Code for: Adaptive Label Error Detection: A Bayesian Approach to Mislabeled Data Detection

This Zenodo repository contains the code and data outputs used to generate the results found in the paper "Adaptive Label Error Detection: A Bayesian Approach to Mislabeled Data Detection", which can be found here: doi:10.48550/arXiv.2601.10084. Adaptive Label Error Detection (ALED) is also implemented in the statlab Python package, which can be found on pypi here (or through Github, here). The jupyter notebook files provided correspond to the data displayed in the tables of the paper as follows: Table I ("PRETRAINED DENSENET121 WITH 5% MISLABELING") and Figure 1: DenseNet-Pretrained.ipynb Table II ("PRETRAINED RESNET50 WITH 5% MISLABELING") and Figure 2: ResNet-Pretrained.ipynb Table III ("RANDOMLY INITIALIZED DENSENET121 WITH 5% MISLABELING") and Figure 3: DenseNet-Untrained.ipynb Table IV ("RANDOMLY INITIALIZED RESNET50 WITH 5% MISLABELING") and Figure 4: ResNet-Untrained.ipynb Figure 5 (varying percentage of mislabeled samples): ResNet-Pretrained_Mislabel_Percentage.ipynb Figure 6 (full ALED workflow: applying ALED algorithm to identify mislabeled classes, then removing them and retraining): ResNet-Pretrained_Full_Workflow.ipynb Table V ("VARYING ALED PROJECTION NUMBER, FIXED ENSEMBLE NUMBER (10) AND LIKELIHOOD RATIO (2)"); Table VI ("VARYING ALED NUMBER OF ENSEMBLES, FIXED PROJECTION NUMBER (2) AND LIKELIHOOD RATIO (2)"); and Table VII ("VARYING ALED LIKELIHOOD RATIO, FIXED PROJECTION NUMBER (2) AND ENSEMBLE NUMBER (10)"): ResNet-Pretrained_HyperParams.ipynb All outputs of the jupyter notebooks (as shown in the last block of each file, if applicable) is provided in results.zip. To reproduce these results, the Python environment file ALED_environment.yml is provided.