V0.2 dev

[dillondaudert]

To Do

Functionality

• Allow multiple "views" of dataset at fit/transform time ( UMAP#206, UMAP#601, UMAP docs )
  • Discrete metric data with discrete distances ( UMAP#624 )
• Support passing precomputed distances
  • As KNNGraph
  • As matrix
• Helper functions to construct fit/transform config

Tests

• config.jl
• utils.jl
• membership_fn.jl
• neighbors.jl
• simplicial_sets.jl
• embeddings.jl
• optimize.jl
• fit.jl
• transform.jl

Docs

• UMAPConfig component structs and how to use
• UMAPResult/UMAPTransformResult
• What is public API vs. unstable/in-progress functionality
• Tutorials for basic vs. advanced usage
• Documenter (?) - we want docs that auto-build, and have some guarantee to be up-to-date...

[dillondaudert]

Need to find the right code path for categorical (e.g. supervised) views. Ultimately, we want to use the proper simplicial set intersection logic (in coalesce_views - general_simplicial_set_intersection.)

This function takes two views' fuzzy simplicial set representations and combines them. If one (or both?) of these is based on a categorical metric, the pairwise edge probabilities are updated by ...

1. If vertices are same class, technically multiply by exp(0) = 1.
2. If one of the vertices' classes is unknown, multiply by exp(-unknown_dist)
3. If vertices are different classes, multiply by exp(-far_dist).

Some of the complexity comes from representing this fuzzy simplicial set as a matrix - normally, they are sparse matrices, but here this is a dense matrix. The fast_intersection function in the python implementation optimizes this.

In order to dispatch to the proper logic, we may need to create relatively simply FuzzySimplicialSet structs that wrap the graphs. Then fuzzy_simplicial_set((knns, dists), knn_params, src_params::SourceViewParams) can dispatch on knn_params to create the appropriate FuzzySimplicialSet struct.

Might first investigate only creating such a struct for the categorical metrics, dispatching appropriately for the simplicial set intersection.