Model exploration is useful for explaining fitting algorithms in the classroom situation. For practitioners, visualisation offers a way of examining model behaviour, checking model stability and juxtaposing multiple model fits. Such exploration will lead to improved understanding and possibly discovery of improved fits.
This talk will discuss model exploration techniques which show fitted models on sections of data space, together with observed data near the section. Sections are either chosen interactively by the user, or calculated to reveal some feature of the model. Visualisations then show a path or tour through the selected sections. These techniques are applicable to regression and classification settings, and for fits ranging from linear models, to Bayesian fits to black-box models. The ideas presented here are implemented in the R package condvis. Examples demonstrating model exploration using condvis will be given.
The work is joint with Mark O’ Connell and Katarina Domijan.