Evento acontece no dia 21 de maio, às 10h, no Anfiteatro do CIn
O Centro de Informática (CIn) da UFPE receberá o Seminário de Pesquisa de título “Machine Learning and Compositional Data: Integrating Constrained Optimization in the Aitchison Space”, que será ministrado pelo professor Michele Gallo, da Universidade de Naples, Itália. A apresentação abordará os benefícios de utilizar os Dados Composicionais (CoDa) como uma variedade Riemanniana, garantindo coerência e invariância de escala. O evento acontecerá no dia 21 de maio, às 10h, no Anfiteatro do CIn, sem necessidade de inscrição prévia para participar.
Confira abaixo mais informações sobre o seminário e o palestrante:
Abstract
Compositional data (CoDa) — defined as strictly positive vectors representing relative proportions of a whole — pose unique challenges in predictive modeling and Machine Learning. Traditional approaches rely on extrinsic transformations, such as the centered or isometric log-ratio, to map constrained data from the simplex into a real Euclidean space. However, when CoDa regression problems lack closed-form analytical solutions, standard Euclidean optimization becomes inadequate and risks generating results that violate the fundamental constant-sum and positivity constraints.
To overcome these structural limitations, this work proposes an algorithmic paradigm shift that natively models the CoDa sample space as a Riemannian manifold. By fully embracing Riemannian geometry, parameter estimation is reformulated as a constrained optimization problem directly within the Aitchison space. This intrinsic approach absorbs the strict constraints of the simplex directly into the geometry of the space itself, ensuring scale invariance and subcompositional coherence by construction. Furthermore, by deriving analytical gradients with respect to the Riemannian metric, the proposed Interior Point Flow (IPF) algorithm provides a mathematically rigorous solution to continuously evolve parameters strictly within the permissible manifold. Crucially, empirical validations demonstrate that this framework elegantly bypasses severe computational bottlenecks—such as covariance matrix singularity—enabling robust estimation even in high-dimensional (p > n) scenarios, paving the way for geometrically coherent Machine Learning architectures.
Shortbio
In 1995, he graduated in Economics and Business from the University of Naples – Federico II, where he subsequently earned a Doctorate in “Total Quality Management” in 2000. He is full professor of Statistics from October 2020 at the University of Naples – L’Orientale. Member of Board of Directors (from November 2024). Delegate of the Rector for Orientation and Tutoring Services (from 2020 to 2022). Elected member of Board of of the Italian Statistical Society (SIS) (from July 2024). Elected treasurer of the Italian Statistical Society (SIS) (from July 2020 to 2024). Vice-Director of the Department of Human and Social Science (from July 2019 to December 2019). President of the degree ‘Political Science and International Relations (L-36)’ (from December 2011 to November 2015). Coordinator of the PhD course ‘Institutions, law and economics of public services’ (from November 2008 to October 2014). Delegate of the Rector for Informatics issues (from 2012 to 2014). President of the Interdepartmental Service Centre for Telematics and Informatics (from January 2007 to December 2011). Quality Assurance Manager for the Scientific Research (from June 2004 to December 2010). PRIMARY RESEARCH INTERESTS: Multivariate data analysis, Compositional data analysis, Rasch analysis, Tensor decomposition.
Comentários desativados