Introducing ShaTS: A Shapley-Based Method for Time-Series Models

Standard Shapley-based methods for model explainability often fail on time-series data by ignoring temporal dependencies and becoming computationally expensive. A novel method called ShaTS (Shapley-based Time Series) addresses these limitations by grouping features before calculating their importance values. ShaTS offers three distinct grouping strategies—Temporal, Feature, and Multi-Feature—to analyze data based on time instants, individual features over time, or logical feature sets. This approach improves computational efficiency and delivers more accurate, interpretable explanations for time-series models, as demonstrated on the SWaT dataset for anomaly detection.