Prof. Yi Mei

Prof. Yi Mei

Professor, Fellow of Engineering New Zealand, IEEE Senior Member

Explainable AI

AI is increasingly trusted with consequential decisions, from scheduling hospital resources to detecting cyber attacks. People need to understand and trust those decisions. My research makes AI explainable in two ways: building models that are interpretable by design, and explaining the decisions of complex models after training.

18papers in this theme
2award-winning papers
1papers with 100+ citations (OpenAlex)
160+citations of our XAI-by-GP survey (OpenAlex)

Introduction

Two routes to explainability

A black-box model, such as a deep neural network, can be explained post hoc — which features mattered, or what would have to change to get a different outcome. Alternatively, the model can be interpretable by design: a short formula evolved by genetic programming can be read directly.

Two routes to explainable AIPOST-HOC EXPLANATIONblack-boxmodelfeature importancecounterfactualslocal surrogatesexplain decisionsafter trainingINTERPRETABLE BY DESIGNpriority = PT + WINQ ÷ NORthe model itself is a readable formula,kept small and dimensionally consistent
Figure 1. Post-hoc explanation versus interpretable-by-design models. Our work spans both.

My Contributions

With my students and collaborators, I have made the following contributions. The papers shown are selected automatically: award-winning papers, papers with 100+ citations, and papers in top venues. Click a title to read the paper, or list entry to see it in my publication list.

1 Genetic programming for explainable AI

We surveyed how GP contributes to explainable AI, both by evolving interpretable models and by producing explanations for black-box models, and use GP to build interpretable local explanations.

2 Explaining evolved decision policies

Even evolved rules can grow large. We explain them by ranking the features that drive each decision and by generating counterfactual what-if scenarios.

3 Counterfactual explanations

A counterfactual tells a user the smallest change that would alter a decision. We evolve sets of diverse counterfactuals so users see several plausible alternatives.

Counterfactual explanationsSeveral minimal changes to the input that each flip the model's decision.decision boundaryoutcome Aoutcome Boriginal input xdiverse counterfactuals x′“If these features had been slightly different, the decision would change.”
Figure 2. Diverse counterfactual explanations: several minimal changes to the input, each of which crosses the decision boundary.

4 Interpretable-by-design heuristics

We keep evolved rules small, simple and physically meaningful by enforcing dimensional consistency, simplifying expressions, and optimising interpretability as an explicit objective.

Interpretability–performance trade-off (schematic)rule size (complexity) →decision error →small and accuratetrade-off front of evolved rules
Figure 3. Schematic interpretability–performance trade-off. Multi-objective GP finds a front of rules; rules near the knee are both small and accurate.

5 Explainable machine learning in applications

We make machine learning explainable in security and healthcare, for example robust SHAP explanations for intrusion detection.

Research Output

Papers per yearStacked bars of journal and conference papers per year.0242017: 0 journal2017: 1 conference120172018: 0 journal2018: 0 conference20182019: 0 journal2019: 0 conference20192020: 0 journal2020: 1 conference120202021: 0 journal2021: 1 conference120212022: 0 journal2022: 3 conference320222023: 1 journal2023: 2 conference320232024: 1 journal2024: 3 conference420242025: 0 journal2025: 2 conference220252026: 2 journal2026: 1 conference32026Journal articles & booksConference papers
Figure 4. My papers in this theme per year (18 in total), identified by title keywords in my publication lists; papers can belong to more than one theme. The 2026 count covers papers published so far this year.

Research Themes

Full publication list →   ← Back to home