Surrogate model approach for efficient conformal prediction approximation

Accurate predictive modeling plays a crucial role in optimizing system performance, preventing failures, and refining maintenance strategies. While traditional modeling techniques, particularly black-box models, offer point predictions, the demand for predictions accompanied by measurable confidence has grown, particularly in high-risk scenarios. The Conformal Prediction (CP) framework provides a means to establish an error rate bound for predictions. Inductive conformal prediction necessitates calibration sets with a sufficient number of instances to uphold the chosen confidence level. This requirement entails a separate set of instances, referred to as the calibration set, which was not utilized during model training. Such a demand presents a challenge, particularly for intricate systems that are expensive to evaluate and have limited datasets. In our study, we introduce a novel framework that seamlessly integrates surrogate modeling with conformal prediction to effectively tackle these obstacles. Our method constructs a surrogate model to interpolate the calibration set and p-value. This approach significantly enhances the model's prediction capacity, facilitating more informed decision-making. To validate the efficacy of our proposed method, we apply it to a numerical example. The results underscore its capability to reduce dependency on large sample size, pushing the boundaries of the field toward more efficient and precise uncertainty quantification.