DualCF: Dual Collaborative Filtering For Recommender System

With the rapid development of the Internet technology,information overload has become one of the most disturbing problems.Recommender systems as an effective way of alleviating information overload problem have been extensively studied.As the most widely used recommending algorithm,Collaborative Filtering relies on user history to learn user interests and it assumes that people who like the same thing will have similar interest.Recently,with the expansion of deep learning,research employing deep learning on Collaborative Filtering emerged.Most of these methods utilized deep neural networks to capture the latent vectors of users and items or complex relations between them.These methods mainly mapped raw input into low-dimensional space to model abstract relations.However,low-dimensional space may make some data indivisible.To solve this problem,we propose a Symmetric Deep Neural Network that contains two Multi-Layer Perceptrons,a forward MLP and a reversed MLP,to learn data mapping in low-dimensional and high-dimensional space jointly.SDNN also contains a lateral connection between two MLPs per layer that allows the reversed MLP to contain more information.We further combine SDNN with a matrix factorization model into a unified framework to learn user-item relationship linearly and non-linearly simultaneously and we name the new model Dual CF.We conducted extensive experiments on three public datasets and we compared the performance of single models and combined models respectively.The results demonstrate the effectiveness of our proposed model.We also learned the influence of hyperparameters.