Fine-Grained Offline Ranking for Short Video Advertisements CTR via Visual- and Audio-Based Features

Traditional methods for the click-through rate (CTR) prediction of short video advertisements (ads) face several challenges, such as dependency on user information, evaluation criteria inapplicable, and classification coarse-grained. In this paper, we propose FOR_CTR (Fine-grained Offline Ranking for short video advertisements CTR via visual-and audio-based features), a framework that utilizes visual and audio features to accurately rank short video ads without relying on user information. It can enhance the ranking process of short video ads by effectively assessing the fine-grained scores of each aspect. Specifically, we segment short video ads into image, audio, and matching degree aspects, which are independent of user information to model CTR of short video ads. To analyze and understand user preferences effectively, we introduce a Hierarchical Intra-modal Feature Disentanglement module, along with a Global-based Auxiliary Loss module for effective supervision. This approach allows us to explore the factors that truly affect the CTR in video content, thereby ensuring the applicability of the evaluation criteria. Moreover, we propose an Inter/Intra-modal Video Rating module to assign score to each feature and regulate their orders using pairwise matching in the Primary Loss module, avoiding coarse-grained classification. Experimental results on diverse real-world datasets confirm the effectiveness of our proposed method in offline CTR ranking. Our codes are released at: https://github.com/ChenrMa/FOR_CTR.