An Improved Collaborative Filtering Similarity Model Based on Neural Networks

At present, collaborative filtering has already become one of the most widely used means to provide personalized recommendation service. The crux of collaborative filtering is to utilize user-rating-data matrix for the sake of figuring out similar neighbor users or items with similar ratings, which can be realized mainly by virtue of similarity algorithms. Recommendation consequences which are gained though the traditional algorithms, like Pearson correlation coefficient, cosine and so forth, might not be satisfactory, especially in the context of comparatively sparse user-rating-data matrix. The thesis based on the improved similarity algorithms proposed by relevant researchers, analyses the shortcomings of the pre-existing algorithms, redefines the similarity algorithm formulas including the concept of weight, and furthermore makes use of the characteristics of neural network to practice in order to the optimal weight. This thesis designs an experiment on the foundation of two real data sets, comparing the recommendation effects between the newly-built similarity algorithm models and the pre-existing ones. The results indicate that the superiority of the new method in different parameter circumstances.