[Paper Review] Diffusion Augmentation for Sequential Recommendation

Qidong Liu, Fan Yan+5 more|arXiv (Cornell University)|Sep 22, 2023

Recommender Systems and TechniquesComputer Science3 citations

TL;DR

This paper proposes DiffuASR, a novel diffusion-based data augmentation framework for sequential recommendation that generates high-quality, preference-aware item sequences to alleviate data sparsity and the long-tail user problem. By adapting a sequential U-Net to model discrete sequence generation and employing two guide strategies for preference alignment, DiffuASR produces augmented data that directly improves SRS models without complex training procedures, demonstrating consistent gains across three datasets and models.

ABSTRACT

Sequential recommendation (SRS) has become the technical foundation in many applications recently, which aims to recommend the next item based on the user's historical interactions. However, sequential recommendation often faces the problem of data sparsity, which widely exists in recommender systems. Besides, most users only interact with a few items, but existing SRS models often underperform these users. Such a problem, named the long-tail user problem, is still to be resolved. Data augmentation is a distinct way to alleviate these two problems, but they often need fabricated training strategies or are hindered by poor-quality generated interactions. To address these problems, we propose a Diffusion Augmentation for Sequential Recommendation (DiffuASR) for a higher quality generation. The augmented dataset by DiffuASR can be used to train the sequential recommendation models directly, free from complex training procedures. To make the best of the generation ability of the diffusion model, we first propose a diffusion-based pseudo sequence generation framework to fill the gap between image and sequence generation. Then, a sequential U-Net is designed to adapt the diffusion noise prediction model U-Net to the discrete sequence generation task. At last, we develop two guide strategies to assimilate the preference between generated and origin sequences. To validate the proposed DiffuASR, we conduct extensive experiments on three real-world datasets with three sequential recommendation models. The experimental results illustrate the effectiveness of DiffuASR. As far as we know, DiffuASR is one pioneer that introduce the diffusion model to the recommendation.

Motivation & Objective

• Address the dual challenges of data sparsity and the long-tail user problem in sequential recommendation systems.
• Overcome limitations of existing data augmentation methods that suffer from poor-quality generated sequences or require complex training procedures.
• Leverage the generative power of diffusion models for discrete sequential data, a novel application in recommendation systems.
• Design a sequence-aware diffusion framework that preserves user preference while generating plausible next items.
• Enable direct training of SRS models on augmented data by ensuring high-quality, semantically coherent generated sequences.

Proposed method

• Propose a diffusion-based pseudo sequence generation framework to bridge the gap between image and sequence generation in diffusion models.
• Design a Sequential U-Net architecture that adapts the U-Net structure for discrete sequence noise prediction, capturing long-range dependencies in interaction sequences.
• Introduce two guide strategies—preference-guided and sequence-guided—using classifier-free guidance to align generated sequences with the user's historical preferences.
• Train the diffusion model end-to-end to generate plausible, diverse, and preference-consistent item sequences for short interaction histories.
• Use the generated sequences as additional training data to fine-tune standard SRS models without modifying their training pipelines.
• Ensure high-quality generation by leveraging the strong modeling capacity of diffusion models and minimizing distribution shift from real sequences.

Experimental results

Research questions

• RQ1 Can diffusion models be effectively adapted to generate high-quality, discrete sequential data for recommendation?
• RQ2 How can the preference of a user’s historical sequence be preserved during the generation of augmented sequences?
• RQ3 Does data augmentation via diffusion generation improve performance on both data-sparse and long-tail user scenarios in sequential recommendation?
• RQ4 How do different guide strategies affect the quality and relevance of generated sequences?
• RQ5 Can the augmented data be directly used to train SRS models without complex iterative or fine-tuned training procedures?

Key findings

• DiffuASR significantly improves sequential recommendation performance across three real-world datasets: Beauty, Sports, and Taobao.
• The method achieves consistent gains on all three benchmark SRS models—Bert4Rec, SASRec, and TransRec—demonstrating its generalization capability.
• Long-tail users, defined as those with fewer than 10 historical interactions, show the most substantial performance improvement, confirming the method’s effectiveness for sparse users.
• The number of augmented items per sequence has a strong impact on model performance, with optimal gains observed at moderate augmentation levels.
• The proposed guide strategies significantly enhance the relevance of generated items, reducing distribution drift and improving alignment with user preferences.
• The ablation study confirms that both the Sequential U-Net and the guide mechanisms are essential for high-quality sequence generation and downstream recommendation accuracy.