A Framework for Multiclass Contour Visualization - Li et al. 2023

VU Visualisierung 2 (186.833)

Our project uses contour plots to visualize the probability of a class prediction by a mutliclass classification model, given two scalar variables. In addition, we provide a domain-specific language ("DSL"), similar to the DSL propsed in the paper by Li et al. This implementation is modeled after the approach taken in the paper. This DSL provides access to the various configuration options included in our implementation, easily copy-able as JSON text.

While our implementation provides two datasets to serve as example data, it is designed to be used with your own data and models. Included in the repository is a Jupyter Notebook with the basic procedures to...

• automatically isolate every combination of two features
• train a Machine Learning model on these two features
• (for the example datasets we opted for a simple MLP)
• build a grid of your feature space
• get predictions for every combination of feature values
• and save a file in the format our implementation expects.

Using these steps, you can easily create your own model from your own data. Due to the nature of contour plots, we are limited to scalar variables and classification problems. If your data fits these criteria, you can explore the impact of two different features on the prediction. For a higher-feature space multiple plots would need to be created, each trained on the two features and, for example, a third feature with n values to produce n datasets.