Fig. 2: The proposed ResInf framework.
From: Deep learning resilience inference for complex networked systems
ResInf leverages real-world observed or simulated network topologies and node activity trajectories to infer the resilience of networked systems without any prior assumption. It contains three modules: the dynamics encoder maps node activity trajectories to dense representations with stacked Transformer encoder layers, which models the complex correlations between node activities and produces representations for node activity dynamics; topology encoder leverages graph neural network to generate discriminating topological representations for each node’s multi-hop neighborhoods; k-space projector aggregates the representations for node activity dynamics and topologies via a virtual global node, and employs a multi-head self-attention network to fuse the representations learned from various trajectories dynamically. Subsequently, it uses a dimension reduction network to project the aggregated representation to a 1-dimensional k-space, facilitating accurate resilience inference with linear classifiers.
