Research On Personalized Recommendation Algorithms Based On Latent Factor Model

Recommendation system is an effective tool to solve the problem of information overload, which can make personalized recommendation according to user history behavior. Latent factor model algorithm is a classic recommendation algorithm which has achieved great success in both research and application, but there are some problems still worth discussing and researching. Firstly, the algorithm has a poor performance in the case of sparse data, and cannot solve the problem of cold start. The second is the Latent factor model recommendation algorithm has poor scalability. The algorithm parameters need to be updated all at one time. It needs large amount calculation and time-consuming.In order to solve low accuracy problem of latent factor model recommendation algorithm in the case of sparse data, a new latent factor model which can make use of user attributes is proposed. The model combines user attributes in traditional latent factor model, which still can provide recommendations in the case of sparse data according to the user attributes. To solve the problem of the poor expansion of latent factor model, a parallel improved latent factor model algorithm is proposed, which has better scalability. The main research contents of this paper are as follows:1. The thesis analyzes the advantages and disadvantages of mainstream recommendation algorithms, including content-based recommendation algorithm, collaborative filtering recommendation algorithm, network-based recommendation algorithm and hybrid recommendation algorithm. Aiming at the poor performance of traditional latent factor model algorithm in the case of sparse data, an improved latent factor model that can make use of user attributes information is proposed. The improved method can make recommendations by user basic attribute information when the data is insufficient. The method measures the importance of attributes by logistic regression and searches the neighborhoods with similar attributes information so that it can make recommendation according to neighborhoods. Experimental results show that the algorithm can solve the problem of data sparsity in some degree.2. In order to improve the poor scalability of traditional latent factor model recommendation algorithm, this thesis designs and implements a parallel latent factor model algorithm based on Hadoop platform. The process of training model, predicting score by attributes and generating recommendation results are decomposed for a series of Map-Reduce tasks so that the algorithm can run simultaneously on multiple machines, which not only greatly reduces the training time of the model but also increases data processing speed. Experiments show that the algorithm has good scalability on the cluster with the increasing data.3. This Thesis designs a Hadoop based movie recommendation system with the recommendation algorithm proposed. The system can suggest films that user may be interested in by user rating history data. The system has good scalability for massive data processing and can make personalized recommendation for different users in a short time.