Towards a Rigorous Evaluation of Explainability for Multivariate Time Series

Explainability for time series forecasting remains underdeveloped: While explainability research has advanced for tabular, text, and image data, the paper identifies that "explainability in time series is still relatively unexplored," highlighting a significant gap in making AI systems transparent across different data types.

LIME and SHAP can help non-experts understand AI predictions: The study found that two model-agnostic explainability techniques (LIME and Shapley additive explanations) "greatly helped lay humans in understanding the predictions made by the machine learning model," suggesting these methods can make complex AI decisions more accessible to business stakeholders without technical expertise.

Human evaluation is essential for assessing explainability: Rather than relying solely on technical metrics, the researchers used "human evaluation of explainability" to validate their approach, emphasizing that true explainability must be measured by whether actual people can understand and trust the model's reasoning.

Explainability enables accountability in business applications: The paper demonstrates a practical use case where explainability helped a consultancy company understand "the effect of their sales related activities on the sales deals closed," showing how XAI can support responsible decision-making by connecting business actions to outcomes.

Model-agnostic approaches offer flexibility: By using techniques that work across different machine learning models rather than being tied to specific algorithms, the study suggests that explainability solutions can be more broadly applied and adapted to various organizational needs.