Title: FlowTime: Advancing Continuous Generative Watch Time Prediction through Flow-based Personalized Priors

Abstract: In the realm of short-video recommendation systems, watch time has become a critical metric for enhancing deep user engagement. Nevertheless, existing Watch Time Prediction (WTP) techniques are constrained by fundamental limitations tied to their specific paradigms. Direct Regression methods frequently encounter mean-collapse, a consequence of relying on unimodal Gaussian assumptions. Ordinal Regression, on the other hand, is impeded by quantization errors stemming from rigid discretization processes. Meanwhile, Discrete Generative Regression contends with significant inference latency and the complexities associated with heuristic vocabulary design.

Beyond these individual shortcomings, a common deficiency across current approaches is the failure to capture the intrinsic multimodality and heterogeneity inherent in User-Item Interaction Patterns. To tackle these issues, we re-examine the WTP problem through a causal lens, identifying these user-specific patterns as structural confounders that influence watch time outcomes. Specifically, identical interests can result in varied watch times depending on diverse user habits.

We then formally introduce a new paradigm—termed Continuous Generative Regression—and present FlowTime, a novel approach that employs a One-step Generative Variational Autoencoder. FlowTime successfully bypasses the latency issues associated with iterative denoising while preserving the expressive power of continuous latent spaces. Additionally, we have developed a Flow-based Personalized Prior. This component utilizes Normalizing Flows (NFs) to transform a standard Gaussian prior into a complex, history-conditioned manifold, allowing for the adaptive modeling of multimodal interaction patterns.

Furthermore, we introduce TimeRec, the inaugural open-source WTP library, accompanied by a new personalization metric designed to set a rigorous benchmarking standard. Comprehensive offline experiments and online A/B tests confirm that FlowTime significantly outperforms state-of-the-art methods.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC