Mario Boley

Mario Boley

Statistical Machine Learning, Optimisation, Applications to Materials Science and Chemistry

Sitemap

A list of all the posts and pages found on the site. For you robots out there is an XML version available for digesting as well.

Pages

Posts

Blog Post number 4

less than 1 minute read

Published:

This is a sample blog post. Lorem ipsum I can’t remember the rest of lorem ipsum and don’t have an internet connection right now. Testing testing testing this blog post. Blog posts are cool.

Blog Post number 3

less than 1 minute read

Published:

This is a sample blog post. Lorem ipsum I can’t remember the rest of lorem ipsum and don’t have an internet connection right now. Testing testing testing this blog post. Blog posts are cool.

Blog Post number 2

less than 1 minute read

Published:

This is a sample blog post. Lorem ipsum I can’t remember the rest of lorem ipsum and don’t have an internet connection right now. Testing testing testing this blog post. Blog posts are cool.

Blog Post number 1

less than 1 minute read

Published:

This is a sample blog post. Lorem ipsum I can’t remember the rest of lorem ipsum and don’t have an internet connection right now. Testing testing testing this blog post. Blog posts are cool.

portfolio

projects

Human-Simulatable Additive Classification Models

Published:

Generalised additive classification models provide probabilistic predictions by adding the output of individual predictors to form a raw score $a=a_1+\dots+a_k$ and then by mapping this raw score to a probability $p(a)$. This process is considered human-interpretable because of the relative ease with which humans can carry out addition, and this is further enhanced by model fitting methods that restrict the individual predictor output to a small set of round values, i.e., $a_i \in \{-5, \dots, 5\}$. However, even when the resulting raw sum is round and easy to compute, the standard classification machinery based around the logistic transformation $p(a)=e^a/(1+e^a)$ renders it hard for humans to convert raw outputs to probabilities. For instance, a raw model output of $a=1$ corresponds to a probability $e/(1+e) \approx$ 0.7311 and $a=2$ corresponds to $e^2/(1+e^2)\approx 0.8808$.

Rethinking the Data-Driven Discovery of Rare Phenomena (DARE)

Published:

Due to its interest in rare phenomena, scientific discovery typically resembles the proverbial hunt for the needle in a haystack. Materials science seeks stable materials that have desired properties within huge combinatorial spaces. Neuroscience looks for accurate and specific explanations of critical neuro-electric phenomena like seizures or alpha-rhythms in infinite parameter spaces of dynamical brain models. Through this project, we enable the data-driven discovery of such important rare phenomena from high-dimensional observational data collections.

Graph Neural Networks for Guiding Chemical Synthesis of Metal–Organic Frameworks

Published:

This project will explore how graph neural networks (GNNs) can be used to probabilistically predict successful crystallization conditions in the reticular synthesis of novel metal–organic frameworks (MOFs). MOFs are rich chemical structures with important applications in carbon capture and storage, water harvesting from air, hydrogen storage, and catalysis—areas central to addressing global energy and sustainability challenges.

publications

Identifying domains of applicability of machine learning models for materials science

Published in Nature Communications, 2020

Demonstrates how statistical rule learning enables the discovery of trustworthy input ranges of machine learning models for materials properties.

Recommended citation: C Sutton, M Boley, LM Ghringhelli, M Rupp, J Vreeken, M Scheffler. (2020). "Identifying domains of applicability of machine learning models for materials science." Nature Communications. 11(1),4428.
Download Paper

Better Short than Greedy: Interpretable Models through Optimal Rule Boosting

Published in Proceedings of the 2021 SIAM International Conference on Data Mining (SDM), 2021

Improving the accuracy / comprehensibility trade-off of additive rule ensembles via exactly optimising the gradient boosting objective for conjunctive rules.

Recommended citation: M Boley, S Teshuva, P Le Bodic, G Webb. (2021). "Better Short than Greedy: Interpretable Models through Optimal Rule Boosting." SDM.
Download Paper

Relative Flatness and Generalization

Published in Proceedings of the 34th International Conference on Neural Information Processing Systems (NeurIPS), 2021

Why do over-parameterized machine learning models like deep neural networks generalize to unseen data? Under certain conditions, the relative flatness of the loss surface around the parameter values found during training is an explanation.

Recommended citation: Petzka, Henning, Michael Kamp, Linara Adilova, Cristian Sminchisescu, and Mario Boley. "Relative flatness and generalization." Advances in neural information processing systems 34 (NeurIPS 2021): 18420-18432.
Download Paper

Interpretable Machine Learning Models for Phase Prediction in Polymerization-Induced Self-Assembly

Published in Journal of Chemical Information and Modeling, 2023

Prodives interpetable models to predict the morphological outcome of polymerization-induced self-assemblies with a performance that suffices to reduce time-consuming experimentations by practitioners.

Recommended citation: Y Lu, D Yalcin, PJ Pigram, LD Blackman, M Boley. "Interpretable machine learning models for phase prediction in polymerization-induced self-assembly." Journal of Chemical Information and Modeling 63, no. 11 (2023): 3288-3306.
Download Paper

Bayes beats cross validation: fast and accurate ridge regression via expectation maximization

Published in Proceedings of the 36th International Conference on Neural Information Processing Systems (NeurIPS), 2023

Ridge regression is one of the most widely used methods to fit linear regression models. Surprisingly, the established method for estimating its optimal shrinkage hyper-parameter value can be rather inaccurate. We show how a Bayesian formulation can address this issue.

Recommended citation: SY Tew, M Boley, DF Schmidt. (2023). "Bayes beats cross validation: fast and accurate ridge regression via expectation maximization." NeurIPS. 36
Download Paper

Orthogonal Gradient Boosting for Simpler Additive Rule Ensembles

Published in International Conference on Artificial Intelligence and Statistics, 2024

Improving the accuracy / comprehensibility trade-off of rule ensembles and other additive models via proper adaption of boosting with weight correction.

Recommended citation: F Yang, P Le Bodic, M Kamp, M Boley. (2024). "Orthogonal Gradient Boosting for Simpler Additive Rule Ensembles." AISTATS.
Download Paper

Gradient Boosting versus Mixed Integer Programming for Sparse Additive Modeling

Published in Joint European Conference on Machine Learning and Knowledge Discovery in Databases (ECMLPKDD 2025), 2025

How accurate is gradient boosting as an approximation algorithm for the best additive model with a limited number of terms? Theoretically, we show that the approximation gap can be as wide as $1/2$, but how representative is that for the practical performance? An empirical comparison with mixed integer programming paints a differentiated picture.

Recommended citation: Yang, Fan, Pierre Le Bodic, and Mario Boley. "Gradient Boosting Versus Mixed Integer Programming for Sparse Additive Modeling." Joint European Conference on Machine Learning and Knowledge Discovery in Databases (ECMLPKDD 2025): 453-470.
Download Paper

Cortical connectivity, local dynamics and stability correlates of global conscious states

Published in Communications Biology, 2025

Abstract: Waking levels of human consciousness are known to be supported by the integrity of complex structures and processes in the brain, yet how they are exactly regulated by neurobiological mechanisms remains uncertain. Here a space-time-resolved inference-based framework is applied to estimate the neurophysiological variables of a whole-cortex model and analyze the neural mechanism correlates of global consciousness by way of a correlation analysis between behavioural and neural variable time-series. Using magnetoencephalography (MEG) data from 15 participants under Xenon-induced anesthesia, interconnected neural mass models (NMMs) were developed and time-evolving regional neurophysiological variables and inter-regional connectivity strengths were inferred from the data. Analyses revealed significant correlations between consciousness levels and inter-regional connectivity, particularly in posterior parietal, occipital, and prefrontal regions. Moreover, results support a parietal, rather than frontal, network backbone to facilitate global consciousness. Regional level analyses further identified correlates of consciousness within the posterior parietal and occipital regions. Lastly, reductions in consciousness were linked to stabilized cortical dynamics, reflected by changes in the eigenmodes of the system. This framework provides a novel, inference-based approach to investigating consciousness, offering a time-resolved perspective on neural mechanism correlates during altered states.

Recommended citation: Zhao, Y., Tsuchiya, N., Boley, M. et al. "Cortical connectivity, local dynamics and stability correlates of global conscious states". Commun Biol 8 (2025): 1391
Download Paper

talks

An Introduction to Subgroup Discovery

Published:

Subgroup discovery (SGD) is a form of local pattern discovery for labeled data that can help find interpretable descriptors from materials-science data obtained by first-principles calculations. In contrast to global modeling algorithms like kernel ridge regression or artificial neural networks, SGD finds local regions in the input space in which a target property takes on an interesting distribution. These local distributions can potentially reflect interesting scientific phenomena that are not represented in standard machine learning models. In this talk, we go over the conceptual basics of SGD, sketch corresponding search algorithms, and show some exemplary applications to materials-science data obtained by first-principles calculations.

Download Download Slides

</p>

teaching

Teaching experience 2

Workshop, University 1, Department, 2015

This is a description of a teaching experience. You can use markdown like any other post.