Skip to content
GraphForge

Research: Network Analysis Notebook Workflow

Scope: SNAP datasets, Cypher metrics, pandas/NetworkX bridge, analyst verbs.

Executive summary

Section titled “Executive summary”

Notebook network analysis works best when GraphForge owns storage and Cypher, analyst verbs own the common centrality/community algorithms, and NetworkX or igraph are used as optional visualization / oracle layers.

What holds up:

  • load_dataset(..., "snap-ego-facebook") loads :Node / :CONNECTED_TO — match those labels in Cypher (not invented :Person / :FRIEND_OF aliases)
  • Degree, triangle, and path queries in openCypher for exploratory metrics
  • forge.rank / forge.cluster with optional write_property for persistence
  • to_networkx() / to_igraph() / to_dataframe() for bridges
  • Parquet project directories for shareable notebook state

Friction to design around:

  • Exported NetworkX node IDs are internal integers unless you pass node_id_property=... — plan write-back with an explicit id property
  • Prefer analyst verbs over per-node Cypher UNWIND write-back for large score maps
  • Align every snippet’s labels with the dataset schema before debugging the engine

Label alignment

Section titled “Label alignment”
from graphforge import GraphForge
from graphforge.datasets import load_dataset


forge = GraphForge("facebook/")
load_dataset(forge, "snap-ego-facebook")


# Dataset schema
forge.execute("MATCH (n) RETURN DISTINCT labels(n) AS labels LIMIT 5")
forge.execute("MATCH ()-[r]->() RETURN DISTINCT type(r) AS t")

Recommendations

Section titled “Recommendations”
  1. Discover labels once after load_dataset; never hard-code a different schema.
  2. Use forge.rank / forge.cluster for catalog algorithms; see analyst verbs at scale.
  3. Pass node_id_property when round-tripping algorithm results through NetworkX.
  4. Persist notebooks with GraphForge("analysis/"), not ad-hoc pickle files.