Recent advances in neural networks have significantly improved electricity theft detection, achieving higher detection accuracy compared to earlier methods (e.g., support vector machine and decision tree). However, the performance of these networks is still restricted by the class imbalance issue, which causes the neural networks to bias toward classifying unknown users as the majority class (i.e., normal users). While previous works have developed oversampling and data augmentation techniques to alleviate this problem at the data level, these techniques usually replicate existing fraudulent samples or generate similar ones, which can lead to overfitting and, thus, limit model performance. To this end, this article aims to mitigate the class imbalance issue from a novel perspective at the algorithmic level. Specifically, a sample-weighted (SW) loss is proposed to efficiently train neural networks by assigning different weights to samples based on their importance, in contrast to most existing works, which treat all samples equally. Notably, the proposed SW loss is independent of any specific model architecture, meaning that it can be seamlessly integrated with various neural networks to update their weights for electricity theft detection. Simulation results on real-world datasets show that the proposed SW loss outperforms baselines (e.g., binary cross entropy loss, class-balanced loss, oversampling, and data augmentation), with an increase of about 0.27% to 9.78% in mean average precision and 0.14% to 2.92% in the area under the curve, respectively.