El-Rec: Enhanced User and News Interaction for News Recommendation

With the development of the Internet, the amount of information data on the Web is growing exponentially. Recommender systems are widely used in online news services. The existing methods usually extract features from news content and historical user behavior and predict news of interest to users. However, these methods only use dot-product to interact news features with user features, making the recommendation system tend to recommend news that is similar to historical news. In fact, this dot-product approach is based only on interest and ignores some biases of the candidate news itself, such as the fact that breaking news is always more eye-catching. To resolve the problem, in this paper we proposed an Enhanced User and News Interaction modeling for News Recommendation. In our method, the click probability is predicted by a gated aggregator which aggregates the news and user representation adaptively. We evaluate El-Rec on two standard news datasets, i.e., MIND-large and MIND-small. The resulting model significantly outperforms previous methods and achieves new state-of-the-art results.