Research notes: analyst verbs at notebook scale

This Advanced note describes how to evaluate native graph algorithms. Start with the runnable network guide; a beginner research project does not require benchmarking.

Supported workflow

forge.rank() and forge.cluster() return Arrow tables and support explicit write_property. paths, analyze, and similar provide other read-only algorithm results. Use the algorithm catalog for names, options, graph projections, and result schemas. Python and Node call the Rust engine; NetworkX and igraph are not fallback execution engines.

Measure an actual workload

Select a dataset you can reproduce and record its source, content identity, node and edge definitions, direction, and preprocessing. The proposed graphforge.datasets catalogue is not required or available as a built-in loader; use graph construction.

Measure import, project reopen, the selected algorithm, and any write-back separately. Record engine version, hardware, memory, storage, input size, and algorithm options. Distinguish first execution from reuse of prepared state. Report failures and resource use alongside duration. Earlier unattributed sub-second examples are not current performance guarantees.

Use scale guidance and the measurement method for the repository’s workload-specific evidence. Keep results on the owning issue or artifact.

Check correctness and interpretation separately

Compare an algorithm with a reference implementation only after matching projection, direction, parameters, and normalization. Public UUIDs or a chosen stable external key must identify the same nodes in both outputs.

Matching implementations support the algorithmic result. They do not validate sampling, establish that a community represents a real social group, or turn a centrality score into a quality measure. Inspect source evidence and exceptions before drawing a substantive conclusion.