C-SHAP for Time Series: An Approach to High-Level Temporal Explanations

Purpose: Explainable (XAI) aims to increase the transparency of AI systems. In literature, many XAI techniques for time series provide explanations using point-based attribution scores. SHAP is one such algorithm that can be used to determine attribution scores. However, localised methods fail to capture the influence of high-level patterns on model reasoning. Concept-based XAI provides explanations in terms of high-level features. In this research, we present C-SHAP, an approach that relies on SHAP to provide attribution scores for concepts.

Contributions: Our first contribution is the presentation of two approaches to concept-based XAI for time series. The first approach is fully model-agnostic and post-hoc. The second approach relies on a concept-informed model, where the model is directly trained on the concepts under inspection. Our second contribution is the implementation of C-SHAP using multiple decomposition algorithms (Discrete Wavelet Transform, Empirical Mode Decomposition, and a custom decomposition) to provide different explanations. Our third contribution is a custom decomposition including components selected for human-centred interpretability: bias, trend, scale, low frequency, change in variance, and high frequency. Our fourth contribution is the validation of C-SHAP in two domains, human activity recognition and predictive maintenance, to showcase its generalizability.

Methods: In our implementation, we construct concepts using time series decomposition. To determine the attribution of concepts, we apply SHAP using concept masking, where we determine the attribution of concepts by replacing them with uninformative concepts. C-SHAP is evaluated on three tasks: the classification of locomotion activities from accelerometer data in the OPPORTUNITY dataset, the detection of anomalous samples from engine noise in the FordA dataset, and the prediction of the remaining useful life of an engine from pressure levels in the PHM Turbofan dataset. We present anecdotal evidence and perform stability and feasibility analyses.

Results: Using our custom decomposition, we find that ‘Bias’ and ‘Scale’ are specifically influential for locomotion models trained on the OPPORTUNITY dataset, ‘High frequency’ is influential for anomaly detection models trained on the FordA dataset, and ‘Trend’ is influential for the RUL prediction models trained on the Turbofan dataset. We argue these explanations match intuition and provide further validation through a feasibility analysis. Additionally, we provide an indication of stability for C-SHAP.

Conclusion: In this research, we present a SHAP-based approach to concept-based XAI. C-SHAP is generally applicable to time series use cases, only requiring the selection of a concept construction algorithm. In future work, the selection of concepts for different domains should be explored and verified through user studies. Advantages of C-SHAP include its high-level explanations, the custom control over concept selection, and its post-hoc, optionally model-agnostic, explanations.