Correcting Class Imbalance in Prior-Data Fitted Networks for Tabular Classification

In three linesPrior-data fitted networks (PFNs) excel at tabular classification but suffer from class imbalance affecting rare classes. This study adapts classical mitigation techniques (thresholding, downsampling) to PFNs, finding thresholding outperforms due to PFN calibration properties, while downsampling provides comparable results with reduced inference cost.

Summary generated by Claude — human-verified