Enhancing performance and interpretability of multivariate time-series model through sparse saliency

Explainable time-series modelling is an essential task for modern intelligent transportation systems (ITS). How-ever, balancing accuracy and interpretability in multivariate time series forecasting presents significant challenges. These challenges arise from the necessity to understand the significance of features and their temporal variations. Factors such as autocorrelation in time series and data processing techniques like sliding windows expand feature sets, thereby complicating pattern recognition using traditional post-hoc explanation methods and making the issue even more complex. To overcome these challenges, in this study, we propose a flexible post-process approach which generates sparse and normalized saliency values based on existing saliency generation methods such as GradientSHAP. Additionally, an optional window aggregation and alignment strategy is introduced to align with the original time series dataset, enhancing the intuitive understanding of feature importance. Furthermore, the potential use of sparse saliency for data augmentation to improve the model is explored. Lastly, we utilize naturalistic data from San Francisco airport to demonstrate our approach for ITS time-series prediction and explanation. The evaluation results indicate that integrating sparse saliency from high-performing models not only boosts the performance of XGBoost models by 10.92% but also simplifies model complexity, facilitating easier interpretation.