Research On Sequential Recommendation Method With High-order Markov Chains

With the rapid development of information technology,people have gradually stepped into the era of information overload from the era of information scarcity.How to help target users find interesting information from massive data has become an urgent problem that needs to be solved in academia and industry.Under this background,recommendation system arises at the historic moment and has become an effective tool to alleviate the problem of information overload.Recommendation system can be regarded as a bridge between information producers and users.It can model users’ preferences by mining users’ historical behavior information,and then provide users with a list of information they may be interested in.The in-depth study of recommendation technology can not only improve the experience of users,but also bring considerable profits to merchants.The traditional recommendation method is based on the potential assumption that users’ interests remain the same,which leads to the failure of obtaining satisfactory recommendation results.Based on the temporal relationship between user behaviors,the sequential recommendation method describes the long-term and short-term interests of users,so as to obtain more accurate recommendation results.The method of decomposing the personalized Markov chain limits the generalization ability of the model due to violation of the triangle inequality.In recent years,a simple and efficient translation recommendation has brought some light to alleviate this problem.However,this method only uses the latest behavioral information of target users to model shortterm interests,which cannot weigh the influence of users’ long-term and short-term preferences on their future decisions.In response to this problem,the contribution of this paper is mainly reflected in the following three aspects:(1)In this paper,sequential recommendation method with high-order Markov chains is proposed.Based on the Trans Rec method,higher-order sequential information of users is explicitly modeled.This method embeds all items and users in a Euclid space,and builds a personalized translation vector in this space with the help of the user’s recent multiple interactions,so that more information about the user’s short-term preferences can be encoded in the translation vector.(2)Aiming at the problem of how to integrate high-order sequential information of users,this paper analyzes the shortcomings of the two common aggregation strategies of maximum pooling and average pooling,and then designs an adaptive weighting mechanism.This mechanism has the ability to learn different contribution degree of different items to the target user’s short-term interest at the moment.(3)Experiments on four real datasets show that the proposed method is significantly superior to current collaborative filtering methods on two evaluation metrics(HR and NDCG),and the necessity and rationality of integrating higher-order sequential information in translation vectors are verified.