Modeling user preference ranking from implicit behavior in short video streaming

In short video platforms, user preferences are often reflected through implicit behaviors rather than explicit ratings. Traditional Quality of Experience (QoE) models relying on subjective Mean Opinion Scores (MOS) or objective transport metrics are insufficient to capture real-world user engagement patterns. This paper presents an implicit feedback-based approach for modeling user preference ranking by constructing pairwise samples from implicit signals such as watch time, likes, and comments. A shallow neural network is trained using a pairwise ranking loss on real-world user interaction logs, with preference labels generated via a heuristic scoring function. Experimental results demonstrate that the proposed method outperforms representative baseline models in ranking accuracy and exhibits strong generalization and interpretability. This study offers an effective and scalable solution for preference modeling in scenarios lacking explicit feedback, and can be extended to applications such as cold-start recommendation or cache prioritization in streaming platforms.