This is a hub for our team dedicated to explainable AI (XAI) for time series, providing valuable resources and insights to enhance understanding and application in this critical field.
We are hiring!
- Research Interns: We are currently looking for passionate graduate students to join our team as interns! We also If you are interested in contributing to the field of explainable AI for time series and making a meaningful impact, check out our job openings
- Alibaba Innovative Research: We are actively seeking professors for exciting collaboration opportunities! This one-year project focuses on explainable AI (XAI) in time-series scenarios. Professors can send 1-2 interns to join our team, enhancing their academic experience while contributing to meaningful research. You may refer to more details at the official website.
If you are interested in exploring these opportunity, please reach out to us at [[email protected]] or [[email protected]]. Let's work together to push the boundaries of XAI!
Our latest work Explainable Artificial Intelligence for Time Series： A Comparative Survey has been released! A comparative review analyzing state-of-the-art XAI methods for time series, with practical guidelines for method selection and implementation. Download PDF to master time series explainability.
- YANG, Linxiao: [[email protected]]Always the core idea contributor in XAI for time series, interactive XAI and rule learning.
- GENG, Linyuan: [[email protected]] Statistics expert, focusing on self-learning to enhance temporal model performance.
- GU, Xinyue: [[email protected]] Working on statistics, operations and causal inference; a theoretical proof provider.
Our survey Explainable Artificial Intelligence for Time Series: A Comparative Survey presents:
Time series data presents unique challenges for explainability:
- Strong temporal dependencies between observations
- Long-term effects and phase shifts from events
- Need to explain both individual points and subsequences
- Complex interactions between temporal and feature dimensions
Our survey introduces how four major categories of XAI methods address these challenges:
- Perturbation-based approaches
- Distribution-aware methods
- Decomposition-based techniques
- Attention mechanisms
Key contributions include:
- Comprehensive Analysis: Systematic review of 50+ XAI methods for time series
- Comparative Study: Carefully designed experiments comparing different methods across various tasks
- Practical Guidelines:
- Guidelines for choosing appropriate XAI methods
- Task-specific evaluation metrics
- Explain Temporal Black-Box Models via Functional Decomposition:
How to explain temporal models is a significant challenge due to the inherent characteristics of time series data, notably the strong temporal dependencies and interactions between observations.
The unique characters of time series data： 1，the high degree of temporal dependency (including the commonly used autocorrelation) inherent in time series data complicates the task of identifying the influence of specific inputs on the model outputs. 2，events occurring at a given time point can have enduring effects on subsequent observations, encapsulating phase shifts that ripple through future data points. 3，we need to consider not only a single time point (e.g., points corresponding to events), but also time series subsequence in explanation.
cite this: @inproceedings{yang2024explain, title={Explain Temporal Black-Box Models via Functional Decomposition}, author={Yang, Linxiao and Tong, Yunze and Gu, Xinyue and Sun, Liang}, booktitle={International Conference on Machine Learning}, year={2024}, organization={PMLR} }
- Interactive Generalized Additive Model and Its Applications in Electric Load Forecasting:
Inaccurate load forecasting may lead to the threat of outages or a waste of energy. How to give accurate electric load forecasting when there is
- limited data or even no data, such as load forecasting in holiday, or under extreme weather conditions.
- a need for model interpretability as high-stakes decision-making usually follows after load forecasting.
We propose an interactive GAM which is not only interpretable but also can incorporate specific domain knowledge in electric power industry for improved performance.
cite this: @inproceedings{yang2023interactive, title={Interactive Generalized Additive Model and Its Applications in Electric Load Forecasting}, author={Yang, Linxiao and Ren, Rui and Gu, Xinyue and Sun, Liang}, booktitle={Proceedings of the 29th ACM SIGKDD Conference on Knowledge Discovery and Data Mining}, pages={5393--5403}, year={2023} }
- Learning Interpretable Decision Rule Sets: A Submodular Optimization Approach:
Rule sets are highly interpretable logical models in which the predicates for decision are expressed in disjunctive normal form (DNF, OR-of-ANDs). We consider a submodular optimization based approach to form an accurate and interpretable rule set. the proposed approach is simple, scalable, and can deal with the exponential-sized ground set of rules.
cite this: @inproceedings{yang2021learning, title={Learning interpretable decision rule sets: A submodular optimization approach}, author={Yang, Fan and He, Kai and Yang, Linxiao and Du, Hongxia and Yang, Jingbang and Yang, Bo and Sun, Liang}, booktitle={Advances in Neural Information Processing Systems}, volume={34}, pages={27890--27902}, year={2021} }
- Mask or Perturbation:
- Explaining time series predictions with dynamic masks Jonathan Crabb ́e and Mihaela van der Schaar. ICML'21
- Learning perturbations to explain time series predictions Joseph Enguehard. ICML'23
- Explaining time series via contrastive and locally sparse perturbations] Zichuan Liu, Yingying Zhang, Tianchuan Wang, Zefan Wang, Dongsheng Luo, Menguan Du, Min Wu, Yi Wang, Chunlin Chen, Lunting Fan, and Qingsong Wen. ICLR'24
- Distribution-aware:
- What went wrong and when? instance-wise feature importance for time-series black-box models Sana Tonekaboni, Shalmali Joshi, Kieran R Campbell, David Duvenaud, and Anna Goldenberg. In Proceedings of the International Conference on Neural Information Processing Systems, volume 33, pages 799–809.2020
- Temporal dependencies in feature importance for time series prediction Kin Kwan Leung, Clayton Rooke, Jonathan Smith, Saba Zuberi, and Maksims Volkovs. In The International Conference on Learning Representations, 2023.
- Feature importance identification for time series classifiers Han Meng, Christian Wagner, and Isaac Triguero. In IEEE International Conference on Systems, Man, and Cybernetics, pages 3293–3298, 2022
- SHAP/LIME-based:
- Feature importance for time series data: Improving kernelshap Mattia Villani, Joshua Lockhart, and Daniele Magazzeni. International Conference on AI in Finance Workshop on Explainable Artificial Intelligence in Finance, 2022.
- Interpreting a recurrent neural network’s predictions of icu mortality risk Long V. Ho, Melissa Aczon, David Ledbetter, and Randall Wetzel. Journal of Biomedical Informatics, 114(C), 2021
- Limesegment: Meaningful, realistic time series explanations Torty Sivill and Peter Flach. In International Conference on Artificial Intelligence and Statistics, pages 3418-3433, 2022
- Structural:
- Temporal fusion transformers for interpretable multi-horizon time series forecasting - Bryan Lim, Sercan ̈O Arık, Nicolas Loeff, and Tomas Pfister. International Journal of Forecasting, 37(4):1748–1764, 2021.
- Ssdnet: State space decomposition neural network for time series forecasting. Yang Lin, Irena Koprinska, and Mashud Rana. In IEEE International Conference on Data Mining, pages 370–378. IEEE, 2021.
- Probabilistic transformer for time series analysisBinh Tang and David S Matteson. In Advances in Neural Information Processing Systems, volume 34, pages 23592–23608. Curran Associates, Inc., 2021.