Dynamic Graph Explanations

Identifies which edges appearing at specific timestamps most influenced the model's output. Unlike static graph explainers, this method accounts for the order and timing of interactions.

• Assigns a score to each (u, v, t) tuple
• Reveals if a prediction was driven by a recent burst of activity or a long-term stable connection
• Essential for fraud detection where the sequence of transactions matters more than the static graph structure

Measures how a node's evolving role over time contributes to predictions. A node's importance is not static—it fluctuates based on its temporal neighborhood and interaction patterns.

• Tracks when a node became influential, not just if it was
• Distinguishes between nodes that are consistently central vs. those with spiking importance during critical events
• Used in social network analysis to identify information spread catalysts

Extracts the minimal time-respecting subgraph that serves as the causal rationale for a prediction. This subgraph preserves both the structural connectivity and the temporal ordering of interactions.

• Must respect temporal causality: an edge at time t cannot be influenced by an edge at time t+1
• Evaluated using temporal fidelity metrics that measure prediction change when the subgraph is removed
• Critical for explaining predictions in dynamic molecular simulations

Attributes importance to discrete temporal events in continuous-time dynamic graphs. Uses techniques adapted from integrated gradients and Shapley values computed over the temporal domain.

• Models the graph as a temporal point process where each interaction is an event
• Explains why a specific ordering of events led to the outcome
• Applied in recommendation systems to explain why a user received a suggestion after a particular sequence of interactions

Traces how attention weights propagate across both graph layers and time steps in temporal graph attention networks. Visualizes the dynamic information pathways that led to a decision.

• Reveals if the model attends more to recent interactions or long-range dependencies
• Identifies temporal attention bottlenecks where information from many timestamps converges
• Provides interpretability for traffic forecasting models that predict congestion based on historical sensor data

Generates the minimal set of temporal edge additions or removals that would change the GNN's prediction. Focuses on actionable recourse in time-evolving systems.

• Answers: 'Which interaction, if it had not occurred or had occurred at a different time, would alter the outcome?'
• Must preserve temporal consistency—counterfactuals cannot violate the arrow of time
• Used in financial compliance to identify the specific transaction that triggered a risk flag

Dynamic graph explainers identify the critical temporal sequence of transactions that triggered a fraud alert. By attributing importance to specific edges at specific timestamps, analysts can trace the exact moment a legitimate account was compromised.

• Temporal Edge Attribution: Pinpoints the single high-value transfer that changed the risk score
• Sequence Analysis: Reveals if the fraud was a slow build-up or a sudden smash-and-grab
• Regulatory Compliance: Provides auditable evidence for Suspicious Activity Reports (SARs)

In molecular dynamics, dynamic GNN explanations reveal which molecular bonds form or break over time to cause a binding event. Researchers use temporal subgraph attribution to understand the reaction pathway rather than just the final docked state.

• Transition State Identification: Highlights the critical intermediate conformations
• Bond Importance Decay: Tracks how the relevance of a hydrogen bond fades over simulation time
• Off-Target Risk: Explains why a drug binds to an unintended protein by comparing temporal motifs

Dynamic explainers map the cascading propagation path of a viral post. By attributing importance to specific retweets and timestamps, analysts distinguish organic virality from coordinated inauthentic behavior.

• Patient Zero Detection: Identifies the originating node and the exact time of the cascade trigger
• Temporal Bottleneck Analysis: Finds the single retweet by an influencer that amplified the reach by 10x
• Bot Network Decomposition: Separates the explanation subgraph into organic and inorganic temporal components

Predicting vehicle trajectories requires understanding temporal interaction graphs between agents. Dynamic GNN explanations show which historical interactions (e.g., a cut-in 3 seconds ago) most influenced the ego-vehicle's braking decision.

• Causal Interaction Attribution: Links a specific past maneuver to a current prediction
• Temporal Receptive Field: Visualizes the look-back window that matters for the decision
• Safety Debugging: Explains why the planner predicted a collision by highlighting the offending agent's trajectory edge

Dynamic graph explainers attribute anomalous behavior in network traffic logs to the specific temporal connection that introduced the threat. This moves security operations from alert triage to root cause analysis.

• Lateral Movement Tracing: Explains the sequence of compromised hosts in a kill chain
• Temporal Anomaly Scoring: Differentiates a sudden spike in traffic from a slow data exfiltration
• Automated Incident Reports: Generates a human-readable timeline of the critical edges in the attack graph

When a shipment is delayed, dynamic GNN explanations trace the ripple effect through the temporal logistics network. The explainer attributes the delay to a specific port closure event and shows how its importance propagates forward in time.

• Bottleneck Propagation: Visualizes how a single node failure cascades through the graph over days
• Counterfactual Simulation: Identifies the minimal temporal edge addition (e.g., an alternative route) that would have prevented the disruption
• Inventory Reallocation: Recommends nodes to reinforce based on their temporal importance scores

A fidelity assessment method that measures the change in a dynamic GNN's prediction after masking or rewiring specific temporal edges. By systematically removing edges from critical timestamps and observing the drop in performance, engineers can identify which transient interactions were most influential. This technique is essential for validating that an explainer has correctly identified the true temporal motifs driving the prediction, rather than spurious correlations.

A principle for learning explainable dynamic GNNs by compressing the temporal graph into a minimal sequence of subgraphs that retains maximal mutual information about the label. The bottleneck discards irrelevant structural noise and transient fluctuations, isolating the causal temporal core of the prediction. This approach is particularly effective for financial transaction networks where only a few time-stamped interactions constitute fraudulent behavior.

A technique for interpreting dynamic Graph Attention Networks by analyzing the propagation and aggregation of attention weights across both layers and time steps. By tracing how a node's representation at time t attends to its historical neighbors, engineers can visualize the temporal receptive field that most influenced a prediction. This reveals whether the model relies on recent events or long-term structural memory.

The minimal temporal edge additions or removals that would alter a dynamic GNN's prediction to a desired outcome. For example, identifying that removing a single transaction at timestamp t-3 would have prevented a fraud alert. This provides actionable recourse in time-sensitive systems, allowing operators to understand precisely which historical interaction was the tipping point for the model's decision.

A game-theoretic approach that assigns a fair importance score to each temporal edge by computing its marginal contribution to the prediction across all possible coalitions of time-stamped interactions. Unlike static Shapley values, this method accounts for the ordering and recency of edges, ensuring that a critical edge appearing early in the sequence is properly credited for its foundational role in the evolving graph structure.

A self-explainable dynamic GNN framework where a generator module extracts a concise, causal temporal subgraph (the rationale) and a predictor module makes a decision based solely on that rationale. The generator is trained to select a minimal set of time-stamped edges that are both sufficient for prediction and coherent as a temporal narrative, enabling inherently interpretable dynamic graph models without post-hoc analysis.