LiFTER: Neuro-Symbolic AI Explains Graph Forecasts

This research introduces LiFTER (Link-Fact Temporal Rule Inducer), a novel neuro-symbolic predictor designed for continuous-time dynamic graphs. While existing models effectively forecast future links by compressing past interactions into neural states, they often lack transparency, obscuring the specific entities and temporal patterns contributing to a prediction. LiFTER addresses this interpretability gap by integrating explainability directly into its predictive architecture. LiFTER operates by preserving observed interactions as "grounded temporal facts" and then applying executable "temporal rules" to pre-query these facts. Each prediction score is derived from a signed sum of rule executions, where the historical facts, entity bindings, and temporal order are explicitly satisfied and traceable. This design allows for the inspection, independent recomputation, and intervention upon the evidence and rules responsible for any given prediction. Evaluations across four continuous-time dynamic graph benchmarks show that LiFTER achieves competitive performance in historical-negative forecasting. Crucially, it excels in explanation accuracy and deletion fidelity, outperforming other methods. The architecture also functions as a "microscope," enabling the separation and tracing of contributions from recurrence, history position, and transition to individual facts. With minimal error in reconstructing logits, LiFTER transforms future-link forecasting into a verifiable, grounded computation.