A Matrix Factorization Recommender System Using Item Similarity And Topic Regression

Recommendation is always a hot topic in Big Data area. Many popular meth-ods are based on model-based collaborative filtering, improving the model and algorithm within the framework of matrix factorization. Some models exploits contents of the items to optimize the latent feature matrix of the items. How-ever, these models tend to prefer the data of the item contents, while ignoring the similarity of the items in rating matrix, though this similairty is the core of the traditional memory-based collaborative fitlering. This defect limits the algorithm’s recommendation results.This paper makes a deep research in the traditional content-based recom-mendation, item-based collorative filtering and model-based collorative filtering, trying to find ways to retain the similarity of items in rating matrix and make use of it. Thus, a Model-based Recommender System Using Item Similairty and Topic Regression, CTR-SIM is proposed. The model makes a matrix factorization in training phase, after which the topic regression is taked into algorithm to refine the latent feature matrix of items, while the model also adds additional contraints to ensure the rating similarity of the items is retained on the latent feature matrix. The similarity in the model can also propagate between the similar items which leads to a overall optimization among items. The experiments are based on wide used lastfm dataset and Epinions dataset. The results show that the method pro-posed by this paper have a considerable improvement over the previous methods.