Explainability and importance estimate of time series classifier via embedded neural network

Area of Science:

• Data Science
• Machine Learning
• Time Series Analysis

Background:

• Time series analysis is complex due to inter- and intra-relationships within data.
• Interpreting feature importance in multivariate time series is challenging.
• Existing statistical methods and neural network (NN) approaches have limitations in feature importance estimation.

Purpose of the Study:

• To adapt the Pairwise Importance Estimate Extension (PIEE) method for time series analysis.
• To develop a method for estimating the importance and interpretability of features (time points and series) in multivariate time series.
• To compare the adapted PIEE method with existing NN and explainable AI (xAI) approaches.

Main Methods:

• Adaptation of the PIEE method using an aggregated Hadamard product for time series.
• Empirical study involving univariate and multivariate time series.
• Comparison with existing embedded NN approaches and an xAI method.
• Verification using ground truth, domain knowledge, and ablation studies (Leave-One-Out, Singleton).

Main Results:

• The adapted PIEE method successfully produced feature importance heatmaps and rankings.
• Results aligned with ground truth, domain knowledge, and ablation study findings.
• The method demonstrated effectiveness in interpreting individual time series and time points within multivariate data.

Conclusions:

• The adapted PIEE method offers a robust solution for feature importance estimation and interpretability in time series analysis.
• This approach enhances the understanding of complex multivariate time series data.
• The method provides valuable insights comparable to established techniques and domain expertise.