Boundary Focal Loss for Class Imbalanced Learning

The problem of class imbalance has become a severe change in data mining. While handling class imbalance, in addition to sampling strategies and cost-sensitive learning methods, designing loss functions is a crucial approach in neural network. The focal loss function had been proven to be an effective method. This paper proposes a novel loss function, named boundary focal loss (BFL) and validate it in three experiments. The experiment results show that BFL has better performance than cross-entropy and FL on imbalance dataset.