Non-Negative Matrix Factorization Based On Sigmoid Function And An Application In Recomender System

We often face the problem of finding what we need among the sea of information nowadays. One of the solutions is the recommender system, which has the ability to recommend personal items to the user based on their character and history activities. But the lack of precision and scalability remains two main flaws in the recommender systems. We try to improve the precision of prediction by proposing a matrix factor-ization method based on sigmoid function. And we also improved and implemented a distributed algorithm for stochastic gradient descent method so that the matrix factor-ization can be solved in a large scale.The main contributions of this article include:1)Proposed a scalable non-negative matrix factorization method with high preci-sion based on sigmoid function. Improved the precision of rating prediction and kept the scalability by combining non-negative matrix factorization with latent factor model.2)Improved and implemented a distributed algorithm for stochastic gradient de-scent method to solve the matrix factorization. Compared with the current Distributed stochastic gradient descent algorithm, costs on communication are reduced and loads among nodes are more balanced.3)Implemented a prototype system for movie recommendation based on the tech-niques mentioned above. Users can browse and rate movies. The system will recom-mend new movies to them based these rating activities.