Machine Learning Models with a Reject Option to Minimize Prediction Error: Application to Optical Properties of Dye...

Predictions of the optical properties of dyes must be exceptionally accurate because even a small prediction error of, for instance, the wavelength can dramatically misrepresent the dye visible color. To address this challenge, we developed a novel model dubbed 'CasRidge', which implements a machine learning concept known as reject option modeling. This model both predicts the target property(s) and provides the prediction confidence score. We show that by setting a threshold to reject predictions with confidence below a certain level, we improve the overall model accuracy. We benchmarked our model against several existing optical property prediction models, both with and without the reject option policy. We observed a significant performance improvement for most optical properties of dyes when using the reject option policy. As part of methodology development, we have explored the correlation between our learned confidence score and the traditional cheminformatics concept of "applicability domain". We posit that the CasRidge approach can be employed for predicting any target property where very high prediction accuracy is especially critical. We dedicate this study to Dr. Terry Stouch, who always promoted high quality research.