Technical Summary: Dynamic Link Prediction with Temporally Enhanced Signed Graph Neural Networks

1. Problem Statement

Temporal signed networks (TSNs) model the evolution of cooperative and adversarial relationships in domains such as social media, trust systems, and financial transactions. While Graph Neural Networks (GNNs) have advanced static and unsigned link prediction, effective learning in temporal signed graphs remains challenging. Existing methods often treat networks as static snapshots or fail to account for the complex interplay between signed relations (positive/negative polarity), evolving structures, and balance-theoretic constraints.

Specifically, current approaches struggle with four key challenges:

1. Temporal Information Decay: The influence of historical signed interactions diminishes over time, yet static Signed GNNs (SGNNs) lack adaptive mechanisms to weight recent versus older evidence.
2. Evolving Balance Structures: Structural balance patterns (e.g., triadic stability) are not fixed; they require continual recalibration as networks grow.
3. Heterogeneous Temporal Dynamics: Different nodes exhibit distinct temporal behaviors (e.g., central nodes retaining long-term relevance vs. peripheral nodes with rapidly decaying influence), necessitating adaptive rather than uniform weighting.
4. Interpretability: Incorporating temporal mechanisms often reduces the transparency essential for trust-critical applications, and existing temporal attention models offer limited insight into how temporal factors drive signed predictions.

2. Methodology

The authors propose a modular temporal enhancement framework designed to integrate historical context into otherwise static SGNN architectures without altering their core structural encodings. The framework is architecture-agnostic but is instantiated here using the Self-Explainable Signed Graph Transformer (SE-SGformer).

Core Component: Historical Context Integration Module (HCIM)

The HCIM extracts temporally informed representations from historical graph embeddings through a three-stage process:

1. Adaptive Temporal Weighting:

   • Applies learnable, recency-aware weighting to historical node embeddings $\{Z(t-T), \dots, Z(t-1)\}$.
   • Uses a weighting coefficient $w_\tau$ defined by an exponential decay factor $\lambda$ and a recency emphasis parameter $\gamma$.
   • Formula: $w_\tau = \frac{\lambda^{T-1-\tau} \cdot e^{\tau/\gamma}}{\sum_{i=0}^{T-1} \lambda^{T-1-i} \cdot e^{i/\gamma} + \epsilon}$.
   • This allows the model to adaptively prioritize informative historical interactions rather than treating all past snapshots equally.

2. Sequential Pattern Modeling:

   • The weighted embeddings are stacked into a temporal sequence and processed by an LSTM network.
   • This stage captures sequential dependencies and long-range temporal trajectories in signed relationships.

3. Attention-Based Context Integration:

   • A multi-head self-attention mechanism is applied to the LSTM outputs to selectively emphasize historically informative temporal patterns.
   • The representation corresponding to the final timestep is projected to generate the final historical context representation ($H_{context}$).

Temporal Fusion Strategies

The extracted historical context is fused with current node embeddings ($Z_{current}$) using one of two strategies:

• Global Weighting: Applies a single learnable parameter $\alpha$ uniformly across all nodes. Suitable for synthetic networks with homogeneous temporal patterns.
• Node-Adaptive Weighting: Uses a Multi-Layer Perceptron (MLP) to generate node-specific fusion weights $\alpha_{node}$. This accommodates heterogeneous temporal dynamics in real-world data by allowing different nodes to dynamically adjust the importance of historical vs. current information.

The final enhanced embeddings ($Z_{enhanced}$) replace the original embeddings in the prediction layers, while the original loss formulation and signed prediction objectives remain unchanged.

3. Key Contributions

• HCIM Framework: A systematic three-stage temporal framework combining learnable recency-aware weighting, LSTM-based trajectory modeling, and multi-head temporal attention.
• Dual Temporal Weighting: Introduction of both global and node-adaptive weighting strategies to handle both homogeneous synthetic patterns and heterogeneous real-world dynamics.
• Interpretability Preservation: The framework maintains model explainability by using explicit learnable parameters for recency bias and fusion weights, allowing practitioners to validate temporal decision factors.
• Architecture-Agnostic Design: The approach extends existing SGNNs (demonstrated on SE-SGformer) without redesigning the backbone, preserving structural foundations like signed random walk encodings and structural attention biases.

4. Experimental Results

The framework was evaluated on real-world datasets (Bitcoin OTC, Bitcoin Alpha, Reddit Hyperlink) and synthetic networks (Barabási-Albert, Watts-Strogatz).

• Performance Gains: The temporally enhanced SE-SGformer consistently outperformed the static baseline across all datasets with statistically significant improvements ($p < 0.001$).
  • Bitcoin OTC: +3.79% relative AUC improvement; 30.32% error reduction.
  • Bitcoin Alpha: +5.76% relative AUC improvement; 40.06% error reduction.
  • Synthetic Networks: The model showed the most dramatic gains on synthetic data, with relative AUC improvements exceeding 13% and error reductions up to 41.77% on Barabási-Albert and Watts-Strogatz networks.
• Robustness: Despite the extreme structural heterogeneity and hub dominance in the Reddit Hyperlink subset, the framework achieved a 1.37% AUC improvement, demonstrating effectiveness even in highly irregular topologies.
• Efficiency: While the HCIM introduces additional computational overhead (scaling linearly with nodes and quadratically with temporal snapshots), empirical runtime increases were moderate (1.48% to 24.80%), remaining tractable for the evaluated datasets.

5. Significance and Claims

The paper positions this work as a critical bridge between temporal graph learning and signed graph neural networks. Its primary significance lies in:

• Bridging the Gap: It addresses the lack of explicit temporal modeling in signed graph learning, enabling accurate signed link prediction in dynamic and adversarial environments.
• Preserving Interpretability: Unlike approaches that redesign architectures entirely, this method injects temporal awareness into existing, interpretable SGNNs, maintaining the ability to explain predictions based on structural balance and status theories.
• Scalability and Flexibility: By being architecture-agnostic, it provides a scalable foundation for future temporal SGNN research.
• Application Impact: The framework advances capabilities in domains where understanding the evolution of cooperative and adversarial relationships is essential, such as financial fraud detection and online community moderation.

The authors conclude that temporal signed interactions contain predictive information beyond static graph structure alone, and that separating historical and structural representations before adaptive fusion is a beneficial strategy for temporal graph learning.