[Paper Review] Perturbation Validation: A New Heuristic to Validate Machine Learning Models

Jie M. Zhang, Mark Harman+3 more|arXiv (Cornell University)|May 24, 2019

Adversarial Robustness in Machine Learning27 references4 citations

TL;DR

This paper proposes Perturbation Validation (PV), a novel heuristic that assesses machine learning model fit by measuring the drop in training accuracy after perturbing training labels, without relying on test data. PV outperforms traditional methods in detecting model fit quality and explains paradoxes in generalization of high-capacity models.

ABSTRACT

This paper introduces Perturbation Validation (PV), a new heuristic to validate machine learning models. PV does not rely on test data. Instead, it perturbs training data labels, re-trains the model against the perturbed data, then uses the consequent training accuracy decrease rate to assess model fit. PV also differs from traditional statistical approaches, which make judgements without considering label distribution. We evaluate PV on 10 real-world datasets and 6 synthetic datasets. Our results demonstrate that PV is more discriminating about model fit than existing validation approaches and it accords well with widely-held intuitions concerning the properties of a good model fit measurement. We also show that PV complements existing validation approaches, allowing us to give explanations for some of the issues present in the recently-debated apparent paradox that high capacity (potentially overfitted) models may, nevertheless, exhibit good generalisation ability.

Motivation & Objective

• To develop a validation method that evaluates model fit without depending on test data.
• To address limitations in traditional statistical validation approaches that ignore label distribution.
• To create a heuristic that better aligns with intuitive expectations of model fit quality.
• To explain the apparent paradox where high-capacity models generalize well despite potential overfitting.

Proposed method

• Perturb labels in the training data to simulate mislabeling.
• Retrain the model on the perturbed training data.
• Measure the rate of decrease in training accuracy as a metric of model fit.
• Use the sensitivity of training accuracy to label perturbations as an indicator of model robustness and fit.
• Compare the resulting accuracy drop rate across different models and datasets.
• Integrate PV with existing validation techniques to enhance interpretability and detection of model issues.

Experimental results

Research questions

• RQ1 How well does PV detect model fit compared to existing validation methods?
• RQ2 Does PV's performance depend on label distribution, unlike traditional approaches?
• RQ3 Can PV explain the generalization paradox in high-capacity models?
• RQ4 How does PV complement established validation techniques?

Key findings

• PV demonstrated superior discrimination in assessing model fit across 10 real-world and 6 synthetic datasets.
• The method showed strong alignment with widely held intuitions about what constitutes a good model fit measurement.
• PV successfully explained the apparent paradox of high-capacity models generalizing well despite potential overfitting.
• The accuracy drop rate under label perturbation served as a reliable indicator of model robustness and fit quality.
• PV was found to be complementary to existing validation approaches, enhancing their interpretability.