Candidate-Aware Dynamic Representation for News Recommendation

Liancheng Xu, Xiaoxiang Wang, Lei Guo, Jinyu Zhang, Xiaoqi Wu, Xinhua Wang

Abstract: With the application of collaborative filtering and deep neural network in news recommendation, it becomes more feasible and easier to capture users’ preferences of news browsing. However, different from traditional recommendation tasks, the collaborative filtering process of news recommendation requires additional consideration of news content. In addition, to capture users’ real intentions from multi granularity interactive behavior information is still challenging. To address above challenges, we propose a hierarchical structure, namely Candidate-Aware Self-Attention enhanced convolution network (CASA). Specifically, we first devise a hierarchical self-attention networks to simultaneously extract the collaborative filtering signals between users and candidate news and their global correlations in multiple dimensions. Secondly, we employ a convolutional networks to map the keywords and important statements of the relevant news into the high dimensional feature space. By considering the additional content-level information, we can further reinforce the the collaborative filtering signals among news. Moreover, we also incorporate time and location relations during the news representation learning to better capture the user’s contextual information. Extensive experiments on two real-world datasets demonstrates that CASA outperforms all the other state-of-the-art baselines on news recomendation task.