INFUSE: Interactive Feature Selection for Predictive Modeling of High Dimensional Data.

Area of Science:

• Data Science
• Machine Learning
• Medical Informatics

Background:

• Predictive modeling is crucial for understanding outcome probabilities.
• High-dimensional data presents challenges in feature selection for predictive models.
• Current feature selection algorithms lack user interpretability, hindering domain expertise integration.

Purpose of the Study:

• To develop a novel visual analytics system, INFUSE, to enhance the understanding of feature selection processes.
• To enable analysts to interpret predictive feature rankings across different algorithms, cross-validation folds, and classifiers.
• To facilitate the integration of domain expertise into predictive modeling.

Main Methods:

• Development of INFUSE, a visual analytics system.
• Evaluation of feature ranking visualization across multiple feature selection algorithms.
• Case study involving clinical researchers predicting patient outcomes from electronic medical records.

Main Results:

• INFUSE provides insights into how predictive features are ranked.
• The system aids analysts in understanding feature importance across different analytical choices.
• Demonstrated utility in a clinical research case study for outcome prediction.

Conclusions:

• INFUSE improves the interpretability of feature selection in predictive modeling.
• The system empowers users to leverage domain knowledge effectively.
• Visual analytics can significantly enhance the application of predictive modeling in fields like clinical research.