VMS: Interactive Visualization to Support the Sensemaking and Selection of Predictive Models

Explore an innovative approach to comparing and selecting machine learning models through interactive visualization in this 26-minute conference talk by Chen He from the Finnish Center for Artificial Intelligence. Learn about VMS (Visualization for Model Sensemaking and Selection), a system designed to support model users in evaluating predictive models from multiple angles. Discover how VMS integrates performance-, instance-, and feature-level analysis, with a particular focus on a feature view that combines value and contribution of numerous features for local and global model comparison. Examine the application of VMS in predicting patients' hospital length of stay using time-series health records. Gain insights from the evaluation results involving 16 medical field participants, which demonstrate VMS's effectiveness in helping users rationalize models and select optimal ones with small sample sizes. Consider future directions for incorporating domain knowledge in model training, such as tailoring feature importance for different patient groups.

Chen He: VMS Interactive Visualization to Support the Sensemaking and Selection of Predictive Models