L2GTX: From Local to Global Time Series Explanations

L2GTX addresses a critical gap in AI explainability: While deep learning models perform well at time series classification, understanding why they make decisions remains poorly understood. This paper tackles the specific challenge that existing explanation methods designed for images and tables don't work well for time series data, which have unique temporal patterns.

The framework prioritizes interpretability through "temporal event primitives": Rather than producing complex visualizations, L2GTX identifies and explains decisions using understandable patterns like "increasing trends," "decreasing trends," and "local peaks/valleys"—concepts that domain experts and stakeholders can readily grasp and verify.

Global explanations are built efficiently from representative examples: Instead of explaining every single prediction, L2GTX strategically selects a small set of representative instances and aggregates their explanations into class-level insights. This approach makes explanations more scalable and actionable for real-world deployment.

Model-agnostic design increases practical applicability: Because L2GTX works with any deep learning model rather than requiring model-specific modifications, it can be applied across different time series applications (medical monitoring, financial forecasting, etc.) without reengineering explanation systems.

Explanations maintain "faithfulness" to actual model behavior: The framework measures whether global explanations accurately reflect what the model learned, helping ensure that explanations aren't misleading—a key concern for responsible AI deployment in high-stakes domains.