Image credit: Unsplash

Interpretable Network Selection

This project develops interpretable model and network selection methods for high-dimensional data. The motivating problem is that many predictive models can achieve similar accuracy, while telling very different scientific stories. Instead of selecting a single apparently optimal model, this line of work studies sets of models with comparable predictive performance and uses their structure to reveal stable, ambiguous, or context-dependent relationships among variables.

The SWAG framework is central to this direction. It is a wrapper method for sparse learning that emphasizes both predictive ability and interpretability. In applications to genomics and biomedical data, this perspective is especially useful because correlated biomarkers can have competing or even antagonistic roles. The resulting selected model sets can be interpreted as networks, making it possible to identify variables that are consistently active, variables whose effects depend on their context, and variables whose apparent importance is unstable across equivalent models.

The project has been applied to gene selection, inflammatory bowel disease, melanoma, breast cancer microRNAs, neuroimaging, stellar blend classification, sleep epigenetics, and medical decision support. Across these applications, the goal is the same: build statistical learning tools that are not only accurate, but also scientifically interpretable.

Selected related publications include:

  • Swag: A Wrapper Method for Sparse Learning
  • A Predictive Based Regression Algorithm for Gene Network Selection
  • A Paradigmatic Regression Algorithm for Gene Selection Problems
  • Evidence of antagonistic predictive effects of miRNAs in breast cancer cohorts through data-driven networks
  • Is Nonmetastatic Cutaneous Melanoma Predictable Through Genomic Biomarkers?
  • Stellar Blend Image Classification Using Computationally Efficient Gaussian Processes
  • A Multi-Model Framework to Explore ADHD Diagnosis From Neuroimaging Data
  • Epigenetic Impact of Sleep Timing in Children: Novel DNA Methylation Signatures via SWAG Analysis
Model Selection
Roberto Molinari
Assistant Professor in Statistics

My research interests include robust statistics, signal processing, model selection and differential privacy.

Publications

International Journal of Molecular Sciences, 26, (21), 10615
This publication contributes to work in applied statistics, model selection, with a focus reflected in its title: Epigenetic Impact of Sleep Timing in Children: Novel DNA Methylation Signatures via SWAG Analysis.
Project

Scientific Meeting of the Italian Statistical Society, 175–180
This publication contributes to work in applied statistics, model selection, with a focus reflected in its title: Machine Learning and Explainable AI for Type-2 Diabetes Management.
Project

arXiv preprint arXiv:2407.19297
This publication contributes to work in applied statistics, model selection, with a focus reflected in its title: Stellar Blend Image Classification Using Computationally Efficient Gaussian Processes.
Project

Journal of Data Science, 22, (2)
This publication contributes to work in applied statistics, model selection, with a focus reflected in its title: A Multi-Model Framework to Explore ADHD Diagnosis From Neuroimaging Data.
Project

Symposium on Data Science and Statistics
This publication contributes to work in applied statistics, model selection, with a focus reflected in its title: Stellar Blend Image Classification Using Computationally Efficient Gaussian Processes (MuyGPs).
Project

Scientific Reports
This work attempts to explain some contradictory results that have been found regarding the oncogenic or protective roles of miRNAs in breast cancer progression
Project Link to paper

Journal of Emergency and Critical Care Medicine
This study highlights how a highly-promoted predictive-score for admitting patients to intensive care is not applicable in Belgium where other predictive models are more appropriate
Project Link to paper

bioRxiv
This publication contributes to work in applied statistics, model selection, with a focus reflected in its title: Chameleon microRNAs in Breast Cancer: Their Elusive Role as Regulatory Factors in Cancer Progression.
Project

arXiv preprint arXiv:2006.12837
This publication contributes to work in applied statistics, model selection, with a focus reflected in its title: Swag: A Wrapper Method for Sparse Learning.
Project

Melanoma Research, 28, (1), 21–29
This publication contributes to work in applied statistics, model selection, with a focus reflected in its title: Is Nonmetastatic Cutaneous Melanoma Predictable Through Genomic Biomarkers?.
Project

Frontiers in Genetics
A new prediction-based objective function for gene selection is proposed, enabling the identification of small, interpretable models with high predictive power, outperforming alternatives while offering a network of models
Project

Human Genomics
This publication contributes to work in applied statistics, model selection, with a focus reflected in its title: Differentiating Inflammatory Bowel Diseases by Using Genomic Data: Dimension of the Problem and Network Organization.
Project

arXiv preprint arXiv:1511.07662
This publication contributes to work in applied statistics, model selection, with a focus reflected in its title: A Paradigmatic Regression Algorithm for Gene Selection Problems.
Project