Interpreting Black-Box Time Series Classifiers using Parameterised Event Primitives

Amidst the remarkable performance of deep learning models in time series classification, there is a pressing demand for methods that unveil their prediction rationale. Existing feature importance techniques often neglect the temporal nature of time series data, focusing solely on segment importance. Addressing this gap, this paper introduces a local model-agnostic method akin to LIME, which generates neighbouring samples by randomly perturbing segments of the original instance. Subsequently, weights are computed for each neighbouring instance based on its distance from the original, elucidating its influence. Parameterised event primitives (PEPs) are then extracted from these perturbed samples, encompassing increasing and decreasing events and local maxima and minima points. These primitives are clustered to form prototypical events that capture the temporal essence of the data. Leveraging these events, computed weights, and black box predictions, a simple linear regression model is trained to provide local explanations. Preliminary experiments on real-world datasets showcase the method's efficacy in identifying salient subsequences and points and their importance scores, thereby enhancing comprehension of produced explanations.