Research map / Graph learning / Graph self-supervised learning
Behavior Importance-Aware Graph Neural Architecture Search for Cross-Domain Recommendation
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes BiGNAS, a bi-level framework that jointly searches a one-shot GNN architecture per domain (supernetwork) and learns source-behavior importance weights (graph-based perceptron) for cross-domain CTR prediction.
Problem
Cross-domain recommendation with GNNs relies on manually designed, suboptimal architectures and suffers negative transfer from noisy source-domain behaviors.
Approach
Combines a retrain-free one-shot GNN architecture supernetwork customized per domain with an outer-level graph-based perceptron that assigns importance weights to source-domain interaction losses, optimized via bi-level optimization with implicit gradients.
Key results
- BiGNAS reports AUC gains up to 10.57% over the best baseline on Amazon (El-Cl) and 4.61% on the WeChat Discover task.
- Ablation: removing source-domain data (NO-SOURCE) drops performance by 5.8% on average; removing the importance perceptron (NO-IMPO) drops it by 4.3%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Remove source-domain data (ablation) | BiGNAS · Four Amazon tasks | performance: average performance drop of 5.8% | Hurt |
| Remove the Graph-Based Behavior Importance Perceptron (ablation) | BiGNAS · Four Amazon tasks | performance: decline of 4.3% | Hurt |
| Replace customized architecture search with equal weights for all candidate operations (ablation) | BiGNAS · Four Amazon tasks | recommendation performance | Hurt |
| Replace continuous relaxation with discrete architectures (ablation) | BiGNAS · Four Amazon tasks | recommendation performance | Hurt |
| Use one fixed manually designed GNN architecture across domains (ablation) | BiGNAS · Four Amazon tasks | recommendation performance | Hurt |
Surprising findings
- The BiGNAS LogLoss is not uniformly best; on Bo-To the strongest baseline HeroGRAPH has lower LogLoss than BiGNAS, reflecting AUC/LogLoss inconsistency.
Limitations
- Multi-domain experiments are limited to the proprietary WeChat dataset, which is not released.
- Amazon multi-domain settings were not used due to limited user overlap.
Experimental setup
Datasets: Amazon Product 5-core (7 domain pairs) WeChat advertising dataset (Channel, Moments, Discover)
Benchmarks: Amazon Product 5-core cross-domain pairs WeChat advertising dataset
Metrics: AUC LogLoss RelaImpr
Baselines: BPR-MF NGCF LightGCN CoNet-B MiNet-B HeroGRAPH DisenCDR EDDA
Builds on: DARTS BIAO LightGCN NGCF EDDA
Closest papers by meaning
- Graph Disentangled Contrastive Learning with Personalized Transfer for Cross-Domain Recommendation (AAAI 2024 · similarity 0.85)
- LS-TGNN: Long and Short-Term Temporal Graph Neural Network for Session-Based Recommendation (AAAI 2025 · similarity 0.85)
- DisCo: Graph-Based Disentangled Contrastive Learning for Cold-Start Cross-Domain Recommendation (AAAI 2025 · similarity 0.85)
- Learning Accurate and Bidirectional Transformation via Dynamic Embedding Transportation for Cross-Domain Recommendation (AAAI 2024 · similarity 0.85)
- Mind Individual Information! Principal Graph Learning for Multimedia Recommendation (AAAI 2025 · similarity 0.84)
- Dynamic Multi-Interest Graph Neural Network for Session-Based Recommendation (AAAI 2025 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.