DimeRec: A Unified Framework for Enhanced Sequential Recommendation via Generative Diffusion Models

Abstract: Sequential Recommendation (SR) plays a pivotal role in recommender systems by tailoring recommendations to user preferences based on their non-stationary historical interactions. Achieving high-quality performance in SR requires attention to both item representation and diversity. However, designing an SR method that simultaneously optimizes these merits remains a long-standing challenge. In this study, we address this issue by integrating recent generative Diffusion Models (DM) into SR. DM has demonstrated utility in representation learning and diverse image generation. Nevertheless, a straightforward combination of SR and DM leads to sub-optimal performance due to discrepancies in learning objectives (recommendation vs. noise reconstruction) and the respective learning spaces (non-stationary vs. stationary). To overcome this, we propose a novel framework called DimeRec (\textbf{Di}ffusion with \textbf{m}ulti-interest \textbf{e}nhanced \textbf{Rec}ommender). DimeRec synergistically combines a guidance extraction module (GEM) and a generative diffusion aggregation module (DAM). The GEM extracts crucial stationary guidance signals from the user's non-stationary interaction history, while the DAM employs a generative diffusion process conditioned on GEM's outputs to reconstruct and generate consistent recommendations. Our numerical experiments demonstrate that DimeRec significantly outperforms established baseline methods across three publicly available datasets. Furthermore, we have successfully deployed DimeRec on a large-scale short video recommendation platform, serving hundreds of millions of users. Live A/B testing confirms that our method improves both users' time spent and result diversification.

• The paper introduces DimeRec, a framework that integrates generative diffusion models to enhance sequential recommendation by focusing on generating user interests, not just predicting items.
• Numerical experiments demonstrate DimeRec's superior performance over baselines on various datasets, showing significant improvements in HR, NDCG, and recommendation diversity across different list sizes.
• The framework offers practical benefits for real-world systems by enhancing diversity and user satisfaction, and contributes theoretically by introducing novel diffusion model applications in recommendation.

DimeRec: A Unified Framework for Enhanced Sequential Recommendation via Generative Diffusion Models

The paper, titled "DimeRec: A Unified Framework for Enhanced Sequential Recommendation via Generative Diffusion Models," introduces a novel approach to addressing the challenges inherent in Sequential Recommendation (SR) systems. It focuses on improving both item representation and diversity by integrating generative Diffusion Models (DM) with SR. This approach is encapsulated in the proposed framework, DimeRec, which outperforms baseline methods with significant numerical evidence.

Framework Overview

DimeRec is designed to enhance sequential recommendations by generating the user's next interest rather than predicting the next specific item. This shift is achieved through a novel combination of two primary modules: the Guidance Extraction Module (GEM) and the Diffusion Aggregation Module (DAM).

• Guidance Extraction Module (GEM): This module extracts the user's interests from their historical interactions. It leverages multi-interest models, such as ComiRec, to create a stable guidance signal that the generative diffusion process can utilize. This approach contrasts with existing models that rely on direct encoding of user behavior sequences, which are non-stationary and less effective for diffusion guidance.
• Diffusion Aggregation Module (DAM): DAM employs a generative diffusion process conditioned on the stable guidance from GEM. It aims to reconstruct user interests represented in a noise-space, allowing for simultaneous optimization of recommendation loss and diffusion model's loss within a newly defined loss space. DAM uses Geodesic Random Walk to map item representations onto a spherical manifold to align optimization objectives across different loss functions effectively.

Numerical Results and Experimentation

The numerical experiments reported in the paper highlight DimeRec's superiority over other methods, such as SASRec, GRU4Rec, and DiffuRec, across several datasets, including YooChoose, KuaiRec, and ML-10M. Notably, DimeRec achieves substantial improvements in Hit Rate (HR) and Normalized Discounted Cumulative Gain (NDCG), particularly when retrieving a larger number of items (e.g., HR@50, NDCG@50). These results underscore its capability to enhance representation learning, as verified by linear probing accuracy experiments, and to offer more diverse recommendations, as evidenced by an increased average number of categories in top recommendations.

Implications and Future Speculations

The implications of DimeRec's framework are twofold:

1. Practical Advancement in Recommender Systems: By adopting the generative paradigm for user interest generation, DimeRec addresses both diversity and exploration-exploitation challenges in SR, thereby improving user satisfaction in real-world applications, such as the short video recommendation platform where it has been deployed with notable success.
2. Theoretical Contributions to Diffusion Models in Recommendation: The introduction of a stable noise space and Geodesic Random Walk adaptation for item representation sets a novel precedent in the use of diffusion processes within RS. This lays a foundation for further research into the interplay between generative models and recommendation algorithms.

In conclusion, DimeRec's unified framework not only advances the state-of-the-art in SR by facilitating richer and more diverse user experiences but also contributes a robust methodological innovation with its integration of diffusion models. The progression of this research could witness expanded applications of diffusion models, potentially leading to more sophisticated personalization strategies in AI-driven recommendation systems.