Abstract: As the rapid development of Internet, danmaku video emerges. This user-video interactive mode has new features, such as real-time dynamic emotion expression and multimodal emotion distribution. Meanwhile, the new features also bring challenges to practical research work, for instance, user portrait depiction is more difficult, video recommendation and advertising need to be more accurate. Existing research has not yet deeply analyzed new features of danmaku video, nor fully explored the academic research value for itself. Based on the theories of psychology, marketing, and some other interdisciplinary frontier knowledge, as well as combined with deep learning, natural language processing technology and system dynamics method, this study attempts to analyze and model danmaku video data from a data-driven perspective, for deeply mining the potential business value of video big data. The objective of the study is to create a high-quality online video intelligent marketing platform. A sentimental recognition approach is proposed to capture the internal relationship between sentimental features of danmaku text and visual sentimental features of video, so as to represent users' sentiment. Considering the influence of users' dynamic emotion on behavior, the dynamic user portrait with sentimental features is constructed to describe full picture of user characteristics, as well as to explore the access logic and behavioral preferences that are hidden behind behavior data. By positioning user roles, the sticky marketing mechanism of online video platform is established. Furthermore, focused on different advertisement types, like creative ads, interstitial ads and oral ads, this study explores the impact of video ads on the emotional and psychological changes of users, reveals the correlation between users' sentiment changes and video ads insertion mode, and thus puts forward the “content customized” dynamic advertising strategy for danmaku video. It not only enriches the existing research, but also provides theoretical guidance and decision-making support for online video platforms to accurately locate and analyze user demands. In particularly, the contributions of the study include: (1) Previous research on marketing strategy of online video platform mainly focused on platform pricing, brand reputation and user satisfaction, ignoring external stimulus factors, such as dynamic factors of user behavior and emotional changes caused by danmaku video advertising. This paper combines big data of danmaku video to analyze users' emotional responses to the external stimuli. Based on multimodal attention fusion, the hierarchical deep association co-attention model is designed to achieve users' sentiment recognition. The paper reveals the correlation between video ads and users' emotional changes, proposes the dynamic intercutting strategy of customized video ads, innovates the video advertising marketing model, and helps advertisers to enhance the brand marketing value; (2) The traditional method of constructing user portrait does not focus on the strong correlation among users' sentiment, natural attribute and behavior. To this end, the study is driven by the sticky marketing of online video platform. Based on three dimensions of users' sentiment, natural attribute and behavior, a dynamic user portrait model with four modules of basic data, behavior modeling, service application and evaluation feedback is established. Considering user demand and dynamic user portrait, the model of sticky marketing mechanism is constructed to recommend rich content options for users, to increase users' stickiness and to prolong users' usage time in online video platform. Thus, it forms a benign marketing mode of flow closed loop and completes the flow transformation between user value and content value in online video platform. Through this study, user needs will be met from a higher quality level, the accuracy of video advertising will be comprehensively improved, and the marketing value can be maximized in online video platform.

Key words: danmaku video, bimodal sentiment fusion, dynamic user profile, sticky marketing mechanism, dynamic advertising strategy