Explainable time-series forecasting with sampling-free SHAP for Transformers

Area of Science:

• Artificial Intelligence
• Machine Learning
• Time-Series Analysis

Background:

• Accurate time-series forecasting is crucial for planning.
• Model explainability is vital for user trust and transparency.
• Existing Shapley Additive Explanations (SHAP) methods can be computationally intensive.

Purpose of the Study:

• Introduce SHAPformer, a novel time-series forecasting model.
• Achieve accurate predictions with enhanced model explainability.
• Develop a faster SHAP-based explanation method for Transformers.

Main Methods:

• Utilized a Transformer architecture for time-series forecasting.
• Integrated Shapley Additive Explanations (SHAP) for model interpretability.
• Developed an attention manipulation technique to eliminate background data sampling for SHAP calculations.

Main Results:

• SHAPformer achieved accurate time-series forecasting performance.
• Generated exact SHAP explanations in under one second (50-1000x speedup over PermutationSHAP).
• Identified key predictors (e.g., past target) and distinct forecasting patterns (e.g., holiday periods).

Conclusions:

• SHAPformer offers a fast, accurate, and explainable solution for time-series forecasting.
• The model provides valuable local and global insights into forecasting behavior.
• Attention manipulation is an effective strategy for efficient SHAP explanations in Transformer models.