Explaining the Unexplainable
Machine Learning meets Statistical Interpretation
I work at the intersection of statistics and machine learning, developing methods to make complex models genuinely understandable. Most of my research revolves around ensemble models and the question of what they actually do.
I am a Doctor of Economics from the University of Naples Federico II. My research sits at the boundary between statistical methodology and machine learning, with a focus on Explainable Machine Learning — developing frameworks that make powerful black-box models transparent and interpretable, in ways that are meaningful for researchers and decision-makers alike.
The main thread of my work is E2Tree (Explainable Ensemble Trees), a method that represents Random Forests and other ensemble models through a single, interpretable tree structure. The goal is to produce explanations that are genuinely faithful to what the model does, in both classification and regression settings. Several papers from this line of work have appeared in leading statistical journals.
Beyond interpretability, I work on applied problems in health economics and social science — predicting depression from national survey data, evaluating hospital quality, and studying the relationship between scientific output and patient outcomes. I am part of the K-Synth Research Lab and contribute to Bibliometrix, an R package used by thousands of researchers worldwide for systematic literature reviews and science mapping.
My main line of research. I develop methods — in particular E2Tree — that translate the internal logic of ensemble models like Random Forests into a single interpretable tree, covering both classification and regression settings.
Random Forests, XGBoost, and other ensemble methods, studied both for their predictive properties and their interpretability. Part of this work focuses on how to measure whether an explanation actually captures what a model does.
Quantitative analysis of scientific literature, citation networks, and research trends. I contribute to bibliometrix and Biblioshiny, open-source R tools used by researchers worldwide for systematic literature reviews and science mapping.
Applying machine learning to health and clinical data: depression prediction from national survey data, hospital service quality evaluation, and the relationship between research output and patient outcomes. Part of two PRIN-funded national projects.
Statistical and machine learning methods applied to economic questions — composite indicators, stochastic evaluation frameworks, and public sector performance analysis. The common thread is bringing methodological rigour to problems that matter for policy.
Applying quantitative methods across disciplines — from social surveys and bibliographic databases to environmental and geophysical data. Some of the most interesting problems sit at the boundary between fields, and that is where I find myself most often.
A live snapshot of the work currently moving forward — papers under review, methods in preparation, and projects in the pipeline.
Defining a principled way to measure how faithfully a global surrogate (E2Tree) reconstructs the predictive behaviour of the underlying ensemble — across classification and regression.
Comparing service quality between research and non-research Italian public hospitals through stochastic composite indicators and genetic matching.
Hierarchical Bayesian errors-in-variables model with Gaussian-process depth dependence for saltwater intrusion in the Volturno River. Temperature and dissolved oxygen identified as robustly associated covariates across seasons.
Extending the E2Tree explainability framework to gradient boosting ensembles (XGBoost, LightGBM), working towards a unified interpretation approach across bagging and boosting paradigms.
Establishing formal statistical foundations for the nLoI index, deriving its asymptotic properties and limit distributions as both sample size and ensemble size grow.
Developing inference tools for conditional Shapley-based feature importance, enabling asymptotically valid confidence intervals without resampling.
Connecting regularization mechanisms in gradient boosting to formal hypothesis testing for pairwise feature interactions, yielding calibrated p-values.
Spatial validation, calibrated uncertainty intervals, and a sampling-requirement curve for machine-learning maps of pyroclastic cover thickness around Somma–Vesuvius and the Phlegrean Fields.
Investigating how socio-sanitary conditions and health system configurations shape national longevity trajectories, examining cross-country patterns in population-level lifespan dynamics.
An R package that generates a single human-readable tree explaining the latent similarity structure learned by a tree ensemble. Dissimilarity between observations is derived from terminal-node co-occurrence frequencies — not a competing predictor, but a faithful reconstruction of the ensemble's grouping logic.
Do you see a point of contact, or have an idea worth developing together? Let's find out.
If any of these threads intersect with your own work — methodological (interpretability, ensembles, statistical frameworks) or applied (health, policy, science mapping) — I'd be glad to explore a joint paper, exchange data, or simply trade ideas over a coffee.
@article{aria2025extending,
title = {Extending Explainable Ensemble Trees to Regression Contexts},
author = {Aria, Massimo and Gnasso, Agostino and Iorio, Carmela and Fokkema, Marjolein},
journal = {Applied Stochastic Models in Business and Industry},
volume = {42},
number = {1},
pages = {e70064},
year = {2025},
publisher = {Wiley},
doi = {10.1002/asmb.70064}
}
@article{aria2025predicting,
title = {Predicting depression in Italy using random forest through the {E2Tree} methodology},
author = {Aria, Massimo and Gnasso, Agostino and Rivieccio, Roberta and Siciliano, Roberta},
journal = {Annals of Operations Research},
year = {2025},
publisher = {Springer},
doi = {10.1007/s10479-025-06758-7}
}
@article{aria2024explainable,
title = {Explainable ensemble trees},
author = {Aria, Massimo and Gnasso, Agostino and Iorio, Carmela and Pandolfo, Giuseppe},
journal = {Computational Statistics},
volume = {39},
number = {1},
pages = {3--19},
year = {2024},
publisher = {Springer},
doi = {10.1007/s00180-022-01312-6}
}
@article{aria2021comparison,
title = {A comparison among interpretative proposals for {Random Forests}},
author = {Aria, Massimo and Cuccurullo, Corrado and Gnasso, Agostino},
journal = {Machine Learning with Applications},
volume = {6},
pages = {100094},
year = {2021},
publisher = {Elsevier},
doi = {10.1016/j.mlwa.2021.100094}
}
@article{adamo2021assessment,
title = {Assessment of Sleep Disturbance in Oral Lichen Planus and Validation of {PSQI}: a case-control multicenter study from the {SIPMO}},
author = {Adamo, Daniela and Gnasso, Agostino and others},
journal = {Journal of Oral Pathology \& Medicine},
year = {2021},
doi = {10.1111/jop.13255}
}
@inproceedings{gnasso2025canyou,
title = {``{Can You Explain That?}'' {E2Tree}, {SHAP}, and {LIME} for Interpretable {Random Forests}},
author = {Gnasso, Agostino and Aria, Massimo},
booktitle = {CLADAG-VOC 2025 -- Studies in Classification, Data Analysis, and Knowledge Organization},
publisher = {Springer, Cham},
year = {2025},
doi = {10.1007/978-3-032-03042-9_20}
}
@inproceedings{gnasso2025fromprediction,
title = {From Prediction to Explanation: Interpreting Risk Factors in Health Survey Analytics},
author = {Gnasso, Agostino and Aria, Massimo and Siciliano, Roberta},
booktitle = {CLADAG-VOC 2025 -- Studies in Classification, Data Analysis, and Knowledge Organization},
publisher = {Springer, Cham},
year = {2025},
doi = {10.1007/978-3-032-03042-9_21}
}
@inproceedings{gnasso2025explainable,
title = {Explainable Decision Tree Ensembles},
author = {Gnasso, Agostino and Aria, Massimo and Iorio, Carmela and Fokkema, Marjolein},
booktitle = {SIS 2024 -- Italian Statistical Society Series on Advances in Statistics},
publisher = {Springer, Cham},
year = {2025},
doi = {10.1007/978-3-031-64447-4_21}
}
Are you a university student and need support with your studies, thesis, or academic projects? I can help you in the following areas:
The e2tree package implements the Explainable Ensemble Trees
(E2Tree) methodology. Rather than fitting a CART tree directly on raw data, E2Tree
learns the relational structure that the ensemble has already established: it extracts a
co-occurrence matrix from the trained model and uses hierarchical clustering to build a transparent,
interpretable dendrogram. The result is a global surrogate that faithfully approximates Random
Forests, XGBoost, and other boosting models.
An R package for quantitative research in scientometrics and bibliometrics. Provides a comprehensive workflow for science mapping analysis, supporting data import from major bibliographic databases including Scopus, Web of Science, Dimensions, OpenAlex, PubMed, Cochrane Library, and Lens. As a core developer team member, I contribute to development, maintenance, and the web interface Biblioshiny.
Mapping Scientific Knowledge about Health for Decision-Making. 24-month PRIN project at the intersection of bibliometrics, knowledge synthesis, and health policy (Code: 2022825Y5E, PI: Prof. Massimo Aria). Junior Researcher.
The value of scientific production for patient care in Academic Health Science Centres. A study on the relationship between research output and clinical outcomes across Italian public hospitals (PI: Prof. Corrado Cuccurullo). Junior Researcher.
I work with researchers across several institutions in Italy and abroad, with shared interests in statistics, machine learning, and the methodological problems that sit between them.
Regression analysis, hypothesis testing, survey design, and advanced multivariate methods tailored to your research or business questions.
Development of classification and regression models using Random Forests, XGBoost, and ensemble methods with a focus on interpretability and robustness.
Making black-box models transparent through interpretable tree structures, SHAP values, LIME, and custom XAI solutions for high-stakes applications.
Systematic literature reviews, citation analysis, research trend mapping, and science mapping using Bibliometrix and Biblioshiny.
Custom scripts, data pipelines, dashboards (Shiny, Streamlit), and R/Python package development for reproducible data analysis workflows.
Publication-quality charts, interactive visualisations, and comprehensive analytical reports to communicate your findings effectively.
We discuss your project goals, data availability, and expected outcomes. This first meeting is free and without commitment.
I prepare a tailored proposal outlining the methodology, timeline, and deliverables for your project.
I carry out the analytical work, keeping you updated with regular progress reports and intermediate results.
You receive the final output — reports, code, models — with documentation and follow-up support as needed.
Let’s talk. Whether it’s for a project or just a virtual coffee, you can find me right here.
agostino.gnasso@unina.itDepartment of Economics and Statistics
University of Naples Federico
II
Monte S. Angelo, Via Cinthia, 80126 Napoli, Italy
Room D-22 · Sector D · 2nd Floor · Building 3