The problem
Time series classification often matters in safety-critical applications, yet it is usually tackled with black-box deep learning models whose decisions are hard for people to understand.
The approach
STELLE (Signal Temporal logic Embedding for Logically-grounded Learning and Explanation) is a neuro-symbolic framework that unifies classification and explanation. Trajectories are embedded directly into a space of temporal logic concepts through a new STL-inspired kernel, which measures how well a raw time series aligns with a set of predefined STL formulae.
The model optimises accuracy and interpretability together, producing:
- local explanations: human-readable STL conditions that justify each prediction;
- global explanations: formulae that characterise each class.
Results
Experiments on diverse real-world benchmarks show competitive accuracy together with logically faithful explanations.