Introducing networkx

Where we left off

Lesson 19 measured exactly where a hand-rolled dict-of-lists graph starts to strain: pairwise operations like normalization grow quadratically, and real graph questions (shortest path, connected components) have no efficient answer in a plain dict at all. This lesson introduces networkx, a pure-Python graph library, no server, no extra account, just pip install networkx (already pulled in as a dependency of other packages this repo uses, and made explicit in pyproject.toml as of this lesson).

What networkx actually is, given what you already built

This is worth saying plainly: networkx is not a new idea, it's a better-engineered version of exactly what Lessons 5-6 built by hand. Nodes, edges, and adjacency lookups, same concepts, same shape, implemented with real indexing structures and a large library of graph algorithms (shortest path, connected components, centrality, and more) that would each take real effort to write correctly by hand. Recognizing add_edge as "the thing I already wrote in Lesson 5" is the whole point of having built it by hand first.

import networkx as nx

G = nx.DiGraph()
G.add_edge("Mia", "humidity sensor", relation="recalibrated")

nx.DiGraph() is a directed graph, edges point one way, matching exactly the (subject, relation, object) shape this course has used since Lesson 4. add_edge takes the two endpoints directly and any extra data (here, the relation label) as a keyword argument stored on the edge, functionally identical to Lesson 5's add_edge(graph, subject, relation, obj), just backed by a proper graph data structure instead of a dict of lists.

What you get for free that you didn't have to write

list(G.successors("Mia"))          # Lesson 6's one_hop, direct neighbors
list(nx.descendants(G, "Mia"))     # every node reachable at any depth
nx.shortest_path(G, "Mia", "greenhouse")  # a real algorithm, not hand-rolled

successors is Lesson 6's one_hop, one method call. descendants is something this course's hand-rolled graph never actually implemented, every node reachable from a starting point, at any depth, computed efficiently. shortest_path answers a question this course has never been able to ask at all: not "what's within two hops," but "what's the shortest connection between these two specific nodes, however many hops that takes."