The problem
Time series classification is usually solved with black-box deep learning, which makes it hard to understand the rationale behind a prediction, a real issue in safety-critical settings.
The approach
A neuro-symbolic framework that unifies classification and explanation by embedding trajectories into a space of Signal Temporal Logic (STL) concepts. A novel STL-inspired kernel maps each raw time series to its alignment with a set of predefined STL formulae, so every prediction comes with the logical concepts that best characterise it.
This enables classification grounded in human-interpretable temporal patterns, with both local and global symbolic explanations.
Results
Early results show competitive performance together with high-quality logical justifications for the model's decisions. The work is developed further in Guided by Stars.