Akademik Veri Yönetim Sistemi
FUTURE GENERATION COMPUTER SYSTEMS, cilt.184, ss.1-20, 2026 (SCI-Expanded, Scopus)
As machine learning models are increasingly deployed in real-world production environments, ensuring their reliability under changing data conditions has become a critical challenge. A particularly common issue is covariate shift, where the data they see during inference differ from the data used for training. This can lead to a drop in performance. Standard solutions such as retraining or test-time adaptation require access to model parameters or labeled data, which is not always feasible in practice. Currently, there is no thorough comparison of lightweight distribution correction methods that work without model access or labels, nor is there a standard way to measure both their performance and efficiency. In this study, we introduce three lightweight, model-agnostic correction methods: Mean-Standard Deviation (Mean-Std), Quantile Mapping, and Cumulative Distribution Function (CDF) Matching. These methods serve as preprocessing steps during inference, aligning new input data with the reference distribution without altering the model or requiring retraining. We compared 25 correction methods across 4 classifiers and 4 datasets, running 3213 experiments. Mean-Std ranked first in F1 (0.736 0.12) and EWF1 (0.130), with only 0.4 ms execution time and 0.9 MB memory use, making it 150 times faster than its closest competitor. CDF Matching (EWF1: 0.069) performed best when the distribution shape changed, especially on the Digits dataset. Quantile Mapping (EWF1: 0.065) was top on the Weather dataset, along with CDF Matching. Neural network-based TTA methods (TENT, EATA, SAR) and output-adjustment methods (LAME, T3A) did not work well on tabular data, but our proposed methods improved results across all classifiers and datasets. These results show that simple statistical alignment can match or outperform more complex distribution correction methods while using much less computing power. This makes them practical, plug-and-play solutions for production machine learning systems.