PaperlayerView exact revision on arXiv

Source-linked AI summary

From Clues to Generation: Language-Guided Conditional Diffusion for Cross-Domain Recommendation

Ziang Lu, Lei Sang, Lin Mu, Yiwen Zhang

arXiv:2604.05365v1cs.IR

TL;DR

Inter-domain recommendation must predict preferences for source-only users, but scarce overlapping users leave limited alignment signals for transfer. LGCD uses LLM-inferred pseudo-interactions and source-conditioned diffusion to generate target representations, and the authors report that it significantly outperforms state-of-the-art methods.

  • Problem

    Inter-domain recommendation must predict target preferences for users with source-domain interactions but no target history, while scarce overlaps leave limited alignment signals for fine-grained transfer.

    Source: §1 · p. 1

  • Method

    LGCD uses LLM reasoning to infer target interests and map them to real items as pseudo-interactions, then applies source-conditioned diffusion with separate real and pseudo-interaction supervision.

    Source: §1 · pp. 1–2

  • Results

    Extensive experiments show that LGCD significantly outperforms state-of-the-art methods in inter-domain recommendation tasks.

    Source: §1 · p. 2, §5 · p. 10

  • Takeaways & Limitations

    Source-domain features can condition diffusion to synthesize target user representations for cold-start preference prediction.

    Source: §1 · p. 2

    The setup assumes two distinct domains with disjoint item sets and a small subset of overlapping users.

    Source: §3.1 · p. 3

Abstract

from arXiv · show

Cross-domain Recommendation (CDR) exploits multi-domain correlations to alleviate data sparsity. As a core task within this field, inter-domain recommendation focuses on predicting preferences for users who interact in a source domain but lack behavioral records in a target domain. Existing approaches predominantly rely on overlapping users as anchors for knowledge transfer. In real-world scenarios, overlapping users are often scarce, leaving the vast majority of users with only single-domain interactions. For these users, the absence of explicit alignment signals makes fine-grained preference transfer intrinsically difficult. To address this challenge, this paper proposes Language-Guided Conditional Diffusion for CDR (LGCD), a novel framework that integrates Large Language Models (LLMs) and diffusion models for inter-domain sequential recommendation. Specifically, we leverage LLM reasoning to bridge the domain gap by inferring potential target preferences for single-domain users and mapping them to real items, thereby constructing pseudo-overlapping data. We distinguish between real and pseudo-interaction pathways and introduce additional supervision constraints to mitigate the semantic noise brought by pseudo-interaction. Furthermore, we design a conditional diffusion architecture to precisely guide the generation of target user representations based on source-domain patterns. Extensive experiments demonstrate that LGCD significantly outperforms state-of-the-art methods in inter-domain recommendation tasks.

1 Introduction

Inter-domain recommendation must predict target preferences for source-only users, yet scarce overlapping users leave these users without explicit alignment signals. LGCD uses LLM-inferred pseudo-overlaps and source-conditioned diffusion to address this challenge.

1 Introduction

  • Scarce overlapping users leave source-only users without explicit alignment signals, making fine-grained preference transfer difficult.These users have no interaction records in the target domain.Source: §1 · p. 1
  • LGCD uses LLMs to infer cross-domain interests from source behaviors, creating pseudo-overlapping anchors and conditioning diffusion on source-domain patterns.The proposed framework targets cold-start preference prediction through conditional generation of target user representations.Source: §1 · pp. 1–2
  • Free-form LLM predictions may not match discrete target items, while semantic pseudo-interactions can miss collaborative signals and add noise.Diffusion transfer also risks losing source-user personalization or copying source patterns that do not fit the target domain.Source: §1 · p. 2
  • LGCD maps LLM-generated content to real target items, uses real overlaps as alignment anchors, and trains real and pseudo-interaction pathways separately with cyclic batching.A cross-attention denoiser incorporates source preferences at each step; MoE fusion balances transferable source preferences with target-specific patterns, and a lightweight guesser supplies a target-space prior at inference.Source: §1 · p. 2
  • Extensive experiments on benchmark datasets show that LGCD significantly outperforms state-of-the-art methods in inter-domain recommendation tasks.The framework combines LLM-based pseudo-interaction construction with conditional diffusion and additional supervision constraints.Source: §1 · p. 2

2 Related Work

Inter-domain recommendation predicts target preferences from source-domain behavior, while much prior work uses overlapping users to align domains. Diffusion methods have also been applied to recommendation, including cross-domain transfer.

2.1 Cross-domain Recommendation

  • Inter-domain recommendation predicts target preferences for users with no target-domain history, unlike intra-domain methods, which assist users already active in the target domain.Mapping and unified-representation paradigms transfer information across domains, with mapping methods learning embedding projections and unified methods learning shared representations.Source: §2.1 · p. 2
  • Both established paradigms rely heavily on overlapping users; methods using non-overlapping users still struggle to achieve optimal inter-domain performance amid scarce anchors.The cited work includes mapping [22] [26] [46] [54] [55], unified representation [3], and non-overlapping-user studies [5, 8, 9, 28, 30, 31, 42].Source: §2.1 · pp. 2–3

2.2 Diffusion Models in Recommendation

  • Diffusion models reconstruct features from noise and have been applied to sequential and cross-domain recommendation, but existing cross-domain methods rely on overlapping users as anchors.DMRec [27] injects source-domain guidance step by step into the reverse process to generate target-domain user representations.Source: §2.2 · p. 3

3.1 Preliminaries

The setup considers two domains with a small overlapping-user subset and users who interact exclusively in one domain. Its objective is to recommend target-domain items to users observed only in the source domain.

3.1 Preliminaries

  • The setting contains two distinct domains, a small subset of overlapping users, and users who interact exclusively in domain A or B.The item sets for the two domains are disjoint.Source: §3.1 · p. 3
  • Each item in either domain has associated textual information, and the setup distinguishes users by whether their histories are single-domain or overlapping.The passage begins defining historical interaction sequences for these user groups.Source: §3.1 · p. 3
  • Training uses interaction data from overlapping and non-overlapping users to recommend target-domain items to users with source-domain interactions only.Source: §3.1 · p. 3

3.2 Overview of LGCD

LGCD combines LLM-based pseudo-interaction generation, conditional diffusion, and MoE fusion. Together, these components construct target-domain preferences and form a final user representation for recommendation.

3.2 Overview of LGCD

  • LLM-based Pseudo-Interaction Generation prompts an LLM with a user's source history, retrieves semantically similar real target items, and adds them to form a pseudo-cross-domain trajectory.Retrieval uses Top-k matching against target-item text embeddings in a shared semantic space.Source: §3.2 · p. 3
  • The Conditional Diffusion Preference Generator fuses source ID-based collaborative and text-based semantic features to guide cross-attention denoising toward target-domain preferences.Training supervises overlapping users with ground-truth target semantics and pseudo-overlapping users with additional alignment constraints; a lightweight guesser supplies a target-space prior at inference.Source: §3.2 · p. 3
  • The MoE Fusion Module uses source-conditioned gating to weight experts processing source and generated target features, producing a final user embedding.The weighted aggregation balances transferable source knowledge with generated target-specific patterns.Source: §3.2 · p. 3

3.3 LLM-based Pseudo-Interaction Generation

LGCD turns single-domain users’ textual histories into target-domain pseudo-interactions and uses them to bridge the gap to real target items. Separate ID and text sequence representations support this workflow.

3.3 LLM-based Pseudo-Interaction Generation

  • LLM-generated target interests are encoded in a shared semantic space, matched to real target items by cosine similarity, and inserted into source sequences as pseudo-interactions.This retrieval step grounds open-ended generations in the target item collection.Source: §3.3 · p. 3, §3.3.2 · p. 4

3.3.1 Pre-trained Sequence Encoder.

  • Separate Transformer encoders represent ID-based collaborative sequences and text-based semantic sequences, whose item text includes titles, categories, and brands.The sequence models are pretrained independently for domains A and B.Source: §3.3.1 · pp. 3–4

3.3.2 Pseudo Interaction Generation by LLM.

  • For single-domain users, prompts use textual interaction histories to generate pseudo-textual interactions, which are encoded and matched to real target items before insertion.Overlapping users have complete interaction sequences from both domains, while single-domain users provide only observed source-domain sequences.Source: §3.3.2 · p. 4, Figure 2 · p. 4

3.4 Conditional Diffusion Preference Generator

The conditional diffusion generator combines source-domain signals with target features, then denoises toward target preferences. Separate handling of pseudo-data, reverse-process initialization, and training constraints address noise and alignment.

3.4 Conditional Diffusion Preference Generator

  • A dual-path strategy distinguishes real from pseudo-overlapping users, masking pseudo-item ID inputs and adding alignment regularization to limit noisy supervision.The objective is to use augmented data without biasing the learned generative distribution.Source: §3.4 · pp. 4–5

3.4.1 Construction of Conditional Signals and Target Features.

  • Conditional signals combine source-domain collaborative ID and text-semantic representations; target features come from the relevant domain, while pseudo-user conditions use authentic source interactions only.The forward process progressively corrupts target features, approaching Gaussian noise at large steps while retaining partial structure at small steps.Source: §3.4.1 · p. 5

3.4.3 Cross-Attention Denoising Network.

  • The denoiser uses cross-attention to condition generation on source signals and directly regresses clean target features rather than estimating noise.The paper motivates direct regression as concentrating capacity on the user-preference manifold and mitigating high-dimensional sequence-modeling difficulty.Source: §3.4.3 · p. 5

3.4.4 Reverse Process.

  • Reverse diffusion starts from a partially noised initialization and uses denoiser outputs with diffusion parameters to iteratively restore the target feature.The start-step ratio controls the starting point, preserving macroscopic semantic contours and refining a coarse target representation.Source: §3.4.4 · pp. 5–6

3.4.5 Guesser network.

  • Lightweight linear guesser networks map conditional signals into target-domain semantic space and provide the reverse process’s test-time initialization.The paper states that this prior reduces required reverse steps and alleviates training difficulty.Source: §3.4.5 · p. 6

3.5 MoE Fusion Module

The MoE fusion module combines generated target features with source-domain conditional signals, using the latter to guide expert selection and support the final task representation.

3.5 MoE Fusion Module

  • The robust source-derived conditional signal alone drives the gating network, while condition-informed experts process concatenated generated features and conditions for weighted aggregation.This design accounts for residual noise in generated features and anchors expert selection in actual interactions.Source: §3.5 · p. 6
  • The MoE output is used as the final representation for task-loss computation.The section specifies task losses for both real overlapping users and single-domain interactive users.Source: §3.5 · p. 6
  • Cyclic sampling places a fixed number of real overlapping instances in each mini-batch, implicitly reweighting their loss to sustain alignment gradients.The strategy addresses the small share of overlapping users that provide primary supervision for cross-domain mappings.Source: §3.5 · p. 6

4 Experiments

Experiments evaluate LGCD on Movie-Book and Food-Kitchen against several classes of recommendation baselines. LGCD achieves the best performance across all metrics in all domains.

4.1.1 Datasets and Evaluation Protocol.

  • The Movie-Book and Food-Kitchen datasets use different time windows and core filters, and evaluation ranks the ground-truth item among 1,000 candidates using HR@N and NDCG@N for N=5 and 10.Source: §4.1.1 · p. 6

4.1.2 Baseline Methods.

  • Baselines cover single-domain, intra-domain, non-overlapping, and inter-domain recommendation methods.Source: §4.1.2 · pp. 6–7

4.2 Overall Comparison

  • LGCD achieves the best performance across all metrics in all domains; Table 1 reports improvements over baselines significant at p<0.001.Source: Table 1 · p. 7, §4.2 · p. 7

4.1.3 Implementation Details.

  • Implementation uses Baichuan2-7B-Chat2 for LLM inference and jina-embeddings-v2-base-en for text encoding, selecting 10 similar items as pseudo-items.Source: §4.1.3 · p. 7

4.2 Overall Comparison

  • Single-domain methods outperform overlapping cross-domain methods on many metrics, while non-overlapping methods are competitive across datasets.Source: §4.2 · p. 7
  • LLMCDSR performs suboptimally and struggles to map single-domain histories into the other domain’s feature space.Source: §4.2 · p. 7

4.3 Further Analysis

Further analyses examine LGCD’s components, parameter choices, guidance, diffusion targets, and sensitivity to overlap. The reported results support combining real and pseudo-overlap signals and conditioning diffusion on source-domain information.

4.3.1 Ablation Studies.

  • Removing diffusion or relying on only real or pseudo-overlapping interactions weakens target-preference restoration; the two interaction types are complementary.Source: §4.3.1 · p. 8
  • Cyclic batching and alignment loss mitigate pseudo-data noise, while the guesser prior and MoE fusion support target-semantic restoration and source-target preference integration.Source: §4.3.1 · p. 8

4.3.2 Hyper-parameter Sensitivities.

  • Optimal settings are nK=10 and T=100, with λ=0.7 for Movie-Book, 0.5 for Food, and 0.3 for Kitchen.Too many pseudo-items add semantic noise, while excessive reverse steps can over-smooth representations and accumulate error.Source: §4.3.2 · p. 8

4.3.3 Impact of Conditional Guidance Methods.

  • Cross-attention guidance outperforms alternatives across most datasets and metrics; ID-only performs worst on Movie–Book, while text-only performs worst on Food–Kitchen.Source: Figure 4 · p. 9, §4.3.3 · p. 9

4.3.4 Study on Training Targets for Diffusion Models.

  • Directly predicting target features performs significantly better than Noise-MSE and Noise-KL, supporting semantic-feature regression for this task.Source: Figure 5 · p. 9, §4.3.4 · p. 9

4.3.5 Influence of Overlap Ratio.

  • LGCD outperforms baselines across most overlap ratios; performance improves as overlap increases and tends to stabilize around a 0.6 ratio.Source: Figure 6 · p. 9, §4.3.5 · p. 9

5 Conclusion

LGCD combines LLM-inferred pseudo-interactions with conditional diffusion to address limited overlap in inter-domain recommendation. Its experiments report improved state-of-the-art performance for cold-start users in cross-domain scenarios.

5 Conclusion

  • LGCD integrates LLMs and conditional diffusion to address the limited-overlap challenge in inter-domain recommendation.Source: §5 · p. 10
  • LLMs infer potential cross-domain preferences and map them to real items, constructing valid pseudo-interactions.Source: §5 · p. 10
  • Cross-attention-driven conditional diffusion generates target user representations from source-domain conditions, with consistency constraints and cyclic batching stabilizing alignment.Source: §5 · p. 10
  • Experiments report that LGCD significantly outperforms state-of-the-art methods for cold-start users in cross-domain scenarios.Source: §5 · p. 10
Loading 2604.05365v1…