Jean-Loup Loyer

Jean-Loup Loyer
BioDescription
Session - Time
An aerospace engineer and data enthusiast, Jean-Loup is a PhD student enrolled at the Instituto Superior Tecnico Lisbon and is currently a research visitor in MIT's AeroAstro Department. For his industrial sponsor Rolls-Royce plc, he is leveraging Machine Learning techniques at a large scale with R, SAS and Python to better predict the maintenance of jet engines. His applied research covers several areas of statistical analysis: econometric modeling, analysis of unevenly spaced and highly correlated multivariate time series, feature extraction and selection, sensitivity and marginal dependence analysis. He has previously held professional positions as an analyst in the French Prime Minister office and as an engineer for a biotech start-up, for a laboratory of theoretical physics and for the French Space Agency.Problems at the Big Data scale regularly involve hundreds to thousands of features and millions to billions of observations. Data scientists are often interested in identifying only a few dozens of the most relevant features in order to generate actionable analyzes. Since the density distributions of the features and their interactions are usually very complex, aggregating the results from several sophisticated feature selection techniques often yields more robust results. In this presentation, we will show how R can be used in practice to select features on a large scale, based on various feature selection techniques - traditional statistical tests (R packages "stats" and "fBasics"), information theory (infotheo, mpmi) and Machine Learning (e1071, class, randomForest, gbm) - coupled with parallel code (foreach, doParallel) and distributed computing (Rmpi). Elementary theoretical aspects will be illustrated on a complex real dataset related to the predictive maintenance of jet engines, covering data visualization, R codes, algorithmic complexities and computational issues.Data Science & Engineering
SAT PM