PaperlayerView exact revision on arXiv

Source-linked AI summary

Cross-domain recommendation via user interest alignment

Chuang Zhao, Hongke Zhao, Ming He, Jian Zhang, Jianping Fan

arXiv:2301.11467v1cs.IRcs.AI

TL;DR

Cross-domain recommendation often separately trains and aggregates domain representations, leaving entity similarity and user-interest alignment unaddressed. COAST combines a unified graph with user-user and user-item interest alignment; the authors report significant improvements over state-of-the-art algorithms on all tasks. They identify validation on more large cross-domain datasets as future work.

  • Problem

    Overlapping-representation methods often train domain embeddings separately and aggregate them, missing cross-domain entity similarity and failing to guarantee aligned user interests amid sparsity and cold-start challenges.

    Source: §1 · pp. 1–2

  • Method

    COAST builds a unified cross-domain heterogeneous graph and aligns overlapping users’ interests through user-user contrastive learning and user-item gradient alignment.

    Source: §1 · p. 2, §3.3.2 · p. 4, §3.4.2 · p. 5

  • Results

    On all tasks, COAST achieves significant improvements over state-of-the-art algorithms for real-world recommendations.

    Source: §1 · p. 2

  • Takeaways & Limitations

    The authors frame partial-user-overlap datasets using explicit and implicit information as a setting with a wider range of application scenarios.

    Source: §5 · p. 9

    The authors identify robustness validation on more large cross-domain recommendation datasets as future work.

    Source: §5 · p. 9

Abstract

from arXiv · show

Cross-domain recommendation aims to leverage knowledge from multiple domains to alleviate the data sparsity and cold-start problems in traditional recommender systems. One popular paradigm is to employ overlapping user representations to establish domain connections, thereby improving recommendation performance in all scenarios. Nevertheless, the general practice of this approach is to train user embeddings in each domain separately and then aggregate them in a plain manner, often ignoring potential cross-domain similarities between users and items. Furthermore, considering that their training objective is recommendation task-oriented without specific regularizations, the optimized embeddings disregard the interest alignment among user's views, and even violate the user's original interest distribution. To address these challenges, we propose a novel cross-domain recommendation framework, namely COAST, to improve recommendation performance on dual domains by perceiving the cross-domain similarity between entities and aligning user interests. Specifically, we first construct a unified cross-domain heterogeneous graph and redefine the message passing mechanism of graph convolutional networks to capture high-order similarity of users and items across domains. Targeted at user interest alignment, we develop deep insights from two more fine-grained perspectives of user-user and user-item interest invariance across domains by virtue of affluent unsupervised and semantic signals. We conduct intensive experiments on multiple tasks, constructed from two large recommendation data sets. Extensive results show COAST consistently and significantly outperforms state-of-the-art cross-domain recommendation algorithms as well as classic single-domain recommendation methods.

1 INTRODUCTION

COAST addresses shortcomings of cross-domain recommendation based on separately trained, plainly aggregated representations by modeling cross-domain entity similarity and aligning overlapping users’ interests. It uses a unified graph and user-user and user-item alignment to support recommendation with partial user overlap.

Motivation

  • Prior overlapping-representation methods train domain representations separately and aggregate them, which can miss higher-order entity similarities and fail to align users’ interests across domains [1].Common aggregations include concatenation, dot products, and pooling.Source: §1 · p. 2
  • Most studies use explicit datasets with fully overlapping users, unlike the partial overlap and rich implicit content the paper targets.The introduction frames this as a mismatch with real-world scenarios.Source: §1 · p. 2

Contributions

  • COAST combines a unified cross-domain heterogeneous graph with user-user and user-item interest alignment to improve dual-domain recommendation with partial user overlap.The framework uses content features to form an implicit dataset and a message-passing mechanism to capture cross-domain similarity.Source: §1 · p. 2
  • The paper reports significant improvements over state-of-the-art algorithms on all tasks.It also identifies dual knowledge transfer based on partial user overlap as a contribution.Source: §1 · p. 2

2 RELATED WORK

The related work reviews cross-domain recommendation paradigms and contrastive learning, positioning COAST as an approach that combines richer content, cross-domain graph modeling, and interest alignment.

2 RELATED WORK

  • COAST’s related-work review spans cross-domain recommendation and contrastive learning, the two areas from which its framework draws.The paper summarizes their paradigms, advantages, disadvantages, and connections to its approach.Source: §2 · p. 2

2.1 Cross-domain Recommendation

  • Prior cross-domain approaches include matrix factorization, mapping [45], knowledge graphs, and combinations of overlapping-entity representations [44].The review notes limitations involving mapping generalizability, graph-method scalability, and the lack of cross-domain similarity in plain representation combinations.Source: §2.1 · p. 2
  • COAST adds content information, a unified cross-domain graph, user-interest alignment, and end-to-end optimization to the overlapping-representation paradigm.The authors say end-to-end optimization avoids potential target inconsistency from two-stage training.Source: §2.1 · pp. 2–3

2.2 Contrastive Learning

  • Contrastive learning aligns views of an instance; COAST draws on clustering to bring different views of the same user toward a shared interest center.The review also describes memory banks and momentum-encoder queues as approaches to scaling negative-example storage.Source: §2.2 · p. 3

3 PROPOSED METHOD

The method formulates dual-domain recommendation with overlapping users and observed interactions, then uses interaction and content data to transfer knowledge and recommend items in both domains.

3 PROPOSED METHOD

  • COAST’s method section introduces the general cross-domain recommendation problem, outlines the framework, and then details its modules and optimization.The stated sequence is problem definition, framework overview, and module and optimization details.Source: §3 · p. 3

3.1 Problem Formulation

  • The formulation assumes two domains, a richer-interaction source and a sparser target, with user sets that share an overlapping subset.It represents each domain’s user-item interactions with a binary matrix whose entries indicate observed interactions.Source: §3.1 · p. 3
  • Given interactions and content from both domains, dual cross-domain recommendation transfers knowledge through overlapping users to recommend items in each domain.The stated objective is to improve recommendation performance in both domains.Source: §3.1 · p. 3

3.2 Overview of COAST Framework

COAST’s overview combines cross-domain graph propagation, overlapping-user interest alignment, and recommendation losses in a joint optimization.

3.2 Overview of COAST Framework

  • A unified cross-domain heterogeneous graph and modified graph-convolution message passing capture similarities between users and items across domains.The framework adds cross-domain similarity to the representation-learning process.Source: §3.2 · p. 3
  • For overlapping users, COAST applies contrastive learning and gradient alignment to align interests from user-user and user-item perspectives.These are the two alignment perspectives named in the framework overview.Source: §3.2 · p. 3
  • Negative-sampling supervision losses for both domains are jointly optimized with the alignment losses.The overview places this after graph construction and user-interest alignment.Source: §3.2 · p. 3

3.3 Cross-domain Graph Convolution

COAST replaces separately trained domain representations with a unified cross-domain heterogeneous graph and message passing designed to capture similarity between users and items across domains.

3.3.1 Construction.

  • A unified cross-domain heterogeneous graph and message-passing mechanism are designed to capture cross-domain similarity between users and items.The graph is built from the two domains’ user-item interactions.Source: §3.3 · p. 3, §3.3.1 · p. 3, §3.3.2 · p. 4
  • Numerical and category attributes are normalized or one-hot encoded, while text attributes are converted to doc2vec vectors with users jointly encoded across domains.For overlapping users, representations from both domains are aggregated by max pooling; sum and averaging showed no significant improvement.Source: §3.3.1 · p. 4

3.3.2 Propagation.

  • Cross-domain user-item interactions are added to graph-convolution messages to enrich embeddings and capture cross-domain collaborative signals.Stacking GNN layers is used to perceive higher-order similarities.Source: §3.3.2 · p. 4
  • The approach forms one graph for interactions across domains and generalizes message passing to cross-domain scenarios.The layer representations are concatenated for users and items.Source: §3.3.2 · p. 4

3.4 User Interest Alignment

COAST addresses plain cross-domain representation aggregation by aligning user interests from user-user and user-item perspectives. It uses contrastive views and gradient alignment to constrain interest consistency across domains.

3.4 User Interest Alignment

  • Plain aggregation of domain representations can ignore user-interest alignment, motivating constraints from user-user and user-item perspectives.These constraints are intended to constrain user representations.Source: §3.4 · p. 4

3.4.1 User-User Alignment.

  • Aggregating users’ second-order neighbors separately in each domain creates contrastive views, whose interest distributions are encouraged to be consistent.The approach uses unsupervised contrastive learning to encourage the views to have the same clustering over interest distributions.Source: §3.4.1 · p. 4

3.4.2 User-Item Alignment.

  • User-item alignment brings different user views closer to interacted-item representations and uses gradient alignment to make their optimization paths consistent.The paper presents gradient alignment as using semantic information; together with contrastive learning, alignment acts as a regularizer against overfitting user representations.Source: §3.4.2 · p. 5

3.5 Model Optimization

COAST jointly optimizes supervised recommendation and two interest-alignment constraints across both domains. Its end-to-end solution is intended to improve both domains while aligning overlapping users’ interests.

3.5 Model Optimization

  • A dual-tower MLP structure captures high-order representations of users and items.Its layer widths are [D, 2D, 4D, 8D, 4D, 2D, D].Source: §3.5.1 · p. 5

3.5.1 Supervised Estimation.

  • Supervised estimation uses randomly selected unobserved user-item interactions as negative instances and optimizes with binary cross-entropy.The negative instances are denoted Y−.Source: §3.5.1 · p. 5
  • The supervised loss is optimized in both domains simultaneously.Source: §3.5.1 · p. 5

3.5.2 Total Loss

  • The total loss combines supervised loss with weighted user-user and user-item alignment constraints, aiming to improve both domains while aligning overlapping users’ interests.λ2 weights the two alignment constraints.Source: §3.5.2 · p. 5

4 EXPERIMENTS

Across four recommendation tasks, COAST outperforms the best baseline, while overlap tests and ablations report robustness and contributions from its graph and interest-alignment components. Experiments use Douban and industrial-platform data, with performance evaluated using Hit Ratio and NDCG.

4.1.1 Data Sets.

  • Experiments cover movie-book, movie-music, book-music, and mall-community tasks, using shared-user links across Douban and industrial-platform domains.The data include user and item content features; users and items with fewer than five interactions are filtered.Source: §4.1.1 · pp. 5–6

4.1.2 Parameter Settings.

  • Evaluation uses leave-one-out testing with 99 sampled negatives and Hit Ratio and NDCG at @10; comparisons include single- and cross-domain baselines.The baseline set includes NMF, LightGCN, MVDNN, DTCDR, DDTCDR, DML, GADTCDR, and CDRIB.Source: §4.1.2 · p. 6, §4.1.3 · p. 6

4.2 Comparison with Baselines (RQ1)

  • Across the four recommendation tasks, COAST achieved 0.32%–10.22% improvement over the best-performing baseline on each task, according to the reported results.Source: §4.2 · p. 6
  • Cross-domain methods outperform single-domain methods in most tasks, and methods using implicit content features outperform interaction-only models.Representation-combination methods also outperform mapping-based methods in the reported experiments.Source: §4.2 · p. 6

4.3.1 Length N.

  • COAST performs best across tested recommendation-list lengths against LightGCN and GADTCDR, with its greatest improvement at N=3.Performance rises for all algorithms as list length increases.Source: §4.3.1 · p. 7, Figure 5 · p. 7

4.3.2 Overlap Ratio M.

  • COAST remains more robust than GADTCDR with 25% user overlap; increasing overlap improves performance, with smaller gains from 75% to 100% than from 25% to 50%.The authors attribute the smaller later gain to the absolute number of overlapping users being sufficient for basic knowledge transfer.Source: §4.3.2 · p. 7, Table 3 · p. 7

4.3.3 Ablation Studies.

  • Ablations report lower performance without the unified graph or either user-interest consistency constraint, while the explicit-only variant still exceeds most baselines.The study attributes graph-module losses to missing cross-domain or user-item interaction signals, and reports alignment as supporting generalization.Source: §4.3.3 · pp. 7–8, Figure 6 · p. 8

4.3.4 Hyper-testing.

  • The reported choices are D=64, K=256, and λ2=0.01; larger λ2 strengthens alignment but can impair domain-specific recommendation.D=64 performs best on all reported metrics; K is sensitive, and the authors describe it as an abstract interest center.Source: §4.3.4 · p. 8, Figure 7 · p. 8, Figure 8 · p. 8

5 CONCLUSION

COAST combines cross-domain similarity and user-interest alignment to improve dual-domain recommendation through bidirectional knowledge transfer. Its experiments report state-of-the-art performance, while the authors identify broader scenarios and datasets for future validation.

5 CONCLUSION

  • COAST combines a unified heterogeneous graph and revised message passing with contrastive learning and gradient alignment for cross-domain similarity and user-interest alignment.The alignment targets overlapping users from user-user and user-item perspectives.Source: §5 · pp. 8–9
  • The framework uses partial user overlap and both explicit and implicit information, which the authors describe as supporting a wider range of application scenarios.The authors also identify richer content information and interest alignment as ingredients in bidirectional knowledge transfer.Source: §5 · pp. 8–9
  • Extensive experiments are reported to demonstrate COAST's state-of-the-art performance and superiority.Source: §5 · p. 9
  • Future work includes overlapping-item and multi-domain scenarios, additional modalities or interactions, and robustness validation on more large cross-domain datasets.Source: §5 · p. 9

A IMPORTANT NOTATIONS

The section is identified as a mathematical-notation table.

A IMPORTANT NOTATIONS

B ALGORITHM

Algorithm 1 takes interaction matrices as input and returns model parameters. Its loop constructs and propagates a graph, computes alignment and supervised losses, and updates parameters until convergence.

B ALGORITHM

  • The algorithm takes interaction matrices A_S, A_T, X_S, and X_T as input and returns parameters Θ.Source: B ALGORITHM · p. 11
  • After parameter initialization and data preprocessing, it constructs the graph G = (U, V, E, H) and performs graph propagation within its training loop.Source: B ALGORITHM · p. 11
  • For overlapping users, the loop computes user-user and user-item interest-alignment losses, combines them with supervised loss, and updates parameters until convergence.Source: B ALGORITHM · p. 11
Loading 2301.11467v1…