Source-linked AI summary
Signed Graph Attention Networks
Junjie Huang, Huawei Shen, Liang Hou, Xueqi Cheng
TL;DR
GATs are effective for graph tasks but are designed for networks with only positive links, leaving signed networks with positive and negative links insufficiently addressed. SiGAT generalizes GAT to signed networks by combining attention with motifs encoding balance and status theories, and experiments show improved performance across three real signed networks.
Problem
GAT is designed for networks with only positive links and does not handle signed networks containing both positive and negative links.
Method
SiGAT generalizes GAT to signed networks by extracting signed motifs and aggregating motif-specific information with graph attention to generate node representations.
Results
Experiments on three real-world signed networks show that SiGAT outperforms different state-of-the-art signed network embedding methods.
Takeaways & Limitations
SiGAT provides a motif- and attention-based framework for modeling signed networks and shows robust results across hyper-parameter settings.
Abstract
from arXiv · showhide
Graph or network data is ubiquitous in the real world, including social networks, information networks, traffic networks, biological networks and various technical networks. The non-Euclidean nature of graph data poses the challenge for modeling and analyzing graph data. Recently, Graph Neural Network (GNNs) are proposed as a general and powerful framework to handle tasks on graph data, e.g., node embedding, link prediction and node classification. As a representative implementation of GNNs, Graph Attention Networks (GATs) are successfully applied in a variety of tasks on real datasets. However, GAT is designed to networks with only positive links and fails to handle signed networks which contain both positive and negative links. In this paper, we propose Signed Graph Attention Networks (SiGATs), generalizing GAT to signed networks. SiGAT incorporates graph motifs into GAT to capture two well-known theories in signed network research, i.e., balance theory and status theory. In SiGAT, motifs offer us the flexible structural pattern to aggregate and propagate messages on the signed network to generate node embeddings. We evaluate the proposed SiGAT method by applying it to the signed link prediction task. Experimental results on three real datasets demonstrate that SiGAT outperforms feature-based method, network embedding method and state-of-the-art GNN-based methods like signed graph convolutional network (SGCN).
1 Introduction
The section motivates extending GNNs and GATs from mainly unsigned networks to directed signed networks. It introduces SiGAT, which uses signed motifs, attention, and sociological theories for signed-network modeling and evaluates the framework on real data.
- Previous GNN methods mainly focus on undirected, unsigned networks, leaving signed, directed, weighted, and other complex networks as an important research direction.
- SiGAT introduces GAT to directed signed networks, using attention to characterize different effects of neighboring nodes on a target node.
- Triad motifs describe balance theory and status theory, providing structural patterns for analyzing signed link prediction.
- The proposed motif-based GNN model for signed networks is named Signed Graph Attention Network, or SiGAT.
- Experiments on real signed social-network datasets evaluate SiGAT's effectiveness for learning node embeddings.
2 Related work
Related work covers signed-network analysis methods and the development of GNN-based graph representation methods. Existing approaches include feature engineering, network embedding, convolution, auto-encoding, and attention mechanisms.
- Signed social networks contain positive and negative social ties and support tasks including community detection, node classification, link prediction, and spectral graph analysis.
- Early signed-network link prediction methods primarily used feature engineering, while later embedding methods learned low-dimensional network representations.
- GNN research has applied convolution, attention, auto-encoding, and related mechanisms to graph analysis problems.
- Representative GNN methods include GCN for node-level semi-supervised classification, GraphSAGE for large inductive graphs, GAE for unsupervised vertex representation, and GAT for attention-based graph processing.
3 Methods
SiGAT models directed signed networks by combining motif-based structural patterns with graph attention. It uses balance and status theories to define motifs, aggregates motif-specific neighborhood messages, and produces node embeddings for signed-network tasks.
- Signed Network Theory: Balance and status theories provide the sociological basis for modeling signed directed networks with distinct structural patterns.Balance theory concerns signed triads, while status theory interprets directed signs as relative status relationships.
- Signed Motifs: SiGAT defines 2 positive/negative-neighbor motifs, 4 directed signed-neighbor motifs, and 32 triangle motifs.These motifs separately encode link signs, directions, and triadic structure.
- Signed Graph Attention Networks: SiGAT concatenates the original node features with all motif-specific messages and feeds them to a two-layer fully connected network to obtain node representation Z_u.The model also uses a mini-batch algorithm for scalability.
- Signed Graph Attention Networks: For each motif, SiGAT extracts the corresponding neighborhood and applies a GAT aggregator to compute motif-specific messages.Attention weights characterize how different neighborhood nodes influence the target node.
- Signed Graph Attention Networks: The unsupervised objective encourages embeddings of friends to be similar and embeddings of enemies to be dissimilar.The loss distinguishes positive and negative neighborhoods and includes a balancing parameter for unequal neighborhood sizes.
4 Experiments
Experiments evaluate signed link prediction on three real-world datasets against feature-based, unsigned and signed embedding, and GNN baselines. SiGAT achieves almost the best performance across the datasets, while parameter analysis indicates relatively robust behavior.
- Link Sign Prediction and Dataset Description: Experiments use three real-world signed social-network datasets and evaluate signed link prediction.The task predicts unobserved signs of existing test edges from training data.
- Link Sign Prediction and Dataset Description: The datasets contain imbalanced positive and negative links, requiring evaluation methods that account for signed-network structure.Bitcoin-Alpha scores are converted to positive when greater than 0 and negative otherwise.
- Baselines: The comparison includes random, unsigned embedding, signed embedding, SGCN, FeExtra, and a simplified SiGAT variant using positive and negative neighbors.All methods use embedding dimension d = 20 for a fair comparison, with learned node representations classified by logistic regression where applicable.
- Results: Signed embedding methods outperform unsigned embedding methods, while FeExtra performs well but relies on manually engineered edge features and has weak generalization to other tasks.SIGNet shows more stable and better results among the signed embedding methods in these experiments.
- Results: SGCN cannot effectively model signed directed networks because it ignores direction and averages neighbors’ hidden states.The authors contrast this with attention, which considers different nodes’ influence on the target node.
- Results: SiGAT achieves almost the best performance on all three datasets, and the simplified SiGAT variant is close to or slightly better than SGCN.The results are attributed to modeling sociological theories with motifs and using attention to distinguish neighbor influence.
- Parameter Analysis: Loss decreases and AUC increases at earlier epochs before converging, with the best epoch varying across datasets.The analysis evaluates epoch effects with d = 20 and dimension effects using dataset-specific epoch settings.
- Parameter Analysis: Performance becomes nearly stable as dimension increases, and the model shows relatively robust results across the analyzed hyper-parameters.The authors relate dataset differences in best epochs to network scale and structure.
5 Conclusion
SiGAT introduces attention and motif-based modeling for signed networks, and experiments on three real-world signed networks show that it outperforms state-of-the-art signed network embedding methods. The paper also reports robustness analysis and identifies broader signed-network tasks for future study.
- SiGAT redesigns GNN architectures to model signed networks using attention mechanisms and different motifs.
- Experiments on three real-world signed networks show that SiGAT outperforms different state-of-the-art signed network embedding methods.
- Hyper-parameter analysis indicates that the proposed methods are robust.
- Future work includes evaluating SiGAT on node classification and clustering and developing more effective signed-network models.