Fig. 5
From: Addressing label noise for electronic health records: insights from computer vision for tabular data
Ablation study showing model performance with NCR with varying NCR hyperparameters for PUH and BH test sets, across varying a NCR starting epochs, b NCR weights, and c number of nearest neighbors (k). Results presented for 20% error in cases and 0.5% error in controls, and 40% error in cases and 0.5% error in controls