Mondrian inductive conformal classifier for the quantitative structure–odor relationship to estimate prediction confidence

Quantitative structure-odor relationship (QSOR) is important in understanding the complex relationship between chemical structures and perceptible odors and has applications in various industries, such as fragrance design, environmental monitoring, and livestock monitoring. Conventional QSOR models provide point predictions, such as predicting fruity or non-fruity odors, with unreliable confidence, limiting their usefulness in decision-making processes. In this work, Mondrian inductive conformal prediction (MICP) is used to increase the reliability of QSOR models. The proposed MICP framework uses conformal prediction to yield prediction sets, allowing prediction confidence to be estimated. An empirical evaluation using the Leffingwell flavor database demonstrated the efficacy of the MICP framework in improving the accuracy and reliability of QSOR predictions, thereby enhancing the overall performance and trustworthiness of odor prediction models. To the best of our knowledge, this is the first time that conformal prediction based on MICP has been used for odor prediction.