Matrix Factorization for Video Recommendation Based on Instantaneous User Interest

As the main field of natural language processing, the recommendation algorithm based on matrix factorization is not only widely studied in academia, but also widely used in industry. Most existing methods only pay attention to the influence of the model on the recommendation result, but ignore the effect of the basic user interest in a certain moment. Thus, in this paper we propose a matrix factorization rating prediction model for video recommendation that combined with the user's instantaneous interest. The model integrates the click-through rate (CTR) prediction network and the rating prediction network mainly to achieve the cross-feature effect. And the two models are combined by the user module and the item module, respectively. First, in the user module of the click-through rate prediction network, we use MLP to replace the random initialization of user features in the traditional matrix factorization network. Then, we use a double convolutional layer to replace the convolution-pooling layer to avoid losing the position information in TextCNN. And by combining the Transformer network that extracts long text features on the basis of the click rate prediction network, we can obtain rating prediction network. Finally, the user's click-through rate prediction results are weighted into the rating prediction network to enhance the interpretability of the model and improve the performance of the entire network by preset confidence factor. The experimental results prove that our model is superior to several compared methods on standard datasets.