Who Trains Matters: Federated Learning under Enrollment and Participation Selection Biases

2026-04-29 · Source: Machine Learning · Field: Technology & Digital — Artificial Intelligence & Machine Learning · Depth: Expert, quick

Summary

Federated learning (FL) often assumes contributing clients represent the target population, but this assumption can fail due to two distinct selection biases. Enrollment bias arises from eligibility rules like device constraints or user consent, determining which clients are reachable for training. Participation bias, more commonly studied, stems from factors such as battery state or network status, influencing which enrolled clients participate in each communication round. This research formalizes FL under a two-stage selection model and introduces \textsc{FedIPW}, an inverse-probability-weighted aggregation scheme designed to recover the target-population mean update. Additionally, it proposes a limited-information aggregate-calibration extension that uses known target-population summaries to reweight the enrolled sample, partially correcting enrollment bias when client-level covariates for non-enrolled clients are unavailable. The study includes an algorithm-agnostic optimization analysis demonstrating that incomplete selection correction can lead to a non-vanishing bias floor, validated by experiments on synthetic federated logistic regression.

Key takeaway

For research scientists developing federated learning systems, understanding and mitigating enrollment and participation biases is crucial. Your models' performance on the true target population can be significantly degraded by these biases, even with existing participation bias corrections. Consider integrating \textsc{FedIPW} or its aggregate-calibration extension to ensure your shared models accurately reflect the target population, especially when client enrollment is non-random.

Key insights

Selection biases in federated learning, particularly enrollment bias, can cause persistent mismatches between training and target-population objectives.

Principles

• Client representativeness is critical in FL.
• Incomplete bias correction leaves a bias floor.

Method

\textsc{FedIPW} uses inverse-probability weighting to correct for two-stage selection bias in federated learning, recovering the target-population mean update. An aggregate-calibration extension uses population summaries for partial correction.

In practice

• Implement \textsc{FedIPW} for biased FL datasets.
• Use aggregate calibration with limited client data.

Topics

• Federated Learning
• Selection Bias
• Enrollment Bias
• Participation Bias
• FedIPW

Best for: Research Scientist, AI Scientist, Machine Learning Engineer

Related on AIssential

• Conflict-Aware Federated Fine-Tuning of Large Language Models with Mixture-of-Experts — FC-MoE resolves expert conflicts in federated MoE LLM fine-tuning, improving stability and performance in non-IID settings.
• Who Uses AI? Platforms, Workforce, and AI Exposure — AI exposure measures from platform data are biased by user demographics, not true workforce exposure, distorting labor market impact estimates.
• Robustness of Refugee-Matching Gains to Off-Policy Evaluation Choices — Algorithmic refugee matching gains are robust across diverse off-policy evaluation methods, confirming original findings.
• Less Data, Faster Training: repeating smaller datasets speeds up learning via sampling biases — Smaller datasets with repetition accelerate learning by inducing beneficial sampling biases that adjust layer-wise growth.
• Recent discoveries on the acquisition of the highest levels of statistical fallacies — Selection bias can create spurious negative correlations among highly selected groups.
• Effects of Training Data Quality on Classifier Performance — Classifier performance degrades universally with training data quality, leading to coincidentally correct but flawed decisions.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Machine Learning.