RESEARCH SUMMARY

Predictive modeling techniques are increasingly being used by data scientists to understand the probability of predicted outcomes. However, for data that is high-dimensional, a critical step in predictive modeling is determining which features should be included in the models. Feature selection algorithms are often used to remove non-informative features from models. However, there are many different classes of feature selection algorithms. Deciding which one to use is problematic as the algorithmic output is often not amenable to user interpretation. This limits the ability for users to utilize their domain expertise during the modeling process. To improve on this limitation, we developed INFUSE, a novel visual analytics system designed to help analysts understand how predictive features are being ranked across feature selection algorithms, cross-validation folds, and classifiers. We demonstrate how our system can lead to important insights  in a case study involving clinical researchers predicting patient outcomes from electronic medical records.

 

teaser

Moore Sloan Poster Research Project