A cost-constrained dismantling framework for hypergraph with dynamic cascading failure using deep reinforcement learning

Abstract

Network dismantling is a fundamental challenge in the study of complex networks, with significant real-world applications such as controlling the spread of diseases and disrupting terrorist networks. However, most existing studies focus on pairwise interaction networks and static dismantling strategies, often overlooking the dynamic nature of cascading failures and the associated cost constraints that are prevalent in real-world scenarios. These limitations restrict their ability to effectively capture the complexity of failure dynamics in systems with higher-order interactions. To overcome these challenges, we utilize hypergraphs to model higher-order relationships within complex systems and introduce a dynamic dismantling approach that explicitly accounts for cascading failures under cost constraints. Building on this foundation, we propose a novel hypergraph dismantling framework, C2HD-RL, which leverages deep reinforcement learning. The framework enables an agent to iteratively explore different node selection strategies in synthetic hypergraphs and adjust its behavior based on the rewards received, ultimately learning an optimal hypergraph dismantling strategy. Comprehensive evaluations on nine real-world hypergraph datasets, compared against seven baseline methods, demonstrate the effectiveness of our approach.