Title: Data and model visualisation for statistical learning problems
Authors: Catherine Hurley - Maynooth University (Ireland) [presenting]
Abstract: Visualization techniques assist in getting to know data prior to modelling, and post-modelling, in exploring and comparing fits, diagnosing lack of fit and understanding predictor effects. We give an overview of some techniques we have developed for visualization in the context of statistical learning, which help address the interpretability deficit. We describe improved methods for exploring predictor importance, predictor interaction and partial dependence model summaries. We use interactive visualisation to dig deeper into model fits by focusing on slices of predictor space, thus investigating local lack of fit, local predictor effects and higher-order interactions. Our techniques are model agnostic and are appropriate for any regression or classification problem. The methods presented are implemented in the R packages vivid and condvis2.