Specified counterfactuals reveal how past changes alter graph predictions
Breaking Predictions Is Not Enough: Specified-Foil Counterfactuals for Temporal Graphs
Artificial Intelligence
Summary
Sometimes, people want to know how changing past events could lead to a different predicted outcome, not just any different outcome. The authors created a way to find specific past changes that would make a prediction switch from one option to a particular alternative. They do this by comparing how the original and alternative outcomes play out over time and identifying minimal changes to past connections or labels. Their method works well on two types of time-changing graphs and is more efficient than previous approaches.
What this means in practice
- •For data engineers: Identify precise past data modifications that lead prediction models to select a specific alternative outcome in time-based graphs.
- •For cybersecurity analysts: Discover exact prior actions or changes that would cause a threat detection system to flag a particular suspicious behavior over another.
Authors
Minwoo Yu, Young-guk Ha
Abstract
Temporal graph counterfactual explanations typically change past events to change or invalidate an original prediction, while leaving its replacement unspecified. Yet a user facing a predicted outcome often asks which past conditions would make a particular alternative occur instead. We formulate this destination-specific question as the Specified-Foil Counterfactual: given an original prediction A and a foil B fixed before search, find a low-cost past-event intervention under which the same predictor selects B as top-ranked. Our trace-guided intervention search contrasts the completed execution of A with a reconstructed incomplete execution of B, maps their difference to DELETE, INSERT, REWIRE, RELABEL, and SHIFT operations, and verifies B through exact replay. We instantiate this principle with LiFTER on continuous-time dynamic graphs and TLogic on temporal knowledge graphs. On CTDGs, the method retains 85.7-93.6% of black-box greedy successes while reducing predictor evaluations by 75.0-80.0%; on TKGs, it reaches the specified foil in 74.8% of 600 comparisons. Executable traces thereby become computational structures for constructing conditions of unselected alternatives, rather than records used only to explain predictions already made.