Research On Recommending Methods About The Changing Of Users’ Interests

With the development of Internet, with the progress of the times, information inthe web is growing extensively, and this makes it more convenient than before toobtain information; but this also exacerbates the difficulties of filtering informationfor users. Traditional information retrieval techniques could only passively responseuser’s requests, and could not actively present information to users in time.Recommender Systems (RS) based on user’s past behaviors, including explicitevaluations and implicit browsing histories and so on, learn user’s interests orpreferences, and then recommend information that might be interested to users. RScould help users find needed information more quickly, and also could help thewebsites sell their products. Therefore, nowadays, many researchers are denotingthemselves to the studies of this field.Conventional recommending algorithms are based on user’s past interests topredict his current interests, which assume that user’s interests are static. However, inthe real life, user’s interests might be changeable as time passes on. As a result, wepropose a recommending model based on users’ changing interests called RegularizedSingular Value Decomposition with Users’ Changing Interests (UCI_RSVD). RSVD(Regularized Singular Value Decomposition) model includes two matrixes; one isuser-feature matrix each element of which represents the extent of preference to thefeature, and the other one is item-feature matrix each of which represents the amountthat item processes. In the UCI_RSVD model, we make user-feature matrix become anon-linear, monotonic decreasing function, the base function of which is a variant ofEbbinghaus Forgetting Curve, and with the incensement of the time span, the functionvalue changes from fast to slow, and finally reach a constant value. The parameters inthe UCI_RSVD model are learnt by using the gradient descent method throughminimizing the objective function.The dataset used in this paper is MovieLens, which collected by GroupLensResearch. The comparative experiments are the user-based collaborative filtering(including two traditional similarity measurements), singular value decompositionbased two imputation methods, Slope One, RSVD model. And from the experimentalresults, we could conclude that our proposed method, UCI_RSVD model, couldsignificantly improve the performance of recommender systems from the degrees ofcomputing efficiency and the predicting accuracy.