Related work¶

tslens builds on a growing body of research in temporal attribution, perturbation, evaluation, and intrinsically interpretable forecasting. This page provides a map of the field and positions the methods available through the toolkit.

Temporal attribution methods¶

Work	Venue	Contribution
Benchmarking deep learning interpretability in time series predictions	NeurIPS 2020	Introduces Temporal Saliency Rescaling (TSR) and evaluation metrics for time and feature importance.
What went wrong and when? Instance-wise feature importance for time-series black-box models	NeurIPS 2020	Introduces FIT, which measures prediction changes using KL divergence.
Temporal Dependencies in Feature Importance for Time Series Prediction	ICLR 2023	Introduces WinIT and delayed feature importance over sliding temporal windows.
Learning Perturbations to Explain Time Series Predictions	ICML 2023	Learns attribution masks and the perturbations applied beneath them.
CGS-Mask: Making Time Series Predictions Intuitive for All	AAAI 2024	Produces binary, time-sensitive explanations with cellular genetic strip masks.
Encoding time-series explanations through self-supervised model behavior consistency	NeurIPS 2023	Trains interpretable surrogates while preserving pretrained-model behavior.

Evaluation and benchmarking¶

Work	Venue	Contribution
Evaluation of post-hoc interpretability methods in time-series classification	Nature Machine Intelligence 2023	Proposes quantitative evaluation criteria for classification explanations.
Evaluation of interpretability methods for multivariate time series forecasting	Applied Intelligence 2022	Benchmarks attribution methods for multivariate, multi-horizon forecasting.
Optimal local explainer aggregation for interpretable prediction	AAAI 2022	Combines local explanations into near-global views through integer optimization.

Interpretable temporal models¶

Work	Venue	Contribution
Temporal Fusion Transformers for interpretable multi-horizon time series forecasting	International Journal of Forecasting 2021	Combines high-performance forecasting with variable selection and temporal attention.
Self-Interpretable Time Series Prediction with Counterfactual Explanations	ICML 2023	Uses variational inference for temporal abduction, intervention, and prediction.
iTrendRNN: An Interpretable Trend-Aware RNN for Meteorological Spatiotemporal Prediction	AAAI 2024	Models evolving meteorological trends with interpretable attention units.

Surveys¶

For a practical comparison of methods available in this package, including model and baseline requirements, see Interpretation methods.