Cross-Domain Empirical Risk Minimization for Unbiased Long-Tailed Classification

Abstract

We address the overlooked unbiasedness in existing long-tailed classification methods: we find that their overall improvement is mostly attributed to the biased preference of "tail" over "head", as the test distribution is assumed to be balanced; however, when the test is as imbalanced as the long-tailed training data---let the test respect Zipf's law of nature---the "tail" bias is no longer beneficial overall because it hurts the "head" majorities. In this paper, we propose Cross-Domain Empirical Risk Minimization (xERM) for training an unbiased test-agnostic model to achieve strong performances on both test distributions, which empirically demonstrates that xERM fundamentally improves the classification by learning better feature representation rather than the "head vs. tail" game. Based on causality, we further theoretically explain why xERM achieves unbiasedness: the bias caused by the domain selection is removed by adjusting the empirical risks on the imbalanced domain and the balanced but unseen domain.

Spatial distribution and predictive factors of antenatal care in Burundi: a multilevel analysis and spatial baseline for the third Burundian demographic health survey

“…During the survey, 9389 women aged 15-49 years in all selected households and 4280 men aged 15-59 years in half of the selected households for women were successfully interviewed. For women with multiple births, we consider the data for the most recent pregnancy with the view of minimising recall bias[13]. Women who had not given birth in the last 5 years prior to the survey and women who had not consulted a health care provider were excluded from the study.…”

“…Burundi has currently 18 provinces grouped into 5 health regions (Bujumbura-Mairie, West, South, North and Centre-East).The new province of Rumonge (which was a commune of the province of Bururi) is the result of the merger between three former communes of Bururi (Rumonge, Buyengero and Burambi) and two former communes of Bujumbura-Rural (Bugarama and Muhuta). During the survey, 9389 women aged 15-49 years in all selected households and 4280 men aged 15-59 years in half of the selected households for women were successfully interviewed.For women with multiple births, we consider the data for the most recent pregnancy with the view of minimising recall bias [13] .Women who had not given birth in the last 5 years prior to the survey and women who had not consulted a health care provider were excluded from the study.Therefore, the study sample size consists of 5063 Burundian women aged 15-49 years who reported at least one live birth in the five years preceding the survey.The figure 1 below shows ,at national level , the clusters distribution according to place of residence. The red dots are urban's clusters and the green dots are rural's clusters.…”

Spatial distribution and predictive factors of antenatal care in Burundi: A spatial and multilevel baseline analysis for the third burundian demographic and health survey

“…During the survey, 9389 women aged 15-49 years in all selected households and 4280 men aged 15-59 years in half of the selected households for women were successfully interviewed. For women with multiple births, we consider the data for the most recent pregnancy with the view of minimising recall bias [17]. Women who had not given birth in the last 5 years prior to the survey and women who had not consulted a health care provider were excluded from the study.…”

Balanced Contrastive Learning for Long-Tailed Visual Recognition

“…NIL focuses on the individual supervised learning for each expert, while NBOD facilitates knowledge transfer among multiple experts. Lastly, xERM [64] aims to develop an unbiased, test-agnostic model for long-tailed classification. Grounded in causal theory, xERM seeks to mitigate bias by minimizing cross-domain empirical risk.…”