2022 Optimization Days

HEC Montréal, Québec, Canada, 16 — 18 May 2022

Schedule Authors My Schedule

WA7 - Optimization and Learning I

May 18, 2022 10:30 AM – 12:10 PM

Location: Quebecor (yellow)

Chaired by James-A. Goulet

4 Presentations

  • 10:30 AM - 10:55 AM

    A Robust Framework For Learning Consumer Preferences

    • Shobeir Amirnequiee, presenter, Ivey Business School
    • Joe Naoum-Sawaya, Ivey Business School, Western University
    • Hubert Pun, Ivey Business School

    Consumer preferences play a critical role in numerous marketing and operational decisions. For years, conjoint analysis has been the leading methodology in preference elicitation, while different estimation methods have been proposed to identify consumer preferences with regard to a product or a service. Support vector machines (SVM) have been commonly used as a data-driven estimation method. This paper builds upon ideas from machine learning and robust optimization, and alleviates the SVM-based estimation models sensitivity to feature and label noise. We propose a framework that guarantees the robustness of the solution against feature noise (i.e., the perturbations caused by consumer misconceptions), and handles label noise (i.e., response errors) using a weighting scheme that determines the relevance of the past choices in predicting the future ones. The proposed models have three appealing characteristics. First, they are built on an SVM-based estimation model that is proved to have high predictive accuracy. Second, the decision maker can choose the type and the level of robustness. Third, the decision maker has the option to focus on the parsimony of the solution (i.e., simultaneous feature selection). We perform extensive experiments, using simulated and real-world choice data, and show that the proposed framework offers better prediction accuracy and lower variability
    in the predictions.

  • 10:55 AM - 11:20 AM

    Améliorer la prise de décision à l'aide de modèles d'apprentissage automatique basés sur les données de production dans un contexte d'industrie 4.0

    • Nicolas LeBlanc, presenter, Université Laval
    • Jonathan Gaudreault, Université Laval
    • Anthony Deschênes, Université Laval

    Lors de la fabrication par usinage de pièces métalliques de hautes précisions en usine, il est nécessaire de minimiser le temps de production par pièce afin de maximiser l’efficacité du temps des opérateurs et des machines. Dans un contexte d'industrie 4.0, plusieurs données propres au contexte de fabrication sont disponibles et peuvent être analysées. Le temps de production par pièce est prédit en utilisant des modèles de régression linéaire, de réseau de neurones et de forêts aléatoires appliqués sur les données de l’opérateur, son expérience et le quart de travail. Ces expérimentations effectuées permettent de diminuer le temps de production par pièce par rapport à la moyenne et cela permet d’obtenir un modèle prescriptif servant à orienter les décisions de l’entreprise sur l’affectation des tâches et le temps nécessaire pour les effectuer. En utilisant les mêmes modèles d’apprentissage automatique avec l’historique des données de production, tels que le nombre de pièces non conformes, les changements d’outils et leur nature, il est possible d’obtenir un modèle prédictif de ces données et les relayer aux superviseurs des opérations afin de mieux planifier les futures opérations.

  • 11:20 AM - 11:45 AM

    Design and implementation of a Machine Learning tool for automatizing abstract screening in Systematic Literature Reviews

    • Julia Isabel Serrato Fonseca, presenter, Université Du Québec À Montréal

    When performing a systematic literature review, processes as searching, screening and data extraction can take extensive time and energy, while inducing human error in the selection of the documents. In recent years, Machine Learning (ML) techniques have been applied to accelerate said processes and to help researchers to be more efficient. Specifically in the abstract screening process, some papers in the literature have intended to guide researchers into constructing ML tools. However, we have found that the papers address the subject superficially, in a very general manner, without presenting a case study to exemplify the construction of the tool, without explaining the problems that one can come across nor how to tackle them. We present in this talk the detailed implementation of Machine Learning tools for automatizing part of the abstract screening process, provide some insight on how to improve its performance and how to deal with the inconveniences encountered.

  • 11:45 AM - 12:10 PM

    Solving optimization problems through analytical inference in Bayesian neural networks

    • James-A. Goulet, presenter, polymtl
    • Luong Ha Nguyen, polymtl

    Learning the parameter of deep neural networks is typically tackled using gradient-based optimization. With the tractable approximate Gaussian inference (TAGI) method, we can change this paradigm by learning through closed-form analytical Bayesian inference. In addition to match or exceed the performance of backpropagation-trained networks, it also allows unprecedented applications. In this presentation we will see how we can solve optimization problems using analytical inference in Bayesian neural networks.