An Empirical Evaluation of Factors Affecting SHAP Explanation of Time Series Classification

Segmentation strategy matters for explaining AI predictions on time series data: The paper found that how time series data is divided into segments significantly affects the quality of SHAP explanations, which are used to understand why AI models make specific predictions about sequential data like stock prices or sensor readings.

Simpler approaches often work better than complex ones: Equal-length segmentation (dividing time series into uniform chunks) consistently outperformed more sophisticated, custom-designed segmentation algorithms, suggesting that practitioners don't need overly complicated methods to get good explanations.

The number of segments is more important than the method itself: The research found that "the number of segments has a greater impact on explanation quality than the specific segmentation method," meaning organizations should focus on finding the right number of divisions rather than debating which segmentation technique to use.

A practical solution to a real computational problem: The paper addresses a genuine barrier to using SHAP for time series—its computational complexity grows exponentially with more data points. By using segmentation strategies, the authors show SHAP becomes more practical for real-world applications involving long time series.

A new technique improves explanation reliability: The authors introduced a "novel attribution normalisation technique that weights segments by their length" that consistently improved explanation quality, offering practitioners a concrete improvement they can implement when explaining time series AI models.