Combining Conformal Prediction and Genetic Programming for Symbolic Interval Regression

Symbolic regression has been one of the main learning domains for Genetic Programming. However, most work so far on using genetic programming for symbolic regression only focus on point prediction. The problem of symbolic interval regression is for each input to find a prediction interval containing the output with a given statistical confidence. This problem is important for many risk-sensitive domains (such as in medical and financial applications). In this paper, we propose the combination of conformal prediction and genetic programming for solving the problem of symbolic interval regression. We study two approaches called black-box conformal prediction genetic programming (black-box CPGP) and white-box conformal prediction genetic programming (white-box CPGP) on a number of benchmarks and previously used problems. We compare the performance of these approaches with two popular interval regressors in statistic and machine learning domains, namely, the linear quantile regression and quantile random forrest. The experimental results show that, on the two performance metrics, black-box CPGP is comparable to the linear quantile regression and not much worse than the quantile random forrest on validity and much better than them on efficiency.