The Research Of Personalized Collaborative Filtering Recommendation Algorithms In Electronic Commerce

Collaborative filtering technology is one of the most widely usedtechnologies in the personalized recommender system, so it has been favoredby many researchers. This paper focused on the collaborative filteringrecommendation algorithm. In order to solve the bottleneck problems such asaccuracy、 scalability、 data sparsity and dynamic in the application ofcollaborative filtering recommendation algorithm, this paper proposed somesolutions to these key problems by combining optimization techniques andmachine learning techniques after deep analysis. The concrete researchcontents are as follows:1、 In order to improve the accuracy of similarity measure andrecommendation performance in the traditional collaborative filteringrecommender system, this paper proposed a collaborative filteringrecommendation algorithm based on two stages of similarity learning. The algorithm took advantage of the nearest neighbor algorithm on the first stageto get candidate neighbors and used the reduced gradient method on thesecond stage to learn similarity. Eventually, the algorithm achieved a higheraccuracy of similarity. The experimental results show that the algorithmproposed, on some conditions, not only outperforms the traditional method interms of the error performance but also has a fast convergence speed, whichcan make dynamic similarity adjustment and distributed calculation possible.2、In order to alleviate the influence of the scalability and data sparsityon the recommendation algorithm, this paper proposed a collaborativefiltering recommendation algorithm based on user representative in subspaceand the two stages recommendation algorithm. This algorithm divided theitems set into subspaces and used fuzzy C-means clustering algorithm togenerate K user representative in each subspace. Then it combined with thetwo stages of similarity learning algorithm to predict ratings and makerecommendation, which alleviated the influence of the scalability withoutloss of accuracy. In addition, an improved method of measuring users’similarity in subspace was proposed to alleviate the influence of data sparsity.Experiments show that this algorithm does improve the scalability withoutlosing the accuracy at the same time, and the new method of similaritymeasure in subspace can relieve the influence of data sparsity. 3、In order to improve the performance of the traditional collaborativefiltering recommender system, a dynamic user-item-time three-dimensionalmodel based on rolling time windows was proposed, which considered thetime sequence problem. Then a special collaborative filtering algorithm wasexplored to work with the model. The ratings at different times were regardeddifferently according to the time sequence and the similarities between userswere composed of components at different times, which increased thetimeliness of the algorithm. In addition, the similarities could also becalculated quickly by an incremental formula deduced in this paper so as toimprove the scalability of the algorithm. At last, some reasonableexperiments have been done and show that the model and algorithmpresented in this paper outperforms the traditional2D collaborative filteringmodel and algorithm in terms of the hit rate.