TSInsight: A Local-Global Attribution Framework for Interpretability in Time Series Data.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Data Science

Background:

• Deep learning models are increasingly used in safety-critical applications, necessitating interpretable methods.
• Interpretability research for time series data lags behind visual modalities due to poor intelligibility of existing techniques.

Purpose of the Study:

• To address the lack of effective interpretability methods for deep time series models.
• To introduce TSInsight, a novel feature attribution framework for time series data.

Main Methods:

• TSInsight attaches an auto-encoder to a classifier, using a sparsity-inducing norm on its output.
• The auto-encoder is fine-tuned using gradients from the classifier and a reconstruction penalty.
• The method learns to preserve predictive features and suppress irrelevant ones, acting as a feature attribution technique.

Main Results:

• TSInsight effectively preserves features crucial for classification while suppressing irrelevant ones.
• The method demonstrated capability in generating both instance-based and model-based explanations.
• Evaluations on eight time series datasets showed TSInsight's superiority over nine other attribution methods.

Conclusions:

• TSInsight is an effective tool for enhancing the interpretability of deep time series models.
• The method naturally achieves output space contraction, a desirable property for feature attribution.
• TSInsight offers a significant advancement in understanding complex time series data through deep learning.