Source-linked AI summary
DA-GCN: A Domain-aware Attentive Graph Convolution Network for Shared-account Cross-domain Sequential Recommendation
Lei Guo, Li Tang, Tong Chen, Lei Zhu, Quoc Viet Hung Nguyen, Hongzhi Yin
TL;DR
SCSR is challenging because shared accounts mix multiple users’ interests across domains, while prior methods inadequately capture entity relationships and explicit cross-domain structure. DA-GCN constructs a cross-domain graph, learns user-specific representations with domain-aware graph convolution, and uses attention to guide message passing. Experiments on two real-world datasets demonstrate DA-GCN’s superiority over several baselines.
Problem
SCSR lacks methods that adequately model mixed user preferences, relationships among multiple entities, and explicit cross-domain graph structure.
Method
DA-GCN constructs a CDS graph and applies domain-aware graph convolution with two attention mechanisms to learn user-specific node representations.
Results
DA-GCN demonstrates superiority over several state-of-the-art baselines in experiments on two real-world datasets.
Takeaways & Limitations
Explicit cross-domain graph structure and attentively weighted message passing provide the paper’s graph-based solution for modeling users, items, and their interactions in SCSR.
Takeaways & Limitations
The model assumes H latent users under each shared account even though the number and identities of those users are unknown.
Abstract
from arXiv · showhide
Shared-account Cross-domain Sequential recommendation (SCSR) is the task of recommending the next item based on a sequence of recorded user behaviors, where multiple users share a single account, and their behaviours are available in multiple domains. Existing work on solving SCSR mainly relies on mining sequential patterns via RNN-based models, which are not expressive enough to capture the relationships among multiple entities. Moreover, all existing algorithms try to bridge two domains via knowledge transfer in the latent space, and the explicit cross-domain graph structure is unexploited. In this work, we propose a novel graph-based solution, namely DA-GCN, to address the above challenges. Specifically, we first link users and items in each domain as a graph. Then, we devise a domain-aware graph convolution network to learn user-specific node representations. To fully account for users' domain-specific preferences on items, two novel attention mechanisms are further developed to selectively guide the message passing process. Extensive experiments on two real-world datasets are conducted to demonstrate the superiority of our DA-GCN method.
1 Introduction
SCSR recommends next items when multiple users share an account across domains, making preference modeling difficult because interactions mix diverse interests and amplify noise. DA-GCN addresses this with an explicit cross-domain graph, domain-aware convolution, attention mechanisms, and experiments on two datasets.
- Motivation: SCSR recommends the next item from interaction histories spanning multiple domains when multiple users share one account.Shared accounts can combine diverse users’ interests within a single interaction sequence.
- Research gap: Few prior methods jointly address shared-account and cross-domain recommendation, while existing approaches overlook explicit structural links between domains.Prior graph-based work may ignore sequential information or require explicit ratings unavailable in both domains.
- Method: DA-GCN constructs a CDS graph whose user and item nodes across domains are connected by association edges.The graph explicitly links accounts and items from two domains.
- Method: DA-GCN uses domain-aware graph convolution and attention mechanisms to learn expressive, user-specific representations for account-sharing users and items.The model assumes H latent users under each account and selectively weights information passed from neighbors.
- Evaluation: Extensive experiments on two real-world datasets report that DA-GCN is superior to several state-of-the-art baselines.The reported comparisons evaluate the model on HVIDEO and HAMAZON.
2 Related Work
Related work covers cross-domain recommendation, shared-account recommendation, and graph-based recommender systems. These lines of research address knowledge transfer, mixed-user behavior, or graph propagation, but are presented as distinct areas.
- Cross-domain recommendation: Cross-domain recommendation methods aggregate or transfer knowledge across domains to address cold-start and data sparsity issues.The literature includes both traditional and deep learning-based approaches.
- Shared-account recommendation: Shared-account recommendation models mixtures of user behaviors, often identifying users before generating recommendations.Prior work uses signals such as user preference and consumption time for identification.
- Graph-based recommendation: Graph-based recommender systems propagate embeddings over interaction graphs to update user and item representations efficiently.Examples include item-item graph convolution in PinSage and high-order proximity modeling in NGCF.
3 Methodologies
DA-GCN formulates shared-account cross-domain sequential recommendation as graph-based next-item prediction, modeling accounts through latent users and a cross-domain sequence graph. Domain-aware and sequence-aware attention guide message passing so representations reflect user-specific preferences and item-order relationships.
- Preliminaries: SCSR predicts the next item in either domain from the account’s historical behavior sequences.
- Cross-Domain Sequence Graph: DA-GCN constructs a cross-domain sequence graph linking users and items across domains and recursively aggregates local-neighbor information.
- Representation Learning with Latent Users: Each shared account is modeled with H latent users whose representations accumulate information from connected items instead of treating the account as one virtual user.
- Representation Learning with Latent Users: Domain-aware attention assigns importance to interacted items for each latent user, while attentive weights control message strength during aggregation across both domains.
- User-specific Item Representation Learning: Item representations aggregate connected users and sequentially related items, using sequence-aware attention to assess their relevance to the target item.
- The Prediction Layer: Sequence-level embeddings are obtained by max pooling item embeddings, and domain-specific next-item objectives are jointly trained with a multi-task negative log-likelihood loss.
4 Experiments
Experiments on HVIDEO and HAMAZON compare DA-GCN with diverse baselines and evaluate sequential information, attention, shared-account modeling, and training efficiency. DA-GCN achieves the strongest reported performance, while ablations and scalability tests support its design choices.
- Experimental Results: DA-GCN achieves the best performance on both HVIDEO domains and outperforms other baselines on HAMAZON in most metrics.The comparison evaluates MRR@5, Recall@5, MRR@20, and Recall@20 using held-out last items as ground truth.
- Experimental Results: SCSR-specific methods outperform other baselines, supporting the importance of modeling shared-account and cross-domain characteristics together.The baselines include traditional, shared-account, cross-domain, sequential, and shared-account cross-domain sequential recommenders.
- Experimental Results: DA-GCN outperforms RNN-based methods π-net and PSJNet, demonstrating stronger modeling of sequential information and complicated interaction relationships.It also outperforms the shared-account and cross-domain baselines VUI-KNN, NCF-MLP++, and Conet.
- Ablation Studies: DA-GCN outperforms GCNOSA and GCNOS, showing that sequential information improves user-preference modeling.GCNOS disables sequential information, while GCNOSA removes both sequential information and attention mechanisms.
- Ablation Studies: DA-GCN outperforms GCNOA, indicating that weighting passed messages differently through attention mechanisms improves node representations.GCNOA disables the attention mechanisms during message aggregation; similar ablation results are reported for HAMAZON, although only HVIDEO results are presented.
- Impact of H: Modeling an account as H latent users performs better than treating it as a single virtual user on HVIDEO.The experiment varies H to examine the shared-account characteristic and reports that the latent-user formulation is more aligned with the observed setting.
- Training Efficiency: DA-GCN has lower training time cost than π-net and PSJNet as the HVIDEO training ratio increases from 0.1 to 1.0.The result is presented as evidence of better scalability to large-scale datasets.