On Efficiency-Effectiveness Trade-off of Diffusion-based Recommenders

TL;DR

This work tackles the efficiency–effectiveness trade-off in diffusion-based sequential recommendations caused by discretization error. It introduces TA-Rec, a two-stage framework that uses Temporal Consistency Regularization during pretraining to enable one-step generation and Adaptive Preference Alignment during fine-tuning to better align with user preferences, both underpinned by theoretical error bounds. Empirical results on multiple real-world datasets show TA-Rec achieves substantial speedups (one-step inference) while delivering superior or competitive recommendation performance compared with both traditional and diffusion-based baselines. The approach holds practical significance for latency-critical deployment and provides a pathway for extending diffusion-based methods to scalable, real-time personalized recommendations.

Abstract

Diffusion models have emerged as a powerful paradigm for generative sequential recommendation, which typically generate next items to recommend guided by user interaction histories with a multi-step denoising process. However, the multi-step process relies on discrete approximations, introducing discretization error that creates a trade-off between computational efficiency and recommendation effectiveness. To address this trade-off, we propose TA-Rec, a two-stage framework that achieves one-step generation by smoothing the denoising function during pretraining while alleviating trajectory deviation by aligning with user preferences during fine-tuning. Specifically, to improve the efficiency without sacrificing the recommendation performance, TA-Rec pretrains the denoising model with Temporal Consistency Regularization (TCR), enforcing the consistency between the denoising results across adjacent steps. Thus, we can smooth the denoising function to map the noise as oracle items in one step with bounded error. To further enhance effectiveness, TA-Rec introduces Adaptive Preference Alignment (APA) that aligns the denoising process with user preference adaptively based on preference pair similarity and timesteps. Extensive experiments prove that TA-Rec's two-stage objective effectively mitigates the discretization errors-induced trade-off, enhancing both efficiency and effectiveness of diffusion-based recommenders.

Figures (5)

• Figure 1: Overall motivation for our work, including the existence of discretization error, high inference costs for diffusion-based recommenders, and the effectiveness and efficiency trade-off.
• Figure 2: The overview of our proposed TA-Rec framework, which alleviates the efficiency and effectiveness trade-off of diffusion-based recommenders with two stages. In the pretraining stage, TA-Rec achieves one-step generation with temporal consistency regularization. In the fine-tuning stage, TA-Rec adopts an adaptive coefficient $\lambda_\beta(t,d)$ to align the generated item with users' preferences.
• Figure 3: Performance of TA-Rec on different $\lambda_c$, demonstrating the sensitivity of TA-Rec to the strength of consistency regularization.
• Figure 4: Performance of TA-Rec on different $\lambda_\beta$, demonstrating the sensitivity of TA-Rec on the strength of preference alignment and the superiority of our Adaptive preference alignment.
• Figure 5: Performance of TA-Rec on multi-step inference settings, demonstrating the generalization ability and robustness of TCR in the multi-step reverse process.