Multi-Source News Recommender System Based on Convolutional Neural Networks

Abstract

The recommender system can help users solve the problem of information overload and find the item which the user requires efficiently. In this paper, we combine the text and image in a given user's browsing news article and classify the article through deep neural network, then we recommend the news article's tag to the user. In the process of extracting text eigenvector, we apply Convolutional Neural Network(CNN) method because of its good performance. On the other hand, VGG method is used to extract the image eigenvector. Meanwhile we imply the Autoencoder(AE) to reduce the dimension of the output vector from VGG and we regularize the output vector. Then we combine the obtained text eigenvector with the image eigenvector and send the spliced eigenvector to Multi-layer Perceptron(MLP) to classify the given news. Finally, we predict and recommend the tag of the article to the user. Experimental results show that our method has obtained good results on the Ifeng News dataset. Compared with traditional CNN, Long Short-Term Memory(LSTM) and other methods that only recommend news based on the text eigenvector, our method achieves good results. For example, our method's accuracy in the test set is 16.1% higher than CNN and 58.2% higher than LSTM.