Author: Chen He
Data Visualization tool to Support the Sensemaking and Selection of Machine Learning Models
Good visualizations can make it easier to interpret complex machine learning models and help in exploring and understanding the underlying patterns, distributions, and relationships in the data. Researchers in DataLit project have developed an interactive visualization tool to compare and explore machine learning models. The tool is designed to be accessible to non-AI experts.
When choosing the best machine learning models, looking at performance numbers alone isn’t always enough—especially when different parts of the data matter differently for the tasks. To help people better understand and choose models, this study introduces VMS (Visualization for Model Sensemaking and Selection), an interactive tool designed for end-users, not just technical experts.
VMS combines multiple perspectives for model comparison—how models perform overall (A), how they handle specific data points (B & C), and how individual features influence predictions (D & F). A key feature of VMS is its ability to visualize the contributions of hundreds of features (D), helping users compare models at both detailed and broad levels. Plus, users can focus on a data subset using filters to explore models based on the parts relevant to the task at hand (E).
The researchers tested VMS by using it to compare machine learning models that predict how long patients might stay in the hospital using time-based health data. With 16 medical professionals as end-users, researchers found that VMS helped users understand the models better and choose the optimal ones for different patient groups, even with limited amount of data. Feedback also suggested adding ways to incorporate medical expertise into model training, such as prioritizing certain features for different patient groups.
You can read the article, explore the prototype and or watch a video demo to learn more.