Use of SHAP for Interpreting a Drone ClassificationModel Based on Statistical Radar Features
Seetharaman Raja, Subhra Kanti Das, Yung Sze Gan, Kah How Teo · 2023
The aim of this paper is to interpret a Light Gradient Boost-based model developed for generic radar classification problems. Light gradient boosting machine (LGBM) being a means of accelerating gradient-boosting decision trees, is apparently helpful in real-time radar-based applications. In this paper, LGBM is developed around statistical features which include the maximum, average, standard deviation, skew, kurtosis, and Shannon entropy of the real-valued vectors. Vectors are complex and real-valued data structures that are generated through the extraction of target range cells along the Doppler dimension at the target range. A conventional Explainable Artificial Intelligence (XAI) method has been used to analyze the trained LGBM model, thereby leading to some recommendations for arriving at features that required to be tuned. SHAP (Shapley Additive Explanations), a post-hoc model agnostic XAI method, is used to provide a global explanation by considering the entire set of data instances followed by an explanation with respect to the local impact of particular data instances on the overall score.