Research On Sequential Recommendation Method Based On Long-Term And Short-Term Preference And Time Interval Information

With the rapid development of the Internet and e-commerce technology,a large number of users have registered many platform accounts,resulting in an explosive increase in data information,making it impossible for users to quickly and accurately find the products related to and needed by themselves.Recommender systems are created to deal with this problem,which analyzes the user’s differentiated characteristics,predicts the subsequent commodity items that he/she may be interested in,and actively recommends items that meet their personalized interests to the user.In real-world scenarios,user preference patterns are dynamic and evolve over time.Specifically,the slow or violent development of the user’s own interests,the emergence and disappearance of hot spots on the external network,hot-selling products,and popular trends will affect the probability that the current recommended items are accepted by the user.Therefore,mining behaviors from the user’s historical sequence another type of sequential recommender system that captures the evolution of users’ dynamic preferences has gradually become popular among scholars.Existing sequential recommendation work still faces the following three problems: First,most of the existing sequential recommendation models only capture the user’s intrinsic interest preference from the user’s long-term behavior pattern,or only analyze the user’s recent behavior pattern to infer the current demand preference,but the impact of the product on the user at different time nodes is differently,existing methods ignore the user’s product feature dependencies and associated elements at different time intervals,resulting in an incomplete understanding of user intention.Second,some existing work often ignores the time context information implied in the user interaction sequence on the time axis when obtaining the commodity order information,which results in the influence of time interval information on user preference modeling can be considered before extracting the sequence information,greatly reducing the accuracy of user preference in deep learning.Third,when users’ long-and short-term preferences are fused,most of the existing works in modeling connect users’ long-and short-term preference features linearly,which results in that users’ long-and short-term preferences cannot play their different roles effectively and affect the accuracy of recommendations.According to the above-mentioned problems,this study proposes a sequential recommendation method based on long-and short-term preferences and time intervals,aiming at the characteristics of dynamic evolution of user preferences in sequence recommendation,based on the user’s long-and short-term sequences,and the time interval in the sequence as auxiliary information,improve the accuracy of recommendation models.The main work of this paper is as follows:1.A sequential recommendation model(MLSUR)based on the representation of users’ long-term and short-term preferences is proposed.This model mainly learns the user’s long-term and short-term interest preferences,and considers the influence of popular items on the user’s recommendation list when integrating preferences.First,the historical sequence of user interaction is divided into two parts,the long-term sequence and the short-term sequence;then the self-attention sequence recommendation method and the gated recurrent unit are used to learn the user’s short-term preference and long-term preference respectively.Finally,the user preference fusion is performed.At the same time,the similarity gate of hot-selling commodities is introduced to calculate the different weights of long-term and short-term preferences,so as to exert different contributions of long-term and short-term preferences.The experimental results show that using deep learning to learn users’ long-term and short-term preferences can improve the representation of user preference modeling and the accuracy of capturing user intention.2.Propose the FTISR algorithm model based on the MLSUR model to further strengthen the prediction accuracy compared with MLSUR.In order to make full use of the time interval information between items in the user’s historical interaction sequence as auxiliary information to model the evolution of users’ dynamic interests,and to explore the influence of different historical periods on user preferences,a sequence recommendation model fused with time interval information(FTISR)is proposed.FTISR first introduces the concept of time interval gate,then uses the time interval gate to update the self-attention network and GRU network to capture the temporal dynamics of user preferences,and finally uses an attention mechanism-based method to balance the user’s long-term preferences and the latest intentions to get the user Final preference representation.The experimental results show that the FTISR model that uses the time interval information in the sequence to jointly model user preferences outperform other high-performance baseline models.accuracy.The research results can provide theoretical and methodological basis for user personalized recommendation and decision support technology under big data.