The development of Internet has taken us significant changes in daily life,especially the evolution of web2.0, which changes the mode of information publishingand brings people more convenient online experience. And the e-commerce website, asa principle force, has been in a surging development. In e-commerce website, arecommender system is often applied to mimic people’s shopping habits, and helpmaking some personalized recommendations for what they may like. The good use ofrecommender system can not only raise the sales, but also ease the users’ trouble ofinformation overload by personalized service of recommendation. The promisingdevelopment prospect of e-commerce promotes the research and application ofrecommendation technologies, which many researchers focus on at present.Actually research of recommender system is about the work concentrated onrecommendation algorithm. And nowadays the research of algorithm is mainly carriedout from these aspects: overcoming the shortcomings of classical algorithms, adoptingnew methodologies in algorithms, algorithm adaption in different applicationcircumstances, hybrid methods. There are kinds of algorithms, which are based oncollaborative filtering, content, knowledge discovery, interactions and so on. Amongthese methods, collaborative filtering has been widely used in recommender system,which attracts more focus by researchers. This paper firstly learns some popularrecommendation algorithms and comprehends them by some comparison, and thenfocuses on the research of collaborative filtering algorithm, after that it seeks ways toimprove it and finally proposes an improved user-model-based collaborative filteringalgorithm. The algorithm is improved because, first, it normalizes the ratings bydecoupling normalization method, since different users may have different rating habits.Second, it takes rating time into consideration, and introduces a time weight to ratingvalues so that to mimic users’ interest drift. Third, for the computation of user similarity,an effective weighting factor is added to Pearson correlation similarity computation toget more accurate neighbor users. Furthermore, the user model is constructed offline,which can make the online recommendation much efficient.The proposed algorithm is tested on ml dataset from MovieLens, and the resultsare measured by Mean Absolute Error (MAE). The experiments are done by two groups, one is for the testing of prior parameters, and the other is comparison with otheralgorithms. Results of experiments shows that the algorithm proposed in this paperperforms better in both accuracy and efficiency. |