Weak Supervision for Real World Graphs

Pratheeksha Nair, Reihaneh RabbanyarXiv 2025 · DOI · arXiv

TL;DR

This paper introduces WSNET, a novel framework that effectively leverages "weak" supervision signals (noisy, indirect labels) for graph representation learning. By combining programmatic weak supervision with graph contrastive learning, WSNET achieves state-of-the-art performance on challenging real-world graph tasks where obtaining clean labels is impractical.

What Problem Does This Paper Address?

Many real-world applications, such as detecting human trafficking or misinformation, involve complex graphs where obtaining accurate, comprehensive labels is extremely difficult and expensive. Traditional graph neural networks (GNNs) rely heavily on clean, abundant labels, limiting their applicability in these high-stakes domains. While weak supervision (using heuristics, distant supervision, etc.) offers a way to generate labels cheaply, these labels are inherently noisy and can even distort learned representations. This paper tackles the challenge of effectively utilizing these imperfect weak labels to train robust GNNs for node classification.

Methods

WSNET is a weakly supervised graph contrastive learning framework that integrates three key components:

  • Weak Label Classification Component: This part uses a cross-entropy loss to align predicted labels with majority-vote aggregated weak labels. Crucially, it introduces a weighting mechanism to account for label noise. Node weights are determined by the agreement among weak labels (higher agreement means more reliable labels) and the node's representativeness in the embedding space.
  • Weak Label Contrastive Component: This component leverages the InfoNCE contrastive learning loss. It aims to pull nodes with similar weak label distributions closer in the embedding space and push apart nodes with dissimilar distributions. The intuition is that similar weak label patterns often indicate nodes belonging to the same class, even if the labels themselves are noisy.
  • Structure-based Contrastive Component: This component enhances robustness by maximizing mutual information between a node's embedding and its graph community. It uses a contrastive loss that contrasts real and corrupted graph counterparts, acting as a regularizer to mitigate the impact of label noise and capture mesoscale patterns within the graph.

The final loss function is a linear combination of these three components.

Datasets

The authors evaluated WSNET on several datasets:

  • Real-world Use Cases:
    • LIAR-WS: A graph-based resource built on the LIAR dataset for misinformation detection. It includes original statements and paraphrases as nodes, with text embeddings as features. Weak labels were generated using GPT-4 and Claude 3-Haiku with RAG.
  • ASW-REAL and ASW-SYNTH: Datasets for detecting human trafficking from online escort advertisements. ASW-REAL is a real-world dataset, while ASW-SYNTH is a synthetic version. These graphs connect ads via shared contact information, and node features include cluster-level statistics. Weak labels were generated using 12 domain-expert defined labeling functions (LFs).
  • CORA-WS: A weakly supervised extension of the benchmark Cora dataset for research paper topic classification. Nodes are papers, edges are citations, and TF-IDF vectorized abstracts are node features. Weak labels were generated using two rule-based LFs based on keywords in abstracts and titles.
  • Benchmark Datasets: The paper also mentions using five benchmark datasets for synthetic experiments: Citeseer, Disease, ogbn-arxiv, Amazon Product, and Coauthor Physics. These datasets were used to simulate varying levels of label noise.

Key Insights and Ideas

  • Weak labels are valuable, even when noisy: The core insight is that even imperfect, noisy labels can be effectively leveraged to improve graph representation learning, outperforming methods that ignore them or rely solely on clean labels.
  • Weak label agreement is a proxy for reliability: Nodes with higher agreement among their weak labels tend to have more reliable aggregated labels, a finding supported by empirical evidence. This allows for intelligent weighting of supervision signals.
  • Contrastive learning can mitigate label noise: The paper demonstrates that contrastive learning, particularly when guided by weak label similarity and graph structure, can help denoise representations and improve robustness in the presence of noisy labels.
  • Combining multiple signals is key: WSNET's success stems from its ability to integrate diverse signals: graph structure, node features, weak labels, and contrastive objectives. This multi-faceted approach leads to more robust and generalizable embeddings.

Key Results

  • WSNET consistently outperforms all baselines across four real-world datasets (LIAR-WS, ASW-SYNTH, ASW-REAL, CORA-WS) and five benchmark datasets.
  • On LIAR-WS, WSNET achieved an 88% performance, surpassing the highest weak label accuracy (87%).
  • On ASW-SYNTH and ASW-REAL, WSNET achieved competitive performance (78% and 74% respectively), on par with or closely approaching state-of-the-art results.
  • WSNET shows a significant advantage over baselines, particularly under low label accuracy conditions in synthetic experiments, highlighting its robustness to noisy labels.
  • Ablation studies show that removing the weak label cross-entropy loss causes the most significant performance drop, emphasizing the importance of even noisy supervision.

Limitations and Future Work

  • Community structure relevance: The paper acknowledges that the impact of WSNET's methodology on graphs with less relevant community structures (e.g., irregular biological graphs) warrants further investigation.
  • Clustering alignment: Improving the alignment between underlying class structure and weak label-informed node embedding clusters is suggested as a future direction, potentially through semi-supervised or constrained clustering techniques.
  • Broader implications: The authors suggest that WSNET's success in integrating weak supervision with contrastive learning opens new pathways for GNN research, encouraging the development of models capable of handling diverse and imperfectly labeled data across various domains.
  • Ethical considerations: The paper highlights the importance of ethical application, emphasizing the need for inclusive and cross-verified labeling functions, and being mindful of potential biases introduced by domain experts or keyword-based labeling. Future work on weak label aggregation should also address bias towards specific labeling functions.

Summary updated

Abstract

Node classification in real world graphs often suffers from label scarcity and noise, especially in high stakes domains like human trafficking detection and misinformation monitoring. While direct supervision is limited, such graphs frequently contain weak signals, noisy or indirect cues, that can still inform learning. We propose WSNET, a novel weakly supervised graph contrastive learning framework that leverages these weak signals to guide robust representation learning. WSNET integrates graph structure, node features, and multiple noisy supervision sources through a contrastive objective tailored for weakly labeled data. Across three real world datasets and synthetic benchmarks with controlled noise, WSNET consistently outperforms state of the art contrastive and noisy label learning methods by up to 15% in F1 score. Our results highlight the effectiveness of contrastive learning under weak supervision and the promise of exploiting imperfect labels in graph based settings.
BibTeX
@misc{nair2025weak,
  author = {Pratheeksha Nair and Reihaneh Rabbany},
  title = {{Weak Supervision for Real World Graphs}},
  year = {2025},
  eprint = {2506.02451},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2506.02451},
}
Figure