CAMRI Loss: Improving the Recall of Prescribed Classes without Sacrificing Accuracy

Abstract

In real-world applications of multiclass classification models, misclassification of important classes (e.g., stop sign) can be significantly more harmful than misclassification of other classes (e.g., no parking). Therefore, it is essential to improve the recall of important classes while maintaining overall accuracy. To achieve this, we have empirically found that concentrating on improving the separability of important classes is an effective way. Existing methods are not suitable for the purpose because they cannot specifically improve the separability of important classes. Then, we propose a loss function, the Class-sensitive Additive Angular Margin (CAMRI) loss, which explicitly gives loss for the feature space. CAMRI loss relatively reduces the variance of important classes by adding a penalty to the angle between important class features and corresponding weight vectors. Experiments on multiple datasets showed that CAMRI loss can improve the recall of specific classes without sacrificing accuracy, with an improvement of up to 9%.