Research On Recommendation Algorithom Based On Collaborative Filtering

With the rapid development of Internet technology,society has entered the information age,which has brought great convenience to people's lives,but also brought some problems.”Information overload” is one of the problems that plague people's lives.The data on the Internet is growing at an exponential rate,and the explosive growth of data makes it very difficult for people to find the information they need in this huge amount of data.The recommendation system is an important tool to solve the problem of ”information overload”,so it has become a major research hotspot.The recommendation system analyzes the user's historical behavior data to establish a user's interest model,and then recommends content that may be of interest to the user,without requiring the user to input his needs,which effectively solves the ”information overload” problem.However,there are some problems with the recommendation system,such as cold starting and low recommendation accuracy.Recommendation algorithms are the core of the recommendation system.In order to make the recommendation system produce a satisfactory recommendation effect,a good recommendation algorithm is needed.Collaborative filtering algorithm is a classic recommendation algorithm.Its algorithm ideology is that using users' history data to calculate similarities,and good recommendation results make it being widely used in recommendation system field.This thesis studies the related collaborative filtering recommendation algorithms to improve the performance of the recommendation system.The specific research work is as follows:First,the Slope One algorithm of collaborative filtering recommendation algorithm is improved,and a Slope One algorithm based on item similarity is proposed.On the basis of the traditional Slope One algorithm,item similarity information is added,and the weight of the item deviation is determined according to the similarity information between items.The item similarity information reflects the importance of the deviation between different items and the target item in the prediction of the score.Experimental simulation results show that the proposed algorithm has higher prediction accuracy than the traditional one.Second,aiming at the cold starting problem in the collaborative filtering algorithm,a scheme using the population statistics information of users is proposed.The user behavior information in the original database is used to calculate the user similarity,and the weight of each statistical feature of a user is calculated according to the user similarity information.Then,the similar users according to the weight of each statistical feature are searched,and recommendations are provided for a special user.Simulation results show that the scheme can solve the user cold starting problem to some extent.Finally,a SVD(Singular Value Decomposition)algorithm for feature extraction using neural networks is studied for collaborative filtering algorithms.The user information and movie information in the Movie Lens 1M dataset are adopted to extract the feature vectors of the users and the movies,and a neural network of three layers is employed to process user information.Meanwhile,a CNN(Convolutional Neural Networks)of two layers is employed to process the text information in movie titles.Experimental results show that compared with the traditional SVD algorithms,the SVD algorithm using neural networks for feature extraction has higher prediction accuracy.