A multiple perceptual news recommendation method based on dynamic multi-convolution kernel CNN and attention mechanism

Abstract

News recommendation is one of the important tasks in recommendation tasks. The existing recommendation process is essentially a system loop feedback process. The content fed back by the recommendation system will be greatly affected by the user's click behavior, while the news recommendation model usually learns from the user's click behavior, which will lead to the system recommending news that users have clicked/liked to watch with high probability and pushing other types of news with low probability, resulting in the problem of insufficient diversity. To effectively address this type of problem, we first analyzed the news data statistically and found the bias of "viewing frequency", and then proposed a multi-perceptual news recommendation method based on dynamic multi-convolution kernel CNN and attention mechanism. We use CNN with convolution kernels of different sizes to extract headline features in the news feature extractor. And then combine them in an attentive way. Form title features with more representations. The user feature extractor is divided into biased user feature extractor and unbiased user feature extractor. One is to learn biased user attributes and perceive user characteristics biased and the other is to learn unbiased user attributes and perceive user characteristics without prejudice. Then we use the cosine similarity loss, that is, to increase the difference between the two user characteristics through cosine similarity. The most important point is that ensure that the accuracy of the model is not reduced as much as possible while debiasing, we combine the two features in an attentive manner, and finally propose a ranking method to improve the diversity of recommendations. Extensive experiments show that our method can not only remove biases, but also achieve a high accuracy of the model.