Temporal Feature Impact Explanation via Dynamic Sliding Window Sampling

Model interpretation methods are essential for understanding the predictions of machine learning models. How-ever, when applied to dynamic multivariate time series data, existing approaches encounter significant challenges, primarily due to the temporal correlations within the same feature at different time points. Ignoring temporal dependencies of time series data during interpretation usually leads to inadequate explanation results. To address these challenges, we propose a novel algorithm called Dynamic Sliding Window Sampling (DSWS) for explaining time series models. This algorithm dynam-ically generates importance scores for the same feature across various time steps by iteratively removing feature sets from specific time periods. Leveraging a sliding time window with an adaptable length, DSWS effectively captures the temporal dependencies of dynamic features and delineates their influence boundaries. Experiments on both synthetic and real world data sets validate DSWS’ applicability for time series data with dynamic temporal dependencies. Furthermore, it outperforms state-of-the-art methods in computational efficiency.