Infinite random forests for imbalanced classification tasks (2408.01777v1)

Abstract: This paper investigates predictive probability inference for classification tasks using random forests in the context of imbalanced data. In this setting, we analyze the asymptotic properties of simplified versions of the original Breiman's algorithm, namely subsampling and under-sampling Infinite Random Forests (IRFs), and establish the asymptotic normality of these two models. The under-sampling IRFs, that tackle the challenge of the predicting the minority class by a downsampling strategy to under-represent the majority class show asymptotic bias. To address this problem, we introduce a new estimator based on an Importance Sampling debiasing procedure built upon on odds ratios. We apply our results considering 1-Nearest Neighbors (1-NN) as individual trees of the IRFs. The resulting bagged 1-NN estimator achieves the same asymptotic rate in the three configurations but with a different asymptotic variance. Finally, we conduct simulations to validate the empirical relevance of our theoretical findings.