LiFTER: Neuro-Symbolic Model for Continuous-Time Dynamic Graph Forecasting
A recent study presents LiFTER (Link-Fact Temporal Rule Inducer), a neuro-symbolic model designed for continuous-time dynamic graphs (CTDGs). Published on arXiv under ID 2608.06765v1, this research tackles the lack of transparency in current neural models that condense previous interactions into states, which masks the contributions of entity sharing and temporal patterns. LiFTER maintains observed interactions as grounded temporal facts and utilizes executable temporal rules on pre-query facts, generating scores through signed sums of rule executions that include clearly satisfied historical facts, entity bindings, and temporal sequences. This framework enables predictions to be analyzed, recalculated independently, and adjusted as needed. LiFTER demonstrates strong historical-negative forecasting and the best macro-averaged performance across four CTDG benchmarks, making it significant for the digital art and AI technology fields by providing interpretable predictions for dynamic networks.
Key facts
- LiFTER is a neuro-symbolic predictor for continuous-time dynamic graph forecasting.
- It preserves observed interactions as grounded temporal facts and applies executable temporal rules.
- Each score is a signed sum of rule executions with explicitly satisfied historical facts, entity bindings, and temporal order.
- The model allows predictions to be inspected, independently recomputed, and intervened upon.
- LiFTER achieves competitive historical-negative forecasting across four CTDG benchmarks.
- It achieves the highest macro performance on the benchmarks.
- The paper is available on arXiv with ID 2608.06765v1.
- The approach treats interpretability as a property of the predictive architecture.
Entities
Institutions
- arXiv