Computo

Published on September 9th, 2026 In this work, Manon Perrignon, Magalie Houée-Bigot, Romain Jeantet, Thomas Croguennec and Mathieu Emily tackle a practical challenge in multi-objective optimization: knowing when an evolutionary algorithm has actually converged, instead of stopping after an arbitrary, user-defined number of generations. They compare four existing stopping criteria and propose a new one designed to balance computational cost against solution quality. 👉 Key strengths of their approach: It is benchmarked on classical multi-objective test problems as well as on cases with known optimal solutions, allowing a direct assessment of solution quality. It is stress-tested on a real industrial food process optimization case, not just synthetic problems. The proposed criterion offers a clear trade-off between stopping too early (suboptimal solutions) and stopping too late (wasted computation). 🔎 Why this matters: Evolutionary algorithms are widely used to solve real-world multi-objective problems, but practitioners rarely have a principled way to decide when to stop. This work provides a practical, validated criterion that can be readily applied beyond the food-process context studied here. 👉 Dive into the full article here: 🔗 https://lnkd.in/eZtUuff9 #MultiObjectiveOptimization #EvolutionaryAlgorithms #StoppingCriteria #Statistics

Published on July 7th, 2026 In this work, Emilie Devijver and Adeline Leclercq Samson tackle a subtle but important issue in functional data analysis: how to build confidence bands for an unknown function estimated from repeated noisy observations, when the choice of approximation dimension can silently introduce bias. They revisit the classical method of Sun and Loader and propose an explicit bias correction, along with a data-driven criterion to select the approximation dimension. 👉 Key strengths of their approach: It provides a corrected, unbiased confidence band, addressing a bias issue that naive approaches overlook. It introduces a model selection criterion that automatically picks the approximation dimension, correcting the bias as part of the procedure. It is validated through an extensive simulation study and illustrated on real data. 🔎 Why this matters: Confidence bands are central to communicating the reliability of a functional estimate, but a poorly chosen dimension can make them misleading. This work shows — somewhat counterintuitively — that a properly debiased band is often wider than the naive one, and gives practitioners a principled way to get it right. 👉 Dive into the full article here: 🔗 https://lnkd.in/e-erC5Jj #FunctionalDataAnalysis #ConfidenceBands #DimensionSelection #Statistics

Published on July 15th, 2026 In this work, Jules B., Einar Holsbø, Kristian Svendsen, Vittorio Perduca and Etienne Birmele tackle a critical challenge in pharmacovigilance: detecting drug combinations — beyond simple pairs — that are overrepresented in adverse event reports, using large-scale spontaneous reporting databases like the FDA's FAERS. They propose a computational framework combining disproportionality analysis with the hierarchical structure of drug classifications. 👉 Key strengths of their approach: It leverages the Anatomical Therapeutic Chemical (ATC) classification to explore drug combinations hierarchically, going beyond pairwise interaction analysis. It uses a hypergeometric risk metric together with an MCMC algorithm for robust empirical p-value estimation, and a genetic algorithm to efficiently search the huge space of possible drug cocktails. A post-treatment step based on penalized logistic regression distinguishes genuine pharmacological interactions from effects merely combining individual drugs. 🔎 Why this matters: Rare, high-order drug interactions are notoriously hard to detect with clinical trials and standard signal-detection methods. This framework, implemented in the R package emcAdr and validated on synthetic and real FAERS data, gives pharmacovigilance researchers a scalable tool to flag combinations that warrant closer investigation. 👉 Dive into the full article here: 🔗 https://lnkd.in/es7DuBhT #Pharmacovigilance #ComputationalStatistics #DrugInteractions #PublicHealth

Published on September 9th, 2026 In this work, Manon Perrignon, Magalie Houée-Bigot, Romain Jeantet, Thomas Croguennec and Mathieu Emily tackle a practical challenge in multi-objective optimization: knowing when an evolutionary algorithm has actually converged, instead of stopping after an arbitrary, user-defined number of generations. They compare four existing stopping criteria and propose a new one designed to balance computational cost against solution quality. 👉 Key strengths of their approach: It is benchmarked on classical multi-objective test problems as well as on cases with known optimal solutions, allowing a direct assessment of solution quality. It is stress-tested on a real industrial food process optimization case, not just synthetic problems. The proposed criterion offers a clear trade-off between stopping too early (suboptimal solutions) and stopping too late (wasted computation). 🔎 Why this matters: Evolutionary algorithms are widely used to solve real-world multi-objective problems, but practitioners rarely have a principled way to decide when to stop. This work provides a practical, validated criterion that can be readily applied beyond the food-process context studied here. 👉 Dive into the full article here: 🔗 https://lnkd.in/eZtUuff9 #MultiObjectiveOptimization #EvolutionaryAlgorithms #StoppingCriteria #Statistics

Published on July 7th, 2026 In this work, Emilie Devijver and Adeline Leclercq Samson tackle a subtle but important issue in functional data analysis: how to build confidence bands for an unknown function estimated from repeated noisy observations, when the choice of approximation dimension can silently introduce bias. They revisit the classical method of Sun and Loader and propose an explicit bias correction, along with a data-driven criterion to select the approximation dimension. 👉 Key strengths of their approach: It provides a corrected, unbiased confidence band, addressing a bias issue that naive approaches overlook. It introduces a model selection criterion that automatically picks the approximation dimension, correcting the bias as part of the procedure. It is validated through an extensive simulation study and illustrated on real data. 🔎 Why this matters: Confidence bands are central to communicating the reliability of a functional estimate, but a poorly chosen dimension can make them misleading. This work shows — somewhat counterintuitively — that a properly debiased band is often wider than the naive one, and gives practitioners a principled way to get it right. 👉 Dive into the full article here: 🔗 https://lnkd.in/e-erC5Jj #FunctionalDataAnalysis #ConfidenceBands #DimensionSelection #Statistics

Published on July 15th, 2026 In this work, Jules B., Einar Holsbø, Kristian Svendsen, Vittorio Perduca and Etienne Birmele tackle a critical challenge in pharmacovigilance: detecting drug combinations — beyond simple pairs — that are overrepresented in adverse event reports, using large-scale spontaneous reporting databases like the FDA's FAERS. They propose a computational framework combining disproportionality analysis with the hierarchical structure of drug classifications. 👉 Key strengths of their approach: It leverages the Anatomical Therapeutic Chemical (ATC) classification to explore drug combinations hierarchically, going beyond pairwise interaction analysis. It uses a hypergeometric risk metric together with an MCMC algorithm for robust empirical p-value estimation, and a genetic algorithm to efficiently search the huge space of possible drug cocktails. A post-treatment step based on penalized logistic regression distinguishes genuine pharmacological interactions from effects merely combining individual drugs. 🔎 Why this matters: Rare, high-order drug interactions are notoriously hard to detect with clinical trials and standard signal-detection methods. This framework, implemented in the R package emcAdr and validated on synthetic and real FAERS data, gives pharmacovigilance researchers a scalable tool to flag combinations that warrant closer investigation. 👉 Dive into the full article here: 🔗 https://lnkd.in/es7DuBhT #Pharmacovigilance #ComputationalStatistics #DrugInteractions #PublicHealth

In this work, Nils Baillie, Antoine Van Biesbroeck and Clément GAUCHY tackle a long-standing challenge in Bayesian statistics: how to compute objectively defined “reference priors” — which aim to minimize subjectivity in prior choice — when analytic expressions are intractable. They propose a novel algorithm that leverages variational inference with neural networks to approximate reference priors in a flexible, computationally efficient way. 👉 Key strengths of their approach: - It finds tractable approximations of reference priors over large parametric families defined by neural networks. - It supports constraints to ensure the resulting prior is proper (i.e. valid for posterior inference), avoiding issues with improper priors. arXiv - Thanks to a simple MCMC-based sampling strategy, it remains practical even when the prior density cannot be written in closed form. arXiv 🔎 Why this matters: In many Bayesian applications — especially when data are scarce — the prior has a strong influence; this method reduces the need for arbitrary prior choices. The article combines classical Bayesian “objective prior” theory with modern deep-learning tools, opening the door to scalable Bayesian inference in complex or high-dimensional models. 👉 Dive into the full article here: 🔗 https://lnkd.in/efguetqs #BayesianStatistics #VariationalInference #NeuralNetworks #ReferencePriors #Computo #MachineLearning #ReproducibleResearch

💡 New article in Computo (November 18, 2025): REG-MMD by Pierre Alquier & Mathieu Gerber The authors introduce REG-MMD, a method combining regularisation with the Maximum Mean Discrepancy (MMD) to more robustly compare or align probability distributions. 🔎 The approach aims to: Improve robustness via an additional regularisation term, Leverage MMD as a flexible and powerful distributional distance, Provide a principled compromise between statistical fidelity and computational tractability. 📈 Potential applications include domain adaptation, distribution alignment, generative modelling, and hypothesis testing—any setting where comparing probability distributions is central. 👉 Discover the full article here: 🔗 https://lnkd.in/epFHk3zR #statistics #machinelearning #MMD #regularization #distributionLearning #Computo

💡 New article in Computo (November 5, 2025) : Fast confidence bounds for the false discovery proportion over a path of hypotheses by Guillermo Durand This article introduces a computationally efficient method to compute confidence bounds on the False Discovery Proportion (FDP) across a structured “path” of hypotheses. 👉 The key contribution: speed. Thanks to a tailored algorithmic design, the method efficiently produces FDP upper bounds even for large or complex hypothesis families, which is often a bottleneck in classical multiple-testing workflows. 🔎 Practical impact: In fields like genomics or high-dimensional data analysis, where thousands of hypotheses are tested simultaneously, fast and reliable FDP estimates help ensure strong statistical control while keeping analyses actionable. 📚 Why it matters: Enhances robustness in large-scale testing, Helps reduce false positives in scientific discoveries, Provides a pragmatic tool for statisticians and data scientists facing multiple-testing inflation. 👉 Full article available here: 🔗 https://lnkd.in/egVniX85 #statistics #multipleTesting #FDP #falseDiscoveries #research #Computo

💡 New article in Computo (September 12, 2025): Draw Me a Simulator by Sandrine Boulet & Antoine Chambaz In this article, the authors revisit how we design data simulators for statistical studies, using Variational Auto-Encoders (VAE) trained directly on real data. Instead of relying on simplified parametric models, they explore how neural networks can generate more realistic synthetic datasets. 👉 They introduce rigorous evaluation criteria for simulator quality: beyond matching marginal distributions, they compare how well downstream algorithms trained on synthetic data perform relative to those trained on real data. 🔎 At the heart of the discussion: How do we balance realism and overfitting? What does it mean to “simulate” reality in a statistically faithful way? What philosophical or ethical questions arise when models imitate the real world? 📚 Why it matters: Improves the reliability of simulation studies, Opens doors to simulators closely grounded in empirical data, Invites reflection on what “modelling the real world” truly means. 👉 Read the full article here: 🔗 Draw Me a Simulator — https://lnkd.in/e7AwV4fb #statistics #machinelearning #simulation #VAE #research #Computo