Correcting Class Imbalance in Prior-Data Fitted Networks for Tabular Classification
Researchers have adapted classical techniques to address class imbalance in Prior-Data Fitted Networks (PFNs) for tabular classification. They found that thresholding performs exceptionally well due to PFNs' calibration characteristics. Downsampling also proved effective, offering reduced computational cost for inference alongside comparable performance, leveraging PFNs' strong limited-data capabilities. AI
IMPACT Introduces novel techniques to improve the performance of AI models on tabular data, particularly in scenarios with imbalanced datasets.