Technical Summary: FlowTime

Problem Definition and Motivation

Watch time prediction (WTP) is a critical metric for optimizing user engagement in short-video recommender systems, serving as a proxy for satisfaction and retention. However, existing WTP methods suffer from inherent paradigm-specific limitations that prevent them from capturing the intrinsic multimodality and heterogeneity of user-item interaction patterns.

The paper identifies three dominant paradigms with specific flaws:

1. Direct Regression: Optimized via Mean Squared Error (MSE), this approach implicitly assumes a unimodal Gaussian prior. This leads to "mean-collapse," where the model predicts an average value that falls in a low-density region of complex, multimodal distributions, failing to reflect the true ordinal structure.
2. Ordinal Regression: This method discretizes continuous time into intervals for classification. It suffers from rigid quantization errors and erroneously assumes conditional independence among intervals, ignoring the sequential dependencies of watch time.
3. Discrete Generative Regression: Adapting language modeling to generate quantized time tokens, this approach is hampered by heuristic vocabulary design and high inference latency due to autoregressive decoding.

Beyond these specific flaws, a shared deficiency is the inability to model User-Item Interaction Patterns as structural confounders. The paper argues via causal analysis that identical user interests can manifest as distinct watch time distributions (e.g., bimodal vs. unimodal) depending on specific user habits and item characteristics. Traditional methods largely neglect these pattern-driven distribution shifts.

Methodology: FlowTime

To address these challenges, the authors propose FlowTime, a novel framework operating under a new paradigm: Continuous Generative Regression. This approach aims to directly fit complex, multimodal watch time distributions in a continuous space.

Core Architecture

FlowTime is built upon a One-step Generative Variational Autoencoder (VAE) architecture, designed to balance expressivity with the low-latency requirements of online systems.

• Probabilistic Encoder: Approximates the posterior distribution of latent variables given user/video features and the target watch time, utilizing a reparameterization trick for end-to-end optimization.
• Generative Decoder: Reconstructs the watch time distribution from the latent space. Unlike iterative diffusion models, FlowTime employs a single-step generation process to ensure real-time inference efficiency.

Flow-based Personalized Prior

A key innovation is the replacement of the standard isotropic Gaussian prior with a Flow-based Personalized Prior.

• Pattern Encoding: The model encodes historical watch time sequences for both users and items using quantile statistics. These sequences are processed via a Transformer-based architecture to capture global distributional shapes and generate contextual embeddings.
• Manifold Warping: Leveraging Normalizing Flows (NFs), the model learns a sequence of invertible, bijective transformations conditioned on these historical patterns. This warps a simple base Gaussian distribution into a complex, pattern-specific latent manifold.
• Adaptive Modeling: This personalized prior allows the model to adaptively capture diverse and complex preference distributions, enabling the decoder to yield precise predictions that respect the specific interaction patterns of the user-item pair.

Optimization Objectives

The framework employs a joint multi-objective loss function:

1. Accuracy Anchoring: Uses Huber loss in the observation space to constrain point estimates and improve regression fidelity.
2. Latent Space Regularization: Uses KL divergence to align the approximate posterior with a standard Gaussian, preserving continuity and identifiability in the latent space.
3. Distribution Alignment: Introduces a Wasserstein distance-based loss on quantile representations to align the generated watch time distribution with the historical interaction patterns of users and items.

Key Contributions

1. New Paradigm: The paper formally proposes Continuous Generative Regression as a fourth paradigm for WTP. It provides a theoretical analysis demonstrating how this paradigm overcomes the mean-collapse of direct regression and the quantization errors of ordinal/discrete generative methods.
2. FlowTime Framework: The authors introduce FlowTime, the first framework for this new paradigm. It integrates one-step generation with a Flow-based Personalized Prior to reshape the prior space into pattern-driven manifolds, achieving adaptive modeling without sacrificing inference efficiency.
3. TimeRec Library: To address inconsistencies in data preprocessing and evaluation, the authors built TimeRec, the first open-source Watch-Time Prediction Library. It standardizes evaluation protocols and integrates reproduced State-of-the-Art (SOTA) methods to establish a rigorous benchmarking standard.
4. Novel Metric: The paper introduces Personalized Distributional Fidelity (PDF), a metric based on Jensen-Shannon Divergence to quantify how well a model captures personalized multimodal distributions.

Experimental Results

The authors conducted extensive offline experiments on three datasets (KuaiRand-Pure, KuaiRec, and a large-scale industrial dataset) and online A/B tests.

• Offline Performance: FlowTime consistently outperformed SOTA baselines across all metrics.
  • Compared to Direct Regression, it significantly reduced Mean Absolute Error (MAE) and improved ranking capability (XAUC), effectively mitigating mean-collapse.
  • Compared to Ordinal and Discrete Generative methods, it avoided quantization errors and heuristic vocabulary limitations.
  • It demonstrated superior flexibility over mixture-density-based models (e.g., EGMN) in modeling complex interaction patterns, evidenced by significant improvements in the PDF metric.
• Efficiency: FlowTime achieved a superior efficiency-performance trade-off. While marginally slower than direct regression, it was significantly faster (nearly 30% reduction in latency) and more accurate than iterative generative baselines like GR.
• Online A/B Testing: Deployed on a platform with over 400 million Daily Active Users (DAUs), FlowTime yielded statistically significant gains:
  • +1.044% in Video Play Time.
  • +1.027% in App Usage Time.
  • +1.236% in Share Rate.
• Ablation Studies: Experiments confirmed the necessity of Normalizing Flows (removing them degraded performance significantly) and the critical role of user history in personalization. The study also validated that the one-step architecture offers a robust inductive bias for high-sparsity scenarios compared to iterative denoising.

Significance

The paper claims that FlowTime represents a significant advancement by shifting the WTP problem from point estimation to continuous generative modeling. By causally identifying interaction patterns as structural confounders and utilizing Normalizing Flows to warp the prior space, FlowTime successfully captures the intrinsic heterogeneity of user behavior that previous methods missed. Furthermore, the release of the TimeRec library is presented as a crucial step toward establishing fair, consistent, and reproducible benchmarking standards for the field of watch time prediction.