Misclassification Analysis: Lasso Feature and Annotations

The application currently is able to conduct unsupervised classification on the full-sized embeddings, and then color code the embeddings 2D projection accordingly.

During workshops, while evaluating the unsupervised classification accuracy, we often interested in "boundary" area where most misclassification occurs. The exploration/analyzing steps are typically done by

1. swap color code between the unsupervised labels & the true labels
2. identify the misclassified region(s), and the color code boundaries on the 2D plot
3. click through the points around that region to browse the misclassified image + metadata. these are essentially hard cases and we'd like to understand why they failed, and how can we improve our classification from here.

As indicated above, there's no way to annotate the misclassified data points.

Ideally, in addition to the existing metadata column, the user should be able to add additional text-based columns during analysis.

1. create new column
2. select data points from 2D plane, preferably using lasso for batch selection
3. add annotation to the field values of the selected observation indexes
4. Repeat step 1-3

An additional note on the reproducibility:

A box on a t-SNE plot selects a projected neighborhood, the same rows may sit elsewhere under UMAP/PCA or a different seed. This annotation feature requires better design on provenance, exporting

image uuid
...
projection_x | projection_y
kmeans labels
annotation columns

with these info, the user is now fully able to reproduce the scatter plot and analysis outside of the application session.

@thompsonmj please feel free to comment and add suggestions, thanks!