Source-linked AI summary
DisCo: Graph-Based Disentangled Contrastive Learning for Cold-Start Cross-Domain Recommendation
Hourun Li, Yifan Wang, Zhiping Xiao, Jia Yang, Changling Zhou, Ming Zhang, Wei Ju
TL;DR
Cold-start recommendation lacks enough observed interactions for new users, and source-domain similarity may not reflect their target-domain interests. DisCo uses graph-based intent modeling and contrastive learning to filter irrelevant information; the paper reports that it outperforms existing methods on public cold-start CDR evaluations.
Problem
Cold-start users lack sufficient observed interactions, and users similar in the source domain may have different interests in the target domain.
Source: Introduction · pp. 1–2Method
DisCo uses a multi-channel graph encoder and intent-wise contrastive learning guided by user similarities to capture intents and filter irrelevant source-domain information.
Source: Introduction · p. 2Results
The paper reports that DisCo outperforms existing methods in extensive public-dataset evaluations of cold-start CDR.
Source: Introduction · p. 2Takeaways & Limitations
The proposed framework targets negative transfer by capturing detailed user intents and filtering irrelevant collaborative information.
Source: Introduction · p. 2
Abstract
from arXiv · showhide
Recommender systems are widely used in various real-world applications, but they often encounter the persistent challenge of the user cold-start problem. Cross-domain recommendation (CDR), which leverages user interactions from one domain to improve prediction performance in another, has emerged as a promising solution. However, users with similar preferences in the source domain may exhibit different interests in the target domain. Therefore, directly transferring embeddings may introduce irrelevant source-domain collaborative information. In this paper, we propose a novel graph-based disentangled contrastive learning framework to capture fine-grained user intent and filter out irrelevant collaborative information, thereby avoiding negative transfer. Specifically, for each domain, we use a multi-channel graph encoder to capture diverse user intents. We then construct the affinity graph in the embedding space and perform multi-step random walks to capture high-order user similarity relationships. Treating one domain as the target, we propose a disentangled intent-wise contrastive learning approach, guided by user similarity, to refine the bridging of user intents across domains. Extensive experiments on four benchmark CDR datasets demonstrate that DisCo consistently outperforms existing state-of-the-art baselines, thereby validating the effectiveness of both DisCo and its components.
Introduction
Cold-start CDR uses interactions from a source domain to recommend in a target domain, but source-domain similarity may not reflect target preferences. DisCo addresses this risk by modeling fine-grained user intents and filtering irrelevant collaborative information.
Introduction
- Cold-start CDR leverages interactions from related domains to improve prediction in a target domain for users lacking target-domain interactions.Traditional embedding-and-mapping methods may overlook diverse user-specific preferences.Source: Introduction · p. 1
- Users with similar source-domain preferences may differ in the target domain, so transferring irrelevant collaborative information can bias target recommendations.The paper identifies this mismatch as a source of negative transfer and sub-optimal performance.Source: Introduction · p. 2
- DisCo combines a multi-channel graph encoder with intent-wise contrastive learning to capture user intents and filter irrelevant source-domain collaborative information.It uses cross-domain user similarity to help preserve target-domain-specific preferences while bridging domains.Source: Introduction · p. 2
- The paper reports extensive evaluations on public datasets and states that DisCo outperforms existing methods for cold-start CDR.Its listed contributions include disentangled intent representations enhanced by interaction graphs and intent-wise contrastive learning.Source: Introduction · p. 2
Related Work
Related work spans cross-domain recommendation and disentangled representation learning. The paper positions DisCo as capturing finer-grained user intent to filter irrelevant collaborative information.
Related Work
- Cross-domain recommendation includes intra-domain methods that transfer information to address sparsity and inter-domain methods that recommend to cold-start users without target-domain interactions.The section names matrix-factorization, deep-transfer, embedding-mapping, and personalized meta-learning approaches.Source: Cross-Domain Recommendation · p. 2
- Disentangled representation learning factorizes representations to distinguish underlying explanatory factors, with applications spanning vision, language, graphs, and recommendation.Examples include factor-wise contrastive graph learning and recommendation models for hierarchical intentions or relation aspects.Source: Disentangled Representation Learning · pp. 2–3
- Prior disentangled CDR methods separate domain-shared and domain-specific preferences, focusing on identifying and transferring shared aspects.The paper instead proposes capturing more detailed user intent and filtering irrelevant collaborative information.Source: Disentangled Representation Learning · p. 3
Notations and Problem Definition
The paper represents source and target domains as user-item interaction graphs, with an overlapping user subset. Cold-start CDR predicts target-domain item engagement for users observed only in the source domain.
Notations and Problem Definition
- Each domain is represented by users, items, and interactions, and its binary interaction matrix defines a user-item bipartite graph.The source and target user sets each include an overlapping subset, alongside their respective non-overlapping users.Source: Notations · p. 3
- The source and target user sets share an overlapping subset, while their remaining users are domain-specific.The notation distinguishes the non-overlapping source users from the non-overlapping target users.Source: Notations · p. 3
- Cold-start CDR predicts target-domain item engagement for a user observed only in the source domain, using disentangled representations that capture diverse intents.The prediction function takes the user, target item, model parameters, and both domain graphs as inputs.Source: Problem Definition · p. 3
The Proposed Framework
DisCo combines domain-specific intent extraction, graph-based user-similarity learning, and intent-wise cross-domain contrast to transfer relevant information while retaining domain-specific preferences. The resulting disentangled intents support recommendation for cold-start users.
The Proposed Framework
- DisCo combines domain-specific intent extraction, multi-step affinity-graph walks, and intent-wise contrastive learning to transfer relevant similarities and predict recommendations for cold-start users.The framework is designed to filter irrelevant source-domain collaborative information.Source: Overview · p. 3
- A graph-disentangled layer follows traditional message passing and uses separate channels to extract detailed user intents, with K denoting the number of intents.The encoder processes user-item interaction graphs in each domain.Source: Disentangled User Intent Graph Encoder · p. 3
- Online and momentum-based target encoders generate per-domain intent embeddings, while multi-step random walks over user-affinity graphs provide high-order similarities for intra-domain contrastive learning.The target encoder is updated by exponential moving average; similarity serves as a pseudo target, with orthogonality encouraging distinct intent representations.Source: Intra-Domain Contrast · pp. 3–4
- For overlapping users, a decoder projects source intent embeddings toward target embeddings while target-domain similarities aggregated across intents supervise inter-domain contrast.This design aims to preserve target-specific collaborative information while capturing domain-shared information.Source: Inter-Domain Contrast · p. 4
- A variational EM procedure estimates latent-intent weights for user-pair similarity, using a mini-batch variational distribution because exact posterior computation would require considering all users.The E-step infers the variational distribution and the M-step optimizes the ELBO.Source: Inter-Domain Contrast · pp. 4–5
- Intent-specific user-item inner products are weighted into an overall preference score, and recommendation training combines source- and target-domain prediction losses with intra- and inter-domain contrastive and orthogonality losses.The recommendation loss uses binary cross entropy and negative sampling; hyperparameters weight the loss components.Source: User Intent Adaptation and Prediction · p. 5, Objective Function · p. 5
Experiment
Experiments evaluate DisCo against recommendation baselines, test its components and hyperparameters, and examine a two-user case study. DisCo excels on game-video and music-movie, remains competitive on sport-cloth and phone-electronic, and mitigates negative transfer in the case study.
Experimental Setup
- The evaluation uses four Amazon domain pairs, holds out 20% of overlapping users as cold-start users, and measures leave-one-out top-10 recommendation with NDCG@10 and HR@10.Baselines include single-domain CF, embedding-and-mapping CDR, and disentangled-representation models.Source: Datasets · p. 5, Evaluation · p. 5, Baselines · p. 5
Performance Analysis
- Across four CDR scenarios, DisCo excels on game-video and music-movie and remains competitive on sport-cloth and phone-electronic against graph-encoder models CDRIB and UniCDR.It also outperforms all other EMCDR-based methods.Source: Performance Analysis · p. 6, Table 1 · p. 6
Ablation Study
- Ablations associate the mapping function, user-specific intent weights, orthogonal intents, and affinity-graph random walks with improved prediction or filtering of irrelevant collaborative information.Removing the cross-domain decoder, using uniform intent weights, omitting the intra-domain contrast loss, or removing affinity graphs and random walks each degrades performance.Source: Ablation Study · p. 6, Table 2 · p. 6
Parameter Analysis
- The best intent count varies by domain, denser video and sport domains benefit from more intents, higher-order walks generally outperform lower-order ones, and d=3–5 performs well.Excessively high λ reduces performance, while the optimal β is usually 0.2–0.4.Source: Parameter Analysis · p. 7, Figure 3 · p. 7, Figure 4 · p. 7
Case Study
- In the music-movie case study, two users with similar movie-source preferences differ in target music preferences, and DisCo mitigates negative transfer where CDRIB does not.The table marks predictions matching target-domain test ground truth in bold.Source: Case Study · p. 7, Table 3 · p. 7
Conclusion
DisCo addresses negative transfer in cold-start cross-domain recommendation by capturing diverse user intents and aligning intent-wise user similarities. Experiments on four benchmark datasets demonstrate its efficacy.
Conclusion
- Experiments on four benchmark datasets demonstrate the efficacy of DisCo for cold-start cross-domain recommendation.Source: Conclusion · p. 7
- DisCo uses a multi-channel graph encoder to capture diverse user intents and multi-step random walks to learn high-order user similarities under each intent.Source: Conclusion · p. 7
- Its intent-wise disentangled contrastive learning retains user-similarity relationships while identifying rationale between source and target domains.Source: Conclusion · p. 7