Research On Content-aware Collaborative Filtering Recommendation With Extreme Residual Connection

In the age of data overloading,in order to facilitate the user's information search process,various information retrieval technologies have been widely deployed.As a typical information push mode,recommender system has become the core service and main monetization means of many user-oriented systems.Collaborative filtering is the most classic recommendation strategy in a recommender system.It simulates user preferences based on a user-item rating matrix.But in reality,interactive data is often sparse,so it faces the problem of low recommendation accuracy.Deep learning methods can improve the performance of recommender systems over traditional methods,especially when text side information is available.To extract information hidden in the text description of an item and fuse it with rating-related information,it has been proposed that the model be stacked with more layers in neural network.However,as the network thus deepens,the problem of the attenuation of the preamble signal can cause the convolutional neural network to break down.The gradient gradually disappears during the back-propagation process,resulting in an inability to adjust the weights and the training process was suspended.Therefore,this thesis explores the application of extreme residual connection in natural language processing,combined with deep learning methods to train recommendation-based prediction models.The main work is as follows:1.A method called eXtreme Residual connected Convolution Collaborative Filtering(xRConvCF)is proposed to predict the personal rating for each item.It creates data branch lines to form a residual module called the eXtreme Residual(xRes)connection,to mitigate the problem of the vanishing gradient and enhance feature reuse.xRConvCF can effectively capture non-linear user and item interactions,and complete the rating prediction task.2.This thesis first uses text side information to augment the data;then uses the special convolution method in the residual module to learn the abstract content of the text with great extent;then,the user and the item latent vector representation are generated through network learning;finally,the fusion feature information of the two is input to the multi-layer perception(MLP)to obtain the final rating.The network performs parameter learning by minimizing the objective optimization function to reduce errors.Therefore,the fusion of scoring data and text information can effectively alleviate the problem of sparse data,and at the same time,deep learning technology can help the model to obtain accurate recommendation.A large number of experiments conducted on two real data sets show the efficiency of the content-aware collaborative filtering recommendation algorithm based on extreme residual connections.In addition,the synergy of the components of the model architecture proposed in this thesis is verified.At the same time,the effects of the three different parameters of the word embedding dimension,the number of hidden layers and the number of residual layers on the performance of the model are discussed.Last but not least,our model shows good recommendation performance compared with the existing methods.