Source-linked AI summary
Exploring Preference-Guided Diffusion Model for Cross-Domain Recommendation
Xiaodong Li, Hengzhu Tang, Jiawei Sheng, Xinghua Zhang, Li Gao, Suqi Cheng, Dawei Yin, Tingwen Liu
TL;DR
Cross-domain recommendation uses source-domain preferences to address cold-start users, but prior mapping-focused methods do not explicitly model preference integration, and existing diffusion recommenders do not directly account for preferences from other domains. DMCDR encodes source interaction history as preference guidance and injects it stepwise during reverse diffusion to generate target-domain user representations. Across three real-world CDR scenarios, the paper reports consistent superiority over SOTA methods and six diffusion-based variants, with improvements up to 32.60%.
Problem
Prior CDR methods generally map source-domain user representations to the target domain without explicitly modeling preference integration, while existing diffusion recommenders do not directly account for preferences from other domains.
Source: §1 · p. 2Method
DMCDR encodes source-domain interaction history as preference guidance and injects it step by step during reverse diffusion to generate personalized target-domain user representations.
Source: §1 · p. 2Results
Takeaways & Limitations
The reported results support preference-guided diffusion as an effective approach to cold-start cross-domain recommendation in the three evaluated scenarios.
Source: §6 · p. 9The setting trains on users overlapping both domains and defines cold-start users as those present only in the source domain.
Source: §2.2 · p. 3
Abstract
from arXiv · showhide
Cross-domain recommendation (CDR) has been proven as a promising way to alleviate the cold-start issue, in which the most critical problem is how to draw an informative user representation in the target domain via the transfer of user preference existing in the source domain. Prior efforts mostly follow the embedding-and-mapping paradigm, which first integrate the preference into user representation in the source domain, and then perform a mapping function on this representation to the target domain. However, they focus on mapping features across domains, neglecting to explicitly model the preference integration process, which may lead to learning coarse user representation. Diffusion models (DMs), which contribute to more accurate user/item representations due to their explicit information injection capability, have achieved promising performance in recommendation systems. Nevertheless, these DMs-based methods cannot directly account for valuable user preference in other domains, leading to challenges in adapting to the transfer of preference for cold-start users. Consequently, the feasibility of DMs for CDR remains underexplored. To this end, we explore to utilize the explicit information injection capability of DMs for user preference integration and propose a Preference-Guided Diffusion Model for CDR to cold-start users, termed as DMCDR. Specifically, we leverage a preference encoder to establish the preference guidance signal with the user's interaction history in the source domain. Then, we explicitly inject the preference guidance signal into the user representation step by step to guide the reverse process, and ultimately generate the personalized user representation in the target domain, thus achieving the transfer of user preference across domains. Furthermore, we comprehensively explore the impact of six DMs-based variants on CDR.
1 INTRODUCTION
Cross-domain recommendation transfers source-domain preferences to improve recommendations for cold-start users, but existing methods emphasize mapping across domains rather than explicitly modeling preference integration. DMCDR uses preference-guided diffusion to generate target-domain user representations, and experiments across three real-world scenarios report superiority over SOTA methods and six diffusion-based variants.
1 INTRODUCTION
- CDR transfers user preferences from an informative source domain to improve recommendations for cold-start users in a target domain.The core task is to draw an informative target-domain user representation from source-domain preferences.Source: §1 · p. 2
- Most prior CDR methods integrate preference into source-domain user representations and then map those representations to the target domain, often leaving integration under-modeled.The neglected integration process may lead to coarse user representations.Source: §1 · p. 2
- Existing diffusion-based recommenders inject information into representations but do not directly account for preferences from other domains, and direct adaptation to CDR is suboptimal.The paper describes diffusion models' feasibility for CDR as underexplored.Source: §1 · p. 2
- DMCDR uses a source-history preference encoder and injects its guidance step by step during reverse diffusion to generate personalized target-domain user representations.The paper also examines six diffusion-based variants for CDR.Source: §1 · p. 2
- Experiments across three real-world CDR scenarios report DMCDR's superiority over SOTA methods and six diffusion-based variants, with improvements up to 32.60%.The introduction does not specify which metric or comparison yields the maximum improvement.Source: §1 · p. 2
2 PRELIMINARY
This section outlines diffusion models and defines the cross-domain recommendation setting. Diffusion models progressively add noise and learn a reverse process, while the task setting distinguishes overlapping users used for training from users present only in the source domain.
2.1 Diffusion Models
- Diffusion models construct a noisy sequence by gradually adding Gaussian noise to an input over T steps, with β_t controlling the noise scale at each step.As T approaches infinity, the representation converges to standard Gaussian noise.Source: §2.1.1 · p. 2
- The reverse process starts from x_T and uses a neural network to denoise over T steps, reconstructing the initial representation x_0.The network can be parameterized by θ; the section notes Transformers and MLPs among possible architectures.Source: §2.1.2 · pp. 2–3
- Training optimizes the data-generating distribution through a variational bound: L_T is independent of θ, L_0 is the negative reconstruction probability, and L_t-1 is a KL-divergence term.The text states that L_T can be ignored during optimization.Source: §2.1.3 · p. 3
2.2 Problem Setting
- The problem setting defines source- and target-domain interaction data and uses overlapping users to train, with cold-start users defined as those present only in the source domain.The supplied setting also distinguishes the target-domain user-item interactions.Source: §2.2 · p. 3
3 METHODOLOGY
DMCDR initializes user and item embeddings, encodes source-domain interaction history as preference guidance, and injects that signal during reverse diffusion to generate personalized target-domain user representations.
3.1 Embedding Layer
- The embedding layer initializes user and item representations in both domains, with d1 denoting the embedding dimension.Source: §3.1 · p. 3
3.2 Preference Encoder
- A Transformer encodes the user's source-domain interaction history, and average pooling produces a guidance signal summarizing personalized preference.The signal guides subsequent diffusion for cold-start recommendation.Source: §3.2 · p. 3
3.3 Diffusion Processes
- The forward process gradually adds Gaussian noise to user representations, using an exponential noise schedule to limit damage to preference modeling.The schedule uses η and lower and upper bounds for added noise.Source: §3.3.1 · p. 4
- Reverse diffusion injects source-domain interaction history as preference guidance step by step, generating a personalized user representation in the target domain.The guided reverse process denoises and reconstructs the user representation.Source: §3.3 · p. 3, §3.3.2 · p. 4
3.4 Training and Inference
- For training, DMCDR scores generated target-domain user representations with an inner product and optimizes rating MSE alongside a preference-guided KL term.The KL term incorporates the guidance signal into the prior distribution.Source: §3.4.1 · p. 4
- Classifier-free guidance jointly trains conditional and unconditional diffusion models by replacing the preference signal with a null token on some training cases.The replacement probability is controlled by p_uncond.Source: §3.4.1 · p. 5
- At inference, cold-start users are initialized from their source-domain representation rather than random Gaussian noise, then reverse-diffused under preference guidance.The guided reverse process runs for T steps to generate the personalized target-domain representation.Source: §3.4.2 · p. 5
4 EXPERIMENTS
Experiments evaluate DMCDR across three real-world CDR scenarios and compare it with existing methods. They also examine model variants, components, hyperparameters, history length, and computational cost.
4.1 Experimental Settings
- Evaluation uses Amazon Book, Movie, and Music domains in three transfer scenarios, MAE/RMSE metrics, and cold-start test fractions of 20%, 50%, or 80%.The scenarios are Movie→Music, Book→Movie, and Book→Music.Source: §4.1.1 · p. 6, §4.1.3 · p. 6, §4.1.4 · p. 6
4.2 Recommendation Performance (RQ1)
- DMCDR significantly and consistently outperforms all baselines across three CDR scenarios, including settings with 80% cold-start users.Table 2 presents the performance comparison for cold-start users in the three scenarios.Source: §4.2 · p. 6, Table 2 · p. 7
4.3 Variant Study (RQ2)
- DMCDR ranks best among six variants, Variant-5 second, and unguided Variant-1 worst; adding too much noise to preference in the forward process can damage personalization.The variants compare different ways of omitting or integrating the source-domain preference during diffusion.Source: §4.3 · p. 7, Figure 3 · p. 7
4.4 Ablation Study (RQ3)
- Removing the diffusion model causes the largest performance drop, while removing preference guidance or the Transformer layer also degrades performance.The ablation results support the roles of preference guidance in the reverse process and the Transformer in capturing personalized preference.Source: §4.4 · pp. 7–8, Table 4 · p. 7
4.5 Hyper-parameter Analysis (RQ4)
- Performance improves as inference steps increase through T′=T; guidance strength helps initially but excess reduces performance, while noise scale first helps then harms and larger T has minimal impact.The study selects T=200 to balance performance and computational cost.Source: §4.5 · p. 8, §4.5.2 · p. 8
4.5.4 Effect of the length of H𝑠
- DMCDR consistently outperforms CDRNP and PTUPCDR across different source-history lengths, and remains strong with short histories.The authors attribute this to generating target-domain representations under preference guidance from limited source-domain interaction history.Source: §4.5.4 · p. 8
4.6 Computational Cost Analysis (RQ5)
- Compared with CDRNP, DMCDR has slightly higher GPU usage and parameter count but much lower per-epoch training time; its efficiencies remain the same order of magnitude.The computational-cost comparison covers three CDR scenarios.Source: §4.6 · pp. 8–9, Figure 7 · p. 8
5 RELATED WORK
Related work distinguishes cross-domain recommendation for overlapping users from cold-start recommendation, and reviews mapping-based approaches alongside diffusion models in recommendation. DMCDR applies diffusion’s information-injection capability to preference transfer for cold-start users.
5.1 Cross-Domain Recommendation
- CDR methods address data sparsity for overlapping users or cold-start recommendation for non-overlapping users; this paper focuses on the latter.CDR aims to use source-domain interaction history to improve target-domain recommendations.Source: §5.1 · p. 9
- Mapping-based approaches transfer preference through learned representations, with meta-learning methods using personalized mappings or preference correlations.Examples include EMCDR and SSCDR, as well as TMCDR, PTUPCDR, REMIT, and CDRNP.Source: §5.1 · p. 9
5.2 Diffusion Models in Recommendation
- Diffusion models support recommendation through denoising, data augmentation, knowledge-graph representation learning, and conditional generation.Examples include DiffRec, DiffKG, Diff4Rec, PDRec, DreamRec, MCDRec, and DiFashion.Source: §5.2 · p. 9
- Unlike prior CDR methods that map features without explicitly modeling preference integration, DMCDR injects preference into user representations and reports new state-of-the-art performance [14, 17, 27].The authors frame this as using diffusion models to address cold-start CDR.Source: §5.2 · p. 9
6 CONCLUSION AND FUTURE WORK
The paper proposes DMCDR, which uses source-domain interaction history to guide diffusion generation of target-domain user representations. Experiments report consistent gains across three real-world CDR scenarios, while future work will explore more challenging tasks such as multi-domain recommendation.
6 CONCLUSION AND FUTURE WORK
- DMCDR uses a user’s source-domain interaction history to guide diffusion generation of a personalized target-domain user representation, transferring preference across domains.The paper also explores six diffusion-based variants for CDR.Source: §6 · p. 9
- Experiments report that DMCDR consistently outperforms state-of-the-art methods and six diffusion-based variants in three real-world CDR scenarios.Source: §6 · p. 9
- Future work will explore diffusion models in more challenging tasks, including multi-domain recommendation.Source: §6 · p. 9