A polarity-aware graph contrastive learning framework for signed link prediction.
Signed networks model relationships that can be either positive or negative — for example, trust/distrust, support/opposition, friendship/hostility, or agreement/disagreement. Unlike ordinary graphs, signed graphs carry polarity, which makes representation learning more challenging.
This project implements a Structural Balance aware Contrastive Learning Framework for Signed Networks. The goal is to learn robust node embeddings for signed link prediction, where the model predicts whether a relationship between two nodes is positive or negative.
Instead of applying generic random augmentations to a graph, this framework uses signed-network-specific structural cues inspired by structural balance theory. The model learns from complementary graph views while preserving the meaning of positive and negative links.
Most graph contrastive learning methods are designed for unsigned graphs. In unsigned graphs, random edge dropping or feature masking can work well as augmentations. However, in signed networks, these operations can easily destroy important polarity patterns.
For example:
- A friend of a friend is often a friend.
- An enemy of an enemy may become a friend.
- A friend of an enemy may indicate a negative relationship.
These patterns are not just graph structure; they carry social meaning. If augmentations ignore edge signs, the model may learn from unrealistic views of the network.
This project focuses on a simple idea:
Signed graph augmentations should preserve signed-network semantics instead of treating all edges as ordinary connections.
Positive and negative edges are processed through separate GNN channels. This prevents positive and negative signals from being mixed too early and helps preserve the meaning of each relation type.
The framework creates meaningful graph views using signed structural priors rather than purely random perturbations.
Main augmentation views include:
- Signed random-walk-based view to capture second-order balance patterns.
- Centrality-based view to highlight influential nodes and polarized hubs.
The model is trained using both supervised and self-supervised objectives:
- Supervised link sign prediction loss for the downstream prediction task.
- Inter-view contrastive loss to align representations across augmented graph views.
- Intra-view contrastive loss to keep final embeddings closer to positive-channel evidence and farther from negative-channel evidence.
Given a signed graph:
G = (V, E, S)
where:
Vis the set of nodes,Eis the set of observed edges,Srepresents the sign of an edge:+1for positive and-1for negative,
predict the sign of an unseen or held-out edge between two nodes.
Signed Graph
│
├── Positive Edge View ──► Positive GNN Encoder ──┐
│ │
├── Negative Edge View ──► Negative GNN Encoder ──┼──► Fusion Head ──► Link Sign Predictor
│ │
└── Augmented Views ─────► Contrastive Objectives ─┘
The model learns node representations by combining:
- polarity-separated message passing,
- signed graph augmentations,
- contrastive alignment across graph views,
- and supervised link sign prediction.
The main experimental dataset is the Bitcoin-OTC signed trust network.
| Dataset | Nodes | Positive Links | Negative Links | Positive Ratio | Sparsity |
|---|---|---|---|---|---|
| Bitcoin-OTC | 5,881 | 32,029 | 3,563 | 0.8999 | 0.1029% |
The dataset is challenging because it is:
- highly sparse,
- heavily imbalanced toward positive links,
- noisy due to real-world trust/distrust behavior,
- and long-tailed in node degree distribution.
The framework was evaluated using an 85/15 train-test split on Bitcoin network experiments.
| Metric | Best / Reported Value |
|---|---|
| Accuracy | ~0.9496 |
| Precision | up to 0.961 |
| Recall | up to 0.9888 |
| Binary F1 | 0.9733 |
| Micro F1 | 0.9493 |
Results may vary slightly depending on the split, random seed, augmentation configuration, and threshold selection.
The model performs strongly on operating-point metrics such as F1, precision, and recall, which are important for practical use cases like trust assessment, fraud detection, and moderation systems.
The project includes experiments to understand the contribution of different components:
- baseline signed GNN setup,
- random-walk augmentation,
- centrality-based augmentation,
- dual-view contrastive learning,
- and sparsity-based robustness analysis.
A key focus is evaluating whether the model remains stable when the graph becomes more sparse or noisy.
- Python
- PyTorch
- Deep Graph Library (DGL)
- NumPy
- Pandas
- Scikit-learn
- Matplotlib
- NetworkX
Signed link prediction has applications in:
- trust and reputation systems,
- fraud and risk analysis,
- social network mining,
- recommendation systems,
- community analysis,
- and moderation pipelines.
In these settings, negative links are often rare but highly informative. A model that ignores polarity or class imbalance may perform well superficially while missing the most important minority patterns.
Planned extensions include:
- evaluating on more signed network datasets such as Slashdot, Epinions, Reddit, and Wiki-RFA,
- extending augmentations to multi-hop signed random walks,
- adding probabilistic thresholds during augmentation,
- improving scalability for larger graphs,
- supporting temporal signed networks,
- and studying weighted or directed signed relationships.
This work was developed as an academic research / thesis project at ABV-IIITM Gwalior under the supervision of Dr. Roshni Chakraborty.
A manuscript based on this work is currently in preparation.
- Nilay Srivastava
- Shreya Vidyadhar
- Aradhya Dixit
- Shruti Chauhan
Supervisor: Dr. Roshni Chakraborty
Built to preserve the meaning of signed relationships, not just the structure of a graph.