Comparative embedding visualizations published at IEEE VIS

I am thrilled to share that our work on comparative embedding visualizations is published at the premier IEEE VIS conference. In the paper, we introduce a powerful framework and computational metrics for comparing embedding visualizations across class-label hierarchies to help discover meaningful differences quickly.

➡️ Test it yourself: https://github.com/OzetteTech/comparative-embedding-visualization

When working with multiple large-scale datasets (e.g., single-cell data), tools like t-SNE, UMAP, PCA, etc. are frequently used for plotting to get an understanding of the class distributions (like cell phenotypes). Unfortunately, when comparing multiple such plots (e.g., from two experimental conditions or two embedding models), the size, complexity, and distribution shifts of the data can make it challenging to identify changes.

Our solution is to compute metrics for groups of points as defined by class labels. This allows us to compare datasets without point correspondences and surface insights at different semantic levels. We do this by constructing a Delaunay neighborhood graph and dividing it into a core, context, and combined region for each class. Using the resulting subgraphs we summarize the class distributions to derive three metrics: confusion, neighborhood, and size. Confusion quantifies the visual intermixing of classes. Neighborhood measures the stability of neighboring class labels. And Size assesses relative class size differences in their local neighborhoods.

To evaluate and apply our framework+metrics we developed and open-sourced a Python-based prototype for Jupyter Notebook-like environments called CEV. Using Rust for computing the metrics, anywidget for the widget architecture, and Jupyter Scatter for rendering, CEV scales to datasets with millions of points. See https://github.com/OzetteTech/comparative-embedding-visualization/tree/main/notebooks for examples.

This fantastic work was led by the wonderful Trevor Manz from Nils Gehlenborg’s lab and started as Ozette’s first graduate research internship. With contributions from our Chief Technology Officer Greg Finak and Chief Data Science Research Officer Evan Greene, this was a fruitful long-term collaboration.

At Ozette we’re trying to push the boundary in single-cell analysis, not just by developing cutting edge machine learning methods but also by expanding the possibilities for efficient data exploration and discovery of actionable insights. This is just the first of several research projects on intelligent and scalable visual data exploration for single-cell data that we’re working on.