Research On Collaborative Filtering Algorithm Of Recommender Systems

Recommender systems play a very important role in e-commerce and other Internetservices. For online shopping sites, recommender systems can encourage users to buy moregoods; for online video, music and news sites, recommender systems can help users findinteresting videos, music and news more easily; for online advertising systems, recommendersystems can improve the CTR.Research on recommender systems is concerned about finding the most accuraterecommendation algorithms. Recommendation algorithms can be classified into three types:1) Collaborative filtering (CF). Recommendations based on historical records of itemsthat the users have viewed, purchased, or rated.2) Content-based filtering (CBF). Recommendations based on information on thecontent of items rather than on other users’opinions or interactions.3) Hybrid Approaches. Combining collaborative filtering and content-based filtering.Collaborative Filtering is one of the most popular approaches among the recommendationalgorithms. Collaborative Filtering can be classified into two types: memory-based CF andmodel-based CF. Memory-based CF use user rating data to compute similarity between usersor items, the similarity is used for making recommendations. Typical examples ofmemory-based CF are user-based k-nearest neighbors CF, item-based k-nearest neighbors CF,Slope One. Model-based CF develops models by using data mining, machine learningalgorithms to find patterns based on training data, the models are used to predictions real data.Typical models of model-based CF are clustering models, matrix factorization (MF),Restricted Boltzmann Machine (RBM).In this paper, we propose several new CF models, they are:1) Neural network base model. We use a three layers neural network model to predictthe rating that user would give to an item.2) Neural tensor network model. Replace neural network base model with neural tensornetwork.3) Neural network base model and knn. Combining neural network base model andk-nearest neighbors.In these methods, we use a vector to represent a user or a movie, vectors and the modelparameters are obtained through training. We use batch gradient descent and error backpropagation algorithm to train the neural network.We test our methods on the movielens-1m datasets. In order to compare with existing methods, such as user-based knn, item-based knn, Slope One, improved item-based knn,ALS-WR, SVD++, we designed some contrast experiments. Contrast experiments are basedon Mahout. The experimental results show that our methods obtained lower RMSE valuesthan existing methods.