Recommender systems analyze user attributes and historical behaviors to understand user needs and filter and select the content they want. Recommender systems have been proposed since the 1990s and have been studied to adapt to various domains. Nowadays Recommendations are now essential for the growing e-commerce, as well as for everyday entertainment such as music, movies, socializing, and networking. In recent years, deep learning technology has become a hot trend, which effectively solves the difficulty of combining features manually in traditional machine learning by automatically learning high-level feature representations. It also brings new challenges and opportunities for the development of recommendation systems. From traditional recommendation models to deep recommendation models, from manual feature engineering to automatic feature engineering, and from low-order features to high-order feature learning, recommender systems have been continuously developed and innovated. In addition, nowadays recommender systems are constantly personalized and scene-oriented, and recommender systems under various scenarios have appeared one after another, for example, social network-based recommendation, scenario-aware recommendation, geographic location-based recommendation, etc., and recommender systems are moving towards diversified and personalized development. In this paper, we explore the feature representation in current recommendation algorithms in the application scenario of video recommendation. Efficient feature representations can mine the user's hidden information and thus improve the accuracy of recommendations. An attention mechanism is introduced to propose a deep collaborative filtering recommendation model based on attention. Extensive experiments on real datasets validate the effectiveness of the model recommendations.