Research map / Graph learning / Knowledge and temporal graphs
Fine-Grained Graph Representation Learning for Heterogeneous Mobile Networks with Attentive Fusion and Contrastive Learning
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 DMGSL, an unsupervised graph structure learning framework that combines hierarchical edge-level attention, LSTM plus temporal multi-head attention over coherence-time snapshots, and contrastive learning to refine wireless data knowledge graphs.
Problem
Automatically constructing and refining heterogeneous, dynamic wireless data knowledge graphs (WDKGs) for mobile networks, which manual construction handles poorly and existing self-supervised GSL methods (designed for static homogeneous graphs) do not address.
Approach
Slice the wireless graph by edge type and fuse slices with edge-level attention, segment time into coherence-time snapshots processed by an LSTM and temporal multi-head attention, and train with a contrastive loss between anchor (expert) and learned graph views.
Key results
- DMGSL achieves 0.7000 accuracy and 0.6436 F1-score on uplink throughput node classification, versus 0.6125 accuracy for the best baseline Sublime.
- Removing the temporal attention module (w/o TAT) drops accuracy to 0.6125, while removing the hierarchical attention module (w/o HAT) yields 0.6625 accuracy.
- Performance peaks at feature masking rates r_l=0.8 and r_a=0.4 and at k=2 nearest data fields.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use DMGSL instead of the strongest baseline Sublime | node classification · Uplink throughput data (35min) | accuracy: 0.6125 +/- 0.0468 -> 0.7000 +/- 0.0250 | Helped |
| add temporal attention (TAT) (ablation) | DMGSL · Uplink throughput data (35min) | accuracy: 0.6125 +/- 0.0250 -> 0.7000 +/- 0.0250 | Helped |
| add temporal attention (TAT) (ablation) | DMGSL · Uplink throughput data (35min) | F1-score: 0.5538 +/- 0.0356 -> 0.6436 +/- 0.0321 | Helped |
| add hierarchical attention (HAT) (ablation) | DMGSL · Uplink throughput data (35min) | accuracy: 0.6625 +/- 0.0306 -> 0.7000 +/- 0.0250 | Helped |
| add hierarchical attention (HAT) (ablation) | DMGSL · Uplink throughput data (35min) | F1-score: 0.6104 +/- 0.0350 -> 0.6436 +/- 0.0321 | Helped |
| set learned-graph feature masking rate to r_l=0.8 and anchor-graph rate to r_a=0.4 | DMGSL · Uplink throughput data (35min) | node classification performance | Helped |
| set bootstrap update parameter τ to 0.99 | DMGSL · Uplink throughput data (35min) | contrastive loss convergence | Helped |
| use learning rate lr=10^-2 rather than lr=10^-3 | DMGSL · Uplink throughput data (35min) | convergence speed | Helped |
| set nearest-neighbor count k to 2 | DMGSL · Uplink throughput data (35min) | node classification performance | Helped |
Limitations
- Evaluated only on a single proprietary uplink throughput dataset from one vehicle-based measurement campaign.
- Only node classification is used as downstream evidence for structure quality; no direct structure ground truth is reported.
- Results have relatively large standard deviations on small test sets.
Experimental setup
Datasets: Uplink throughput data (35min) wireless dataset Uplink throughput data (15min) wireless dataset
Benchmarks: Uplink throughput data (35min) node classification
Metrics: accuracy precision recall F1-score
Baselines: IDGL IDGL-Anch SLAPS SLAPS-2s GEN Sublime
Builds on: IDGL SLAPS Sublime GEN GCN
Closest papers by meaning
- HGOT: Self-supervised Heterogeneous Graph Neural Network with Optimal Transport (ICML 2025 · similarity 0.85)
- Contrastive Auxiliary Learning with Structure Transformation for Heterogeneous Graphs (AAAI 2025 · similarity 0.84)
- HeterSEED: Semantics–Structure Decoupling for Heterogeneous Graph Learning under Heterophily (NeurIPS 2026 · similarity 0.84)
- Bootstrapping Heterogeneous Graph Representation Learning via Large Language Models: A Generalized Approach (AAAI 2025 · similarity 0.84)
- SSTAG: Structure-Aware Self-Supervised Learning Method for Text-Attributed Graphs (NeurIPS 2025 · similarity 0.83)
- An Efficient Subgraph-Inferring Framework for Large-Scale Heterogeneous Graphs (AAAI 2024 · similarity 0.83)
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.