TAESPM: A Learning-Based Spatiotemporal Prediction Framework for Dynamic Influence Maximization

Abstract

In dynamic social networks, identifying critical nodes capable of sustaining influence is paramount. However, existing methodologies frequently succumb to the “Myopic Trap”, leading to suboptimal strategies. We propose TAESPM, a spatiotemporal prediction framework that forecasts high-potential nodes by synergistically modeling evolutionary characteristics. Specifically, we devise a Time-aware Influence Capacity (TIFC) metric and a hybrid encoder (GNN-BiLSTM) with a Temporal Awareness Enhancement Module (TAEM). Experimental results show an accuracy of up to 98% in candidate node prediction, significantly outperforming SOTA benchmarks on the Cumulative Temporal Spread metric.

Key Contributions

• Adaptive Optimization (TIFC): A novel methodology that models historical decay effects and neighbor emergence to measure dynamic nodal influence precisely.
• Hybrid Spatiotemporal Architecture: Seamlessly integrates GNN and BiLSTM to facilitate multi-scale feature fusion across evolving network topologies.
• Temporal Awareness Enhancement Module (TAEM): Explicitly captures long-range dependencies and critical inflection points, overcoming the forgetting issues of traditional RNN/LSTM models.
• Performance & Efficiency: Achieves up to 98% prediction accuracy while drastically curtailing search space and computational overhead in complex dynamic scenarios.

Keywords

• Dynamic Social Networks
• Influence Maximization
• Spatiotemporal Deep Learning
• Predictive Modeling

The contents above are part of a manuscript currently under professional review. For more information, please visit the project documentation.

Recommended citation: Lvyizhuo, et al. (2025). "TAESPM: A Learning-Based Spatiotemporal Prediction Framework for Dynamic Influence Maximization." (Under Review in IEEE TNSE).
Download Paper