Analysis of machine learning prediction reliability based on sampling distance evaluation with feature decorrelation

Abstract

Despite successful use in a wide variety of disciplines for data analysis and prediction, machine learning (ML) methods suffer from a lack of understanding of the reliability of predictions due to the lack of transparency and black-box nature of ML models. In materials science and other fields, typical ML model results include a significant number of low-quality predictions. This problem is known to be particularly acute for target systems which differ significantly from the data used for ML model training. However, to date, a general method for uncertainty quantification (UQ) of ML predictions has not been available. Focusing on the intuitive and computationally efficient similarity-based UQ, we show that a simple metric based on Euclidean feature space distance and sampling density together with the decorrelation of the features using Gram–Schmidt orthogonalization allows effective separation of the accurately predicted data points from data points with poor prediction accuracy. To demonstrate the generality of the method, we apply it to support vector regression models for various small data sets in materials science and other fields. We also show that this metric is a more effective UQ tool than the standard approach of using the average distance of k nearest neighbors ( k = 1–10) in features space for similarity evaluation. Our method is computationally simple, can be used with any ML learning method and enables analysis of the sources of the ML prediction errors. Therefore, it is suitable for use as a standard technique for the estimation of ML prediction reliability for small data sets and as a tool for data set design.

Keywords

• analysis
• learnings
• predictions
• reliability
• samplings
• distance