A1286
Title: SWAG: Interpretation, replicability and statistical inference with multiple simple models
Authors: Roberto Molinari - Auburn University (United States) [presenting]
Stephane Guerrier - University of Geneva (Switzerland)
Nabil Mili - University of Lausanne (Switzerland)
Yagmur Yavuz-Ozdemir - Auburn University (Switzerland)
Samuel Orso - University of Geneva (Switzerland)
Cesare Miglioli - University of Geneva (Switzerland)
Gaetan Bakalli - Emlyon Business School (France)
Abstract: Decision-making based on data analysis goes through an exploratory phase, which usually leads to conclusions that rely on the single "best" interpretation/prediction of the relationships detected in the data. While this approach is standard, in many cases, the uncertainty in the data and/or the presence of latent (unobserved) variables does not justify the reliance on a single model, which may often not respond to the user's needs or may even clash with their domain expertise. Recent literature has focused on the selection of "sets" of simple and accurate models that can be selected according to the needs and expertise of analysts (Rashomon Set Theory). In this direction, a heuristic algorithm is presented called the sparse wrapper algorithm (SWAG), which has been used in various applied studies and addresses different practical needs, starting from how multiple (sparse or simple) models can address issues of replicability and can be interpreted jointly in a network allowing them to be used, individually or jointly, to accurately predict outcomes or optimize different user-defined decision criteria. Newly proposed inference tools for this new multi-model paradigm (Rashomon Inference) are also presented.
