Some scenarios for a hypothetical edge_merge policy, trying to reuse standard relational algebra / sql concepts:

edge_merge policies:

• concat: fully new edges, preserving existing unaltered. add tag umap <-- default
• replace: fully drop old if any, and only put in umap neighbor edges
• inner: only label existing edges, and remove unmatched ones. When multiedges, drop excess.
• left: preserve existing edges, and enrich when there's a match. When multiedges, update just 1.
• upsert: update existing edges, and preserve umatched. Add new as needed. When multiedges, update just 1.
• skip: Preserve initial edges; only enrich nodes w/ the embedding

append new edges:

(g
  .nodes(genes_df, 'id')
  .edges(known_pathways_df, 'src', 'dst')
   .umap(edge_merge='concat', type_tag='umap', col_prefix='umap_')
   .scale(1.0) # should be responsive
).plot()

reuse edges by sticking with either just updating existing (what happens in multiedges, just pick first?) , and optionally adding missing

(g
  .nodes(entities_df, 'id')
  .edges(known_correlations_df, 'src', 'dst')
  .umap(scale=1.0, edge_merge='inner' | 'upsert', type_tag='umap', out_cols={'weight': 'score'}) 

use UMAP for 'clustering on some node columns' but preserve existing edges, or at most augmenting them

(g
  .nodes(entities_df, 'id')
  .edges(some_relns_df, 'src', 'dst')
  .umap(use_cols=['feat1', 'feat2'], use_edges=True, scale=1.0, edge_merge='skip' | 'inner')
)

use UMAP for community detection, without repositioning

(g
  .nodes(entities_df, 'id')
  .edges(some_relns_df, 'src', 'dst')
  .umap(use_edges=True, scale=1.0, edge_merge='skip', position=False)
  .dbscan(use_cols=['x', 'y'], out_cols={'label': 'community', 'weight': 'community_distance'})
)