Explainable Electricity Theft Detection With Gradient-Weighted Class Activation Mapping

Abstract

Neural networks have been widely used for electricity theft detection recently. However, their decision-making process is often not transparent, which limits the understanding of the basis for their decisions. To address this limitation, this letter proposes an explainable electricity theft detection method with gradient-weighted class activation mapping (Grad-CAM). Specifically, Grad-CAM is extended to generate fraud scores by computing the gradient-based importance of input features, highlighting suspicious activities. Simulation results show that the proposed Grad-CAM can provide accurate and reliable decision rationale. Compared with Shapley additive explanations and local interpretable model-agnostic explanations, the balanced detection score of the proposed Grad-CAM increased by 13.38% and 72.53%, respectively.