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

2025-12-23

AI Summary

Time-series forecasting interpretability is hindered by existing methods (e.g., SHAP), which rely on strong feature independence assumptions and computationally expensive counterfactual resampling—compromising real-time applicability and trustworthy attribution. This paper proposes an interpretable Transformer framework tailored for time-series forecasting, introducing the first *sampling-free* SHAP explanation paradigm: it directly models subset feature contributions via attention masking, eliminating independence assumptions and counterfactual sampling. The model integrates temporal embeddings, positional encoding, and attention manipulation to generate both local and global feature importance scores in milliseconds (<1 second—over 1,000× faster than baseline methods). On synthetic benchmarks, explanation fidelity reaches 98.7%; on real-world power load forecasting, it achieves state-of-the-art MAE while accurately identifying critical historical load patterns and holiday-related anomalies.

Technology Category

• Machine Learning: Transparent, Interpretable, Explainable ML
• Humans and AI: Explainable AI (XAI) for Human Understanding
• Computer Vision: Interpretability, Explainability, and Transparency

Application Category

• User Modeling, Personalization and Recommendation: Explainable and interpretable methods for personalization
• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking
• Web Mining and Content Analysis: Large pretrained models with web data

Abstract

Time-series forecasts are essential for planning and decision-making in many domains. Explainability is key to building user trust and meeting transparency requirements. Shapley Additive Explanations (SHAP) is a popular explainable AI framework, but it lacks efficient implementations for time series and often assumes feature independence when sampling counterfactuals. We introduce SHAPformer, an accurate, fast and sampling-free explainable time-series forecasting model based on the Transformer architecture. It leverages attention manipulation to make predictions based on feature subsets. SHAPformer generates explanations in under one second, several orders of magnitude faster than the SHAP Permutation Explainer. On synthetic data with ground truth explanations, SHAPformer provides explanations that are true to the data. Applied to real-world electrical load data, it achieves competitive predictive performance and delivers meaningful local and global insights, such as identifying the past load as the key predictor and revealing a distinct model behavior during the Christmas period.

Problem

• Develops a fast, sampling-free explainable time-series forecasting model
• Addresses SHAP's inefficiency and feature independence assumptions for time series
• Provides accurate local and global explanations for Transformer-based forecasts

Innovation

• Sampling-free SHAP for Transformer time-series forecasting
• Attention manipulation for feature subset predictions
• Generates explanations in under one second