Top-N Recommendation Based On Video Popularity And Video Classification Tags

With the rapid development of information technology and the explosive growth of data volume,“information overload” has become an inevitable problem in the information age.For solving “information overload”,personalized recommendation systems play more and more important roles.The purpose of recommendation system is to predict whether a user will like a particular item(rating prediction problem)or recommend N interesting items for a certain user(Top-N recommendation problem).Collaborative filtering is a kind of recommendation algorithms that only use user rating data.This thesis focuses on the Top-N recommendation problem,and mainly studies the Top-N recommendation algorithm based on collaborative filtering on video datasets.First,this thesis makes a detailed description of the research background,development status,and related concepts of the recommendation system.The classification,advantages and disadvantages of common recommendation algorithms,as well as the accuracy evaluation metrics of recommendation algorithm are introduced in turn.The different algorithm principles(nearest-neighbors,matrix factorization,and neural network)of collaborative filtering algorithm are described in detail.Then,this thesis studies the principle of matrix factorization algorithm for collaborative filtering.In view of the different rating rule of Singular Value Decomposition++(SVD++)algorithm and Asymmetric Singular Value Decomposition(ASVD)algorithm in two stages(training and predicting),along with the constant result of Top-1 ranking probability of List-wise Learning to Rank with Matrix Factorization(ListRank-MF)algorithm when deal with the same video ratings,the list-wise matrix factorization algorithm,based on video popularity,is proposed.The algorithm uses video popularity to optimize calculation formula of the Top-1 ranking probability,so that the videos at the same rating can still be ranked according to video popularity,thus improving the accuracy of the recommendation.The experiment results indicate that the accuracy of the Top-N recommendation on the MovieLens dataset is improved by 5%-12%,the accuracy of Top-N recommendation on the Netflix dataset is improved by 1%-4%.Finally,this thesis studies the collaborative filtering algorithm based on neural network that has emerged in recent years.Autoencoders Recommendation(AutoRec)and Collaborative Denoising Auto-Encoder(CDAE)both are recommendation algorithms based on autoencoder of neural networks.Taking note of the fact that to a certain extent video classification tags reflect the use's preference,the autoencoder algorithm based on video classification tags is proposed.The algorithm is achieved by Gluon deep learning framework.The experiment results on the MovieLens dataset indicate that the Top-N recommendation accuracy of the algorithm is increased by 3-6 percentage points after using the video classification tags.