Automatic monotonicity detection for Gaussian Processes

TL;DR

This paper introduces a novel approach for automatically identifying monotonic relationships in data using Gaussian Processes with virtual derivative observations, enhancing model interpretability and accuracy.

Contribution

The paper presents a new method for detecting monotonicity in Gaussian Process models, improving explainability and performance in regression and classification tasks.

Findings

High accuracy in detecting monotonic directions on synthetic data

Effective in real datasets for model interpretability

Enhances regression and classification performance near data edges

Abstract

We propose a new method for automatically detecting monotonic input-output relationships from data using Gaussian Process (GP) models with virtual derivative observations. Our results on synthetic and real datasets show that the proposed method detects monotonic directions from input spaces with high accuracy. We expect the method to be useful especially for improving explainability of the models and improving the accuracy of regression and classification tasks, especially near the edges of the data or when extrapolating.