FAVE: Flow-based Average Velocity Establishment for Sequential Recommendation

Abstract

Generative recommendation has emerged as a transformative paradigm for capturing the dynamic evolution of user intents in sequential recommendation. While flow-based methods improve the efficiency of diffusion models, they remain hindered by the ``Noise-to-Data'' paradigm, which introduces two critical inefficiencies: prior mismatch, where generation starts from uninformative noise, forcing a lengthy recovery trajectory; and linear redundancy, where iterative solvers waste computation on modeling deterministic preference transitions. To address these limitations, we propose a Flow-based Average Velocity Establishment (Fave) framework for one-step generation recommendation that learns a direct trajectory from an informative prior to the target distribution. Fave is structured via a progressive two-stage training strategy. In Stage 1, we establish a stable preference space through dual-end semantic alignment, applying constraints at both the source (user history) and target (next item) to prevent representation collapse. In Stage 2, we directly resolve the efficiency bottlenecks by introducing a semantic anchor prior, which initializes the flow with a masked embedding from the user's interaction history, providing an informative starting point. Then we learn a global average velocity, consolidating the multi-step trajectory into a single displacement vector, and enforce trajectory straightness via a JVP-based consistency constraint to ensure one-step generation. Extensive experiments on three benchmarks demonstrate that Fave not only achieves state-of-the-art recommendation performance but also delivers an order-of-magnitude improvement in inference efficiency, making it practical for latency-sensitive scenarios.

Figures (5)

• Figure 1: Comparison of generative recommendation paradigms. (Left) Existing "Noise-to-Data" generative methods suffer from prior mismatch and linear redundancy. (Right) Our proposed Fave employs direct trajectory transport from a semantic anchor prior.
• Figure 2: The overall framework of Fave which adopts a two-stage training strategy. It first constructs a basic manifold by learning the instantaneous velocity field from Gaussian noise. The second stage incorporates a semantic anchor prior and average velocity modeling to address the prior mismatch and linear redundancy issue.
• Figure 3: Comparison of inference trajectories for Fave (left) and FMRec (right) on ML-100k. The trajectory points are first $\ell_2$-normalized and then visualized using t-SNE.
• Figure 4: Visualization of embedding distributions on ML-100k for Gaussian noise, the proposed semantic anchor prior, and ground-truth item embeddings, showing that the semantic anchor prior is initialized closer to the target distribution.
• Figure 5: Performance of different hyperparameters. (a) weight $\alpha$ for $\mathcal{L}_{tgt}$, (b) weight $\beta$ for $\mathcal{L}_{src}$, (c) weight $\gamma$ for $\mathcal{L}_{cons}$, and (d) retention rate $\rho$.