Source-linked AI summary
Diffusion Recommender Model
Wenjie Wang, Yiyan Xu, Fuli Feng, Xinyu Lin, Xiangnan He, Tat-Seng Chua
TL;DR
GANs and VAEs have limitations for modeling complex, noisy user interactions, motivating a diffusion-based recommender. DiffRec denoises partially corrupted histories while L-DiffRec and T-DiffRec address scalability and temporal modeling; experiments on three datasets report superiority over baselines.
Problem
GANs suffer unstable training and VAEs have restricted representation ability, limiting accurate modeling of complex user interaction generation.
Method
DiffRec learns recommendations by denoising interaction histories with reduced noise, while L-DiffRec performs latent-space diffusion and T-DiffRec reweights interactions by timestamp.
Results
Experiments on three datasets under multiple settings validate the superiority of DiffRec and its two extensions over competitive baselines.
Takeaways & Limitations
DiffRec establishes a diffusion-based direction for generative recommendation, with extensions for large-scale item prediction and temporal interaction modeling.
Takeaways & Limitations
The authors identify better compression, temporal encoding, conditioning, noise assumptions, and model structures as directions for further exploration.
Abstract
from arXiv · showhide
Generative models such as Generative Adversarial Networks (GANs) and Variational Auto-Encoders (VAEs) are widely utilized to model the generative process of user interactions. However, these generative models suffer from intrinsic limitations such as the instability of GANs and the restricted representation ability of VAEs. Such limitations hinder the accurate modeling of the complex user interaction generation procedure, such as noisy interactions caused by various interference factors. In light of the impressive advantages of Diffusion Models (DMs) over traditional generative models in image synthesis, we propose a novel Diffusion Recommender Model (named DiffRec) to learn the generative process in a denoising manner. To retain personalized information in user interactions, DiffRec reduces the added noises and avoids corrupting users' interactions into pure noises like in image synthesis. In addition, we extend traditional DMs to tackle the unique challenges in practical recommender systems: high resource costs for large-scale item prediction and temporal shifts of user preference. To this end, we propose two extensions of DiffRec: L-DiffRec clusters items for dimension compression and conducts the diffusion processes in the latent space; and T-DiffRec reweights user interactions based on the interaction timestamps to encode temporal information. We conduct extensive experiments on three datasets under multiple settings (e.g. clean training, noisy training, and temporal training). The empirical results and in-depth analysis validate the superiority of DiffRec with two extensions over competitive baselines.
1 INTRODUCTION
DiffRec introduces diffusion-based generative recommendation to address limitations of GANs and VAEs, while L-DiffRec and T-DiffRec target scalability and temporal preference shifts.
- GAN-based recommenders suffer unstable adversarial training, while VAEs trade tractability against the ability to represent heterogeneous user preferences.
- Diffusion Models gradually corrupt and iteratively reconstruct data, offering a tractable forward process and flexible reverse modeling for complex distributions.
- DiffRec learns interaction probabilities by adding scheduled Gaussian noise to interaction histories and iteratively recovering the original interactions.Noise scales are reduced so corrupted histories retain personalized information rather than becoming pure noise.
- T-DiffRec assigns larger weights to later interactions, encoding temporal information to model shifts in user preference.
- L-DiffRec clusters items, compresses each group with a group-specific VAE, and performs diffusion in latent space to reduce parameters and memory costs.
- Experiments on three datasets and multiple training settings report improvements of DiffRec and its extensions over baseline methods.
2 PRELIMINARY
Diffusion Models use a Gaussian forward noising process and a learned reverse denoising process, optimized through an ELBO to reconstruct observed data step by step.
- Forward process: The forward process forms a Markov chain by gradually adding Gaussian noise to x_0 across T steps through q(x_t|x_t−1).The noise scale at step t is controlled by β_t; as T approaches infinity, x_T approaches standard Gaussian noise.
- Reverse process: The reverse process learns p_θ(x_t−1|x_t) to remove noise and recover x_t−1 iteratively from x_T.A neural network predicts the Gaussian distribution’s mean and covariance at each reverse step.
- Optimization: Diffusion Models are optimized by maximizing the Evidence Lower Bound of the likelihood of observed input data.The objective includes reconstruction, prior-matching, and denoising-matching terms.
- Optimization: The denoising-matching terms align the learned reverse transition with the tractable ground-truth transition conditioned on x_t and x_0.
- DiffRec: DiffRec adapts this framework by corrupting user histories forward and learning to denoise them into interaction probabilities.
3 DIFFUSION RECOMMENDER MODEL
DiffRec learns users’ interaction probabilities through iterative corruption and denoising while preserving personalized information. L-DiffRec reduces large-scale prediction costs through latent diffusion, and the training objective combines reconstruction with denoising-matching terms.
- DiffRec framework: DiffRec corrupts interaction histories with scheduled Gaussian noise, then iteratively denoises them to predict users’ interaction probabilities.Its forward process adds noise step by step, while the reverse process recovers interactions through a parameterized neural network.
- Training objective: DiffRec optimizes an ELBO-derived objective containing a reconstruction term and denoising-matching terms across diffusion steps.The reconstruction term measures recovery of x_0, while denoising-matching terms regulate recovery of x_{t−1}; the prior-matching term is omitted as constant.
- Training objective: Importance sampling emphasizes diffusion steps with large loss values, while uniform sampling is used before enough loss estimates are available.The sampling probabilities are estimated by collecting ten L_t values and averaging them.
- Personalized recommendation: Reduced noise scales preserve personalized information instead of driving the latent interaction representation toward pure Gaussian noise.DiffRec reduces training noise through the schedule and uses T′ < T during inference because interactions already contain natural noise.
- L-DiffRec: L-DiffRec clusters items, compresses each cluster’s interaction vector with a group-specific VAE, and performs diffusion in the concatenated latent space.Clustering enables parallel category computation and breaks full connections among encoders, reducing parameters and memory costs for large-scale prediction.
4 EXPERIMENTS
Experiments on three datasets evaluate DiffRec and its extensions across clean, noisy, resource-cost, and temporal settings. Results generally support DiffRec's robustness, L-DiffRec's efficiency, and time-aware variants' effectiveness.
- Experimental settings: Experiments use Amazon-book, Yelp, and ML-1M under clean, noisy, and temporal training settings with full-ranking Recall@K and NDCG@K evaluation.Clean training filters ratings below 4 and splits chronological interactions 7:2:1; noisy training introduces natural or random noise.
- Clean training: Generative methods generally outperform MF and LightGCN, while MultiVAE is especially strong on Amazon-book and Yelp.The comparison includes generative and non-generative baselines, with MultiVAE benefiting from variational inference and multinomial likelihood.
- Noisy training: DiffRec usually surpasses MultiVAE and LightGCN under natural noise, indicating robustness to corrupted interactions.The authors attribute this to iterative recovery of clean interactions from corruption; ML-1M results are comparable because its density makes prediction easier.
- Noisy training: DiffRec remains ahead as random noise rises from 10% to 50%, although all three methods gradually decline.DiffRec is trained under different noise scales at each step, supporting recovery from heavily corrupted interactions.
- In-depth analysis: Using inference step T′ = 0 performs best, while results remain comparable from T/4 to T because small scheduled noises minimally change top-K rankings.The authors link excessive corruption to personalization loss and explain the broader stability through the relatively small noise scale.
- In-depth analysis: A small noise scale such as 1e−4 improves denoising performance, whereas larger scales damage personalization; increasing T has little accuracy effect but raises computation.The study selects T = 5 as a balance between performance and cost, and reports worse ε-ELBO results partly because an MLP struggles to predict sampled noises.
- L-DiffRec: L-DiffRec reduces resources while retaining accuracy, averaging 56.17% fewer parameters and 24.64% lower GPU usage than DiffRec.Against MultiVAE, it reduces parameters by 38.39% and GPU memory by 10.61% on average; clustering and latent diffusion drive the savings.
- T-DiffRec: T-DiffRec and LT-DiffRec outperform DiffRec and L-DiffRec by a large margin, while LT-DiffRec achieves comparable performance to T-DiffRec with fewer parameters.DiffRec-based methods also consume less GPU memory than ACVAE despite having more parameters.
5 RELATED WORK
Diffusion models are presented as a promising alternative for generative recommendation because they offer high-quality generation with lower instability than GANs and broader modeling potential than VAEs. The paper distinguishes DiffRec from prior diffusion-based recommendation by using shared multi-step denoising during inference.
- Diffusion models: Diffusion models have shown low instability and high generation quality across image, text, and audio generation tasks.
- Diffusion models: Recommendation has received limited diffusion-model scrutiny, and CODIGEM predicts interactions using only its first autoencoder during inference.
- Diffusion models: DiffRec differs from CODIGEM by using a shared MLP for multi-step prediction and performing multi-step denoising during inference.
6 CONCLUSION AND FUTURE WORK
DiffRec introduces diffusion-based generative recommendation while adapting the process to preserve personalization, reduce large-scale prediction costs, and model temporal interaction patterns. The paper reports that DiffRec and its extensions outperform baselines in accuracy and resource costs, while identifying several directions for further improvement.
- Conclusion: DiffRec reduces noise scales and inference steps so corrupted interactions retain personalized information rather than becoming pure noise.
- Conclusion: L-DiffRec clusters items for dimension compression and performs diffusion in latent space, while T-DiffRec reweights interactions to capture temporal patterns.
- Conclusion: Experiments on three datasets and varied settings validate DiffRec with both extensions in accuracy and resource costs.
- Future work: Future work includes stronger compression and temporal-encoding strategies, controllable recommendations, alternative noise assumptions, and diverse model structures.