Generate metapath2vec embeddings — rs_metapath2vec • genewalkR

Runs metapath-constrained random walks over a heterogeneous graph and trains the skip-gram model on them (metapath2vec, or metapath2vec++ with per-type negative sampling).

Usage

rs_metapath2vec(
  node_ids,
  node_types,
  from,
  to,
  weights,
  metapath,
  metapath_plus,
  metapath2vec_params,
  embd_dim,
  directed,
  seed,
  verbose
)

Arguments

node_ids: Character vector. Node identifiers.
node_types: Character vector. Node type per node, same length as node_ids.
from: Integer vector. 1-based indices into node_ids for edge origins.
to: Integer vector. 1-based indices into node_ids for edge destinations.
weights: Optional numeric vector. Edge weights.
metapath: String. Hyphen-separated metapath closing on its starting type, e.g. "gene-pathway-gene".
metapath_plus: Boolean. Per-type negative sampling (metapath2vec++).
metapath2vec_params: Named list. Training parameters (walks_per_node, walk_length, num_workers, n_epochs, n_negatives, window_size, lr, sample).
embd_dim: Integer. Embedding dimension.
directed: Boolean. Treat graph as directed.
seed: Integer. Random seed.
verbose: Boolean. Controls verbosity.

Value

A list with:

embedding - Matrix of n_nodes x embd_dim, rows in node_names order.
node_names - Node identifiers in row order.
node_types - Node type per row.
visited - Logical per row. FALSE if no surviving walk touched the node, i.e. its row is the random initialisation.
walk_stats - Named list with the walk generation statistics.