Adaptive Label Cleaning for Error Detection on Tabular Data

Existing supervised methods for error detection require access to clean labels to train the classification model. While the majority of error detection algorithms ignore the harm of noisy labels to detection models. In this paper, we design an effective approach for error detection when both data values and labels may be noisy. Nevertheless, we present AdaptiveClean, a method for error detection on tabular data with noisy training labels. We introduce an effective strategy that can choose the most representative instance to clean. For feature extraction, we use the existing four error detection algorithms for handling multiple types of errors. To reduce the negative effect of noisy training labels on the classification model, we use an adaptive label-cleaning method by training any arbitrary ML models iteratively. Our approach can not only prioritize erroneous instances but also clean noisy labels that affect the classifier primarily. Performance evaluation using five different datasets shows that AdaptiveClean excels over the best baseline error detection system by 0.01 to 0.12 in terms of F1 score.