Adaptive Federated Data Cleaning with Explainability: A Basic Threshold-Driven Approach for Heterogeneous Data Environments

Automated data cleaning is critical for ensuring data quality and robustness of machine learning models. However, modern data environments are increasingly decentralized and heterogeneous, making centralized cleaning methods less viable, particularly when privacy is a concern. In this paper, we propose a novel framework for adaptive data cleaning based on a simple threshold-driven algorithm within a federated learning context. Our approach removes outliers by employing statistical measures (mean and standard deviation) to identify anomalies across distributed nodes. Additionally, we integrate explainability features so that each cleaning decision is transparent to end users. Experimental evaluations on both synthetic and real-world datasets indicate that our method yields notable improvements in data quality while preserving user privacy. We discuss current limitations and outline future avenues for enhancing scalability and extending the framework to handle multimodal data.