Token-level Optimal Transport Graph Alignment — token_ot_graph

Align two graphs (node sets) by representing each node as a **bag of context tokens** and aligning nodes via: 1) multi-view node embeddings (late-fusion similarity), and 2) token-level entropic optimal transport (Sinkhorn) between node bags.

Usage

token_ot_graph_align(data, ...)

# S3 method for class 'hyperdesign'
token_ot_graph_align(
  data,
  anchors = NULL,
  preproc = multivarious::center(),
  ncomp = 20L,
  control = token_ot_graph_align_control(),
  ...
)

# S3 method for class 'list'
token_ot_graph_align(data, ...)

Arguments

data: A `hyperdesign` object (2 domains) or a list of two matrices.
...: Additional arguments (reserved for future extensions).
anchors: Optional unquoted column name in each domain's design table providing anchor identifiers (shared values indicate known correspondences). Use `NULL` for unsupervised alignment.
preproc: Optional preprocessing (default: `multivarious::center()`).
ncomp: Integer number of components stored in `s` (scores) and `v` (loadings). Used for the baseline embedding view when available.
control: Control settings from [token_ot_graph_align_control()].

Value

A `multiblock_biprojector` with subclass `token_ot_graph_align`. Extras include `transport_plan` (sparse coupling), `assignment` (greedy or Hungarian), and `diagnostics`.

Details

This implementation is designed to stay sparse and efficient by generating a top-k candidate set per source node (ANN / kNN) and running Sinkhorn only on the candidate bipartite graph.