Research map / Graph learning / Graph self-supervised learning
Link Prediction in Multilayer Networks via Cross-Network Embedding
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 CGNN, a GNN-based cross-network embedding model that shares a latent space across two partially aligned network layers, using cross-layer message passing plus random-walk and anchor-aware losses for link prediction.
Problem
Link prediction in multilayer networks where layers only partially share nodes (partial interlayer alignment), where prior multiplex methods assume complete node correspondence and single-layer methods ignore cross-layer information.
Approach
A cross-GNN layer lets each node aggregate messages from its same-layer neighbors and from its aligned node in the other layer (weighted by a sigmoid similarity), with shared weights; training combines a random-walk proximity loss with a cosine-based anchor loss that pulls aligned nodes together and pushes hard negatives apart.
Key results
- CGNN reaches AUC 0.9653 on Facebook and 0.9740 on Twitter with 90% interlayer links, outperforming CELP (0.9588 on Facebook side) and other baselines.
- With 0% interlayer links on Facebook, CGNN achieves AUC 0.8989 versus 0.8150 for CELP and 0.8352 for GAE, showing gains with few interlayer links.
- Ablation removing the intralayer loss drops Facebook AUC from 0.9653 to 0.5639, the largest drop among the variants.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use CGNN instead of CELP | Facebook/Twitter · Facebook layer, 90% interlayer links | AUC: 0.9588 -> 0.9653 | Helped |
| use CGNN instead of CELP | Facebook/Twitter · Twitter layer, 90% interlayer links | AUC: 0.9684 -> 0.9740 | Helped |
| use CGNN instead of CELP | Twitter/YouTube · Twitter layer, 90% interlayer links | AUC: 0.9042 -> 0.9353 | Helped |
| use CGNN instead of LPIS | Twitter/YouTube · YouTube layer, 0% interlayer links | AUC: 0.8975 -> 0.9095 | Helped |
| add a cross-GNN layer rather than using a GCN-form layer (ablation) | CGNN · Facebook/Twitter (Facebook layer), 90% interlayer links | AUC: 0.8967 -> 0.9653 | Helped |
| add degree-discounted neighbor weights (ablation) | CGNN · Facebook/Twitter (Facebook layer), 90% interlayer links | AUC: 0.9281 -> 0.9653 | Helped |
| add intralayer loss (ablation) | CGNN · Facebook/Twitter (Facebook layer), 90% interlayer links | AUC: 0.5639 -> 0.9653 | Helped |
| add interlayer loss (ablation) | CGNN · Facebook/Twitter (Facebook layer), 90% interlayer links | AUC: 0.8859 -> 0.9653 | Helped |
| share weights in the second layer (ablation) | CGNN · Facebook/Twitter (Facebook layer), 90% interlayer links | AUC: 0.9629 -> 0.9653 | Helped |
| use negative sampling in interlayer loss (ablation) | CGNN · Facebook/Twitter (Facebook layer), 90% interlayer links | AUC: 0.9648 -> 0.9653 | Helped |
Surprising findings
- Setting alpha=0 (interlayer loss only) yields extremely low performance, indicating intralayer loss is essential.
- Removing negative sampling in the interlayer loss barely changes performance (0.9648 vs 0.9653 on Facebook), suggesting limited contribution.
Limitations
- Evaluated only on a few small social network datasets (Facebook/Twitter, Twitter/YouTube) with two layers
- Relies on known partial interlayer alignments and assumes structural consistency across layers
- Complexity analysis is given but no runtime comparison is reported
Experimental setup
Datasets: Facebook Twitter YouTube
Benchmarks: Facebook/Twitter Twitter/YouTube
Metrics: AUC
Baselines: CN RA SVD node2vec GAE GAT MAA LPIS CLF CELP EG-mini n2v-e
Builds on: GCN GraphSAGE node2vec CELP
Closest papers by meaning
- Cross-links Matter for Link Prediction: Rethinking the Debiased GNN from a Data Perspective (NeurIPS 2023 · similarity 0.89)
- Non-Euclidean Mixture Model for Social Network Embedding (NeurIPS 2024 · similarity 0.86)
- Pure Message Passing Can Estimate Common Neighbor for Link Prediction (NeurIPS 2024 · similarity 0.86)
- A Topological Perspective on Demystifying GNN-Based Link Prediction Performance (ICLR 2024 · similarity 0.86)
- Revisiting Link Prediction: a data perspective (ICLR 2024 · similarity 0.85)
- Networked Inequality: Preferential Attachment Bias in Graph Neural Network Link Prediction (ICML 2024 · similarity 0.85)
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.