Research map / Label-efficient and robust learning / Recognition and long-tail
T-NET: Weakly Supervised Graph Learning for Combatting Human Trafficking
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 T-NET, a weakly supervised graph neural network that classifies clusters of escort ads (HT, ISW, spam) using labeling-function weak labels and contrastive learning, and releases the synthetic ASW-SYNTH dataset.
Problem
Detecting human trafficking activity in adult service website escort ads is hard because labels are costly and slow to obtain, and prior work classifies ads independently and conflates trafficking with independent sex work and spam.
Approach
Cluster ads into micro-clusters, link clusters by shared metadata into a graph, and train a GNN on programmatic weak labels from labeling functions, with node-weighted classification loss and a graph contrastive learning term based on Louvain communities.
Key results
- T-NET achieves weighted average F1 of 0.74 on ASW-REAL, about 7% better than the next best baseline (PI-GNN, 0.67).
- On ASW-SYNTH, T-NET reaches weighted average F1 of 0.78 versus 0.55 for GCN, roughly 23% better than the next best baseline.
- T-NET has the lowest ISW-to-HT conflict while keeping the highest HT accuracy in the ASW-SYNTH analysis.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use T-NET instead of PI-GNN | ASW-REAL node classification · ASW-REAL | weighted average F1 (β = 1): 0.67 ± 0.09 -> 0.74 ± 0.16 | Helped |
| use T-NET instead of GCN | ASW-SYNTH node classification · ASW-SYNTH | weighted average F1 (β = 1): 0.55 ± 0.27 -> 0.78 ± 0.14 | Helped |
| add contrastive learning (ablation) | T-NET on ASW-REAL · ASW-REAL | weighted average F1 (β = 1): 0.57 ± 0.08 -> 0.74 ± 0.16 | Helped |
| use learned node weights rather than setting all node weights to 1 (ablation) | T-NET on ASW-REAL · ASW-REAL | weighted average F1 (β = 1): 0.59 ± 0.12 -> 0.74 ± 0.16 | Helped |
| include weak-label entropy in node weighting (ablation) | T-NET on ASW-REAL · ASW-REAL | weighted average F1 (β = 1): 0.69 ± 0.13 -> 0.74 ± 0.16 | Helped |
| include node-embedding centrality in node weighting (ablation) | T-NET on ASW-REAL · ASW-REAL | weighted average F1 (β = 1): 0.59 ± 0.16 -> 0.74 ± 0.16 | Helped |
| add contrastive learning (ablation) | T-NET on ASW-SYNTH · ASW-SYNTH | weighted average F1 (β = 1): 0.73 ± 0.24 -> 0.78 ± 0.14 | Helped |
| use learned node weights rather than setting all node weights to 1 (ablation) | T-NET on ASW-SYNTH · ASW-SYNTH | weighted average F1 (β = 1): 0.62 ± 0.19 -> 0.78 ± 0.14 | Helped |
| include weak-label entropy in node weighting (ablation) | T-NET on ASW-SYNTH · ASW-SYNTH | weighted average F1 (β = 1): 0.61± 0.22 -> 0.78 ± 0.14 | Helped |
| include node-embedding centrality in node weighting (ablation) | T-NET on ASW-SYNTH · ASW-SYNTH | weighted average F1 (β = 1): 0.75 ± 0.11 -> 0.78 ± 0.14 | Helped |
Surprising findings
- PI-GNN outperforms T-NET on the Spam class on ASW-REAL, attributed to its loss handling class imbalance and disconnected nodes.
- Weak labeling functions have low coverage (about 10.7% on ASW-REAL) and accuracy around 53%, showing weak labels are noisy.
Limitations
- Real-world ASW-REAL dataset cannot be shared due to privacy and sensitivity, so only the synthetic corpus is available.
- Labeling functions are hand-designed and may need updating as trafficking patterns change.
- Spam class performance is weaker than PI-GNN on ASW-REAL, with only 9 spam examples.
Experimental setup
Datasets: ASW-REAL ASW-SYNTH
Benchmarks: ASW-REAL ASW-SYNTH
Metrics: F-beta score (beta=0.5, 1) weighted average F score HT accuracy ISW conflict
Baselines: MLP GCN NRGNN PI-GNN
Builds on: GCN InfoShield NRGNN PI-GNN programmatic weak supervision
Closest papers by meaning
- Barely Supervised Learning for Graph-Based Fraud Detection (AAAI 2024 · similarity 0.84)
- ScamNet: Toward Explainable Large Language Model-Based Fraudulent Shopping Website Detection (AAAI 2025 · similarity 0.83)
- DeDe: Detecting Backdoor Samples for SSL Encoders via Decoders (CVPR 2025 · similarity 0.82)
- CausalDiff: Causality-Inspired Disentanglement via Diffusion Model for Adversarial Defense (NeurIPS 2024 · similarity 0.81)
- From Toxic to Trustworthy: Using Self-Distillation and Semi-supervised Methods to Refine Neural Networks (AAAI 2024 · similarity 0.81)
- Context-aware Graph Neural Network for Graph-based Fraud Detection with Extremely Limited Labels (AAAI 2025 · similarity 0.81)
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.