual information or focused on simulating the time decay effect of historical interactions.However,treating time as contextual information poses a challenge in predicting future behavior during unknown periods,and most time decay methods rely on predefined functions based on prior knowledge and lack domain transferability.To address this issue,this article introduces a neural temporal point process to adaptively learn Temporal dynamics.We design two time-aware models from the perspectives of item category(metadata)and item relationships(external knowledge),respectively,with the following main contributions:(1)Users usually have different dynamic intentions for items of different categories.To adaptively capture users’ dynamic preferences for different categories of items,we propose a Type-wise Temporal Point Process Sequential Recommendation(TTSRec).We introduce and redesign a transformer-based neural temporal point process model that adaptively learns users’ dynamic preferences for different types of items.During the prediction phase,the category strength generated by time intervals is combined with the prediction scores of the sequential recommendation system to improve recommendation performance.(2)Items that users have interacted with in the past will dynamically affect the demand for related items.To learn the temporal dynamics of different relationships,we propose a Temporal Dynamic Item Relations Sequential Recommendation(TDR).The model learns representations of relationships through knowledge graph embedding tasks,and dynamically combines these representations by generating the strength of the relationship using a neural temporal point process.The enhanced representations from the TDR model can flexibly combine with various embeddingbased recommendation models,endowing the model with knowledge and time awareness to improve recommendation performance.(3)We conduct experiments on four public datasets and compare our proposed methods with related classical methods in the field.The results confirm that the proposed method is capable of adaptively capturing temporal dynamics in different domains and that it effectively improves time-aware recommendation systems.