Research On User Dynamic Recommendation Model Based On Interest Drift

The problems of "Information Trek" and "Information Overload" have a long history,the recommendation system can help users to screen in a large amount of product space.However,traditional recommendation algorithms are mainly static algorithms.The basic assumption is that the user's interest is static.This assumption is inconsistent with the facts,leading to many problems such as poor recommendation accuracy and poor interpretability when facing real data.In fact,user interest is not only constantly changing,but also has a certain pattern.In the current research results of user interest modeling,research that considers dynamic changes in user interest often treats user interest attenuation equally,and less distinguishes user interest types.The analysis from the time factor identification did not consider the number of users' attention to the project attributes and the user's activity,item's popularity,etc.which lead to the longterm and short-term interest information representation incomplete,and the user's interest characterization is not comprehensive.In addition,most of the literature in the recommendation field focuses on processing explicit scoring data.However,in many practical situations,especially in e-commerce business recommendation systems,implicit feedback is more common.This paper proposes a hybrid recommendation model LSIMF(Long-term and Short-term Interest and Matrix Factorization for Collaborative Filtering)based on the implicit feedback data set.The main contributions of this paper are:(1)The hybrid model constructed in this paper proposes solutions to the problems of sparse data,inaccurate characterization of user interests,and traditional algorithms that deal with interest drift too much on time factors and ignore active exploration of new interests of users.By replacing the user-item preference matrix with the userattribute preference matrix,the sparsity of the matrix is reduced.Through user preference modeling and subdivision of user interest patterns,user preference documents and user interest distributions are obtained.Through the collaborative filtering of fusion matrix decomposition,the user's new interests are actively explored.(2)Experimental results show that compared with forgetting curve,time window and collaborative filtering algorithm based on matrix factorization,the hybrid recommendation model LSIMF proposed in this paper has significantly improved accuracy,recall,and F1-Score evaluation indicators.(3)The hybrid model constructed in this paper extracts user's interest sets,user activity,item popularity and other related indicators,which can automatically tag users and provide more options for subsequent research on the dynamic evolution of user interests or the expansion of website functions.