FedRBA: Reliability- and Blindspot-Aware Federated Learning for Imbalanced Credit Risk Prediction

Zihang Zeng; Kai Liu

FedRBA: Reliability- and Blindspot-Aware Federated Learning for Imbalanced Credit Risk Prediction

Zihang Zeng
, Kai Liu^*

^*Corresponding author for this work

XJTLU

Research output: Contribution to conference › Paper › peer-review

6 Downloads (Pure)

Abstract

Federated credit-risk learning combines client heterogeneity with severe class imbalance, causing global models to predict no minority defaults even when ranking metrics remain useful. We introduce FedRBA, which estimates dynamic class blindspots from aggregate validation statistics, reweights local learning and server classifier updates, and separates backbone from class-wise classifier aggregation. Across Fashion-MNIST and two credit datasets, a targeted five-seed audit shows minorityrecall gains over FedProx of 14.20, 22.00, and 13.57 percentage points, with collapse reduced by 11, 19, and 15 points. Against static Logit Adjustment, FedRBA improves both Default Credit settings but trails on Give Me Some Credit. Validation-matched operating points remain comparable rather than uniformly superior. Thus, FedRBA recovers minority operating behavior while exposing a ranking–stability trade-off under extreme heterogeneity.

Original language	English
Publication status	Published - 16 Oct 2026
Event	2026 IEEE 5th International Conference on Algorithms, Data Mining, and Information Technology - Wuhan University of Technology, Wuhan, China Duration: 16 Oct 2026 → 18 Oct 2026

Conference

Conference	2026 IEEE 5th International Conference on Algorithms, Data Mining, and Information Technology
Abbreviated title	ADMIT 2026
Country/Territory	China
City	Wuhan
Period	16/10/26 → 18/10/26

Keywords

Federated learning, Credit risk, Class imbalance, Non-IID data, Minority-class collapse, Reliability-aware aggregation

Access to Document

FedRBA: Reliability- and Blindspot-Aware Federated Learning for Imbalanced Credit Risk PredictionFinal published version, 2.11 MB

Cite this

@conference{2c51c5d2dc3e46d7b07af924d2178f43,

title = "FedRBA: Reliability- and Blindspot-Aware Federated Learning for Imbalanced Credit Risk Prediction",

abstract = "Federated credit-risk learning combines client heterogeneity with severe class imbalance, causing global models to predict no minority defaults even when ranking metrics remain useful. We introduce FedRBA, which estimates dynamic class blindspots from aggregate validation statistics, reweights local learning and server classifier updates, and separates backbone from class-wise classifier aggregation. Across Fashion-MNIST and two credit datasets, a targeted five-seed audit shows minorityrecall gains over FedProx of 14.20, 22.00, and 13.57 percentage points, with collapse reduced by 11, 19, and 15 points. Against static Logit Adjustment, FedRBA improves both Default Credit settings but trails on Give Me Some Credit. Validation-matched operating points remain comparable rather than uniformly superior. Thus, FedRBA recovers minority operating behavior while exposing a ranking{\textendash}stability trade-off under extreme heterogeneity.",

keywords = "Federated learning, Credit risk, Class imbalance, Non-IID data, Minority-class collapse, Reliability-aware aggregation",

author = "Zihang Zeng and Kai Liu",

year = "2026",

month = oct,

day = "16",

language = "English",

note = "2026 IEEE 5th International Conference on Algorithms, Data Mining, and Information Technology, ADMIT 2026 ; Conference date: 16-10-2026 Through 18-10-2026",

}

TY - CONF

T1 - FedRBA

T2 - 2026 IEEE 5th International Conference on Algorithms, Data Mining, and Information Technology

AU - Zeng, Zihang

AU - Liu, Kai

PY - 2026/10/16

Y1 - 2026/10/16

N2 - Federated credit-risk learning combines client heterogeneity with severe class imbalance, causing global models to predict no minority defaults even when ranking metrics remain useful. We introduce FedRBA, which estimates dynamic class blindspots from aggregate validation statistics, reweights local learning and server classifier updates, and separates backbone from class-wise classifier aggregation. Across Fashion-MNIST and two credit datasets, a targeted five-seed audit shows minorityrecall gains over FedProx of 14.20, 22.00, and 13.57 percentage points, with collapse reduced by 11, 19, and 15 points. Against static Logit Adjustment, FedRBA improves both Default Credit settings but trails on Give Me Some Credit. Validation-matched operating points remain comparable rather than uniformly superior. Thus, FedRBA recovers minority operating behavior while exposing a ranking–stability trade-off under extreme heterogeneity.

AB - Federated credit-risk learning combines client heterogeneity with severe class imbalance, causing global models to predict no minority defaults even when ranking metrics remain useful. We introduce FedRBA, which estimates dynamic class blindspots from aggregate validation statistics, reweights local learning and server classifier updates, and separates backbone from class-wise classifier aggregation. Across Fashion-MNIST and two credit datasets, a targeted five-seed audit shows minorityrecall gains over FedProx of 14.20, 22.00, and 13.57 percentage points, with collapse reduced by 11, 19, and 15 points. Against static Logit Adjustment, FedRBA improves both Default Credit settings but trails on Give Me Some Credit. Validation-matched operating points remain comparable rather than uniformly superior. Thus, FedRBA recovers minority operating behavior while exposing a ranking–stability trade-off under extreme heterogeneity.

KW - Federated learning, Credit risk, Class imbalance, Non-IID data, Minority-class collapse, Reliability-aware aggregation

M3 - Paper

Y2 - 16 October 2026 through 18 October 2026

ER -