Jan 23, 2024

Better prediction intervals with Perpetual Learning

This blog post explores the advantages of Perpetual Learning, a novel approach that significantly accelerates the learning process and delivers superior prediction intervals compared to traditional methods. We specifically compare Conformalized Quantile Regression (CQR) implemented with Perpetual Learning against the CQR implementation within the MAPIE open-source package.

The table below summarizes the comparison results for a target coverage rate of 0.9 using the California Housing dataset. Both MAPIE CQR and Perpetual CQR surpass the target coverage rate, demonstrating their effectiveness. Notably, Perpetual CQR achieves a remarkable 35% reduction in the average width of prediction intervals, indicating a significant improvement in prediction accuracy.

Table 1: Perpetual CQR vs. MAPIE CQR
Prediction Interval
Average Width
MAPIE CQR 0.900 0.905 2.222
Perpetual CQR 0.900 0.907 1.450

These findings highlight the substantial benefits of Perpetual Learning for unceartinty quantification with conformal prediction. By enabling faster learning and narrower prediction intervals, Perpetual Learning empowers users to make more informed and reliable decisions based on their data.