Quals research talk: Human Centered Model Selection Through Visual Analytics
Abstract
Model selection is a classic machine learning task to choose an
algorithm and its parameters in order to optimize some objective
function. Objective functions generally decompose into two parts - a
measure of loss on the training set that drives our model to match the
training distribution, and some regularizing penalty that drives our
model to generalize well to heldout testing data. However, such
objective functions don't take into account how those models are used.
A model used by an automated machine vs a model used to inform
intervention in a patient's health car plan have wildly different
constraints and costs. This results in a bias towards black box
models that increasingly do not fit the wide usage scenarios of data
science in the big data world.
We propose a formulation of model selection that takes into account
the human using the chosen model by adding a human penalty into our
objective function. We begin by offering examples of data analysis
tasks that don't fit into the classic model selection formulation. We
then highlight three recent and ongoing projects in which visual
analytics is used to drive model selection that is cognizant of the
human penalty. These projects cover classical machine learning models
like classification and regression as well as deep learning.