defaultdict(list) removes the boilerplate of 'check if the key exists, create an empty list if not, then append'. For undirected graphs, every edge (u, v) is added twice (u -> v and v -> u) so both endpoints can find each other. The directed=True switch flips that off. The space cost is O(V + E), not O(V^2) like an adjacency matrix, which is why adjacency lists are the default representation for sparse graphs. Sorting the neighbor lists is purely for readable output; in production you keep insertion order so traversal is deterministic.

For weighted graphs the neighbor list holds tuples of (neighbor, weight) instead of bare neighbors. Adjacency lists naturally handle parallel edges: just append both, and the consumer (Dijkstra, Bellman-Ford, MST) can use whichever weight wins for its question. The same shape extends to multi-attribute edges ((neighbor, weight, capacity, edge_id)) by widening the tuple. If you ever find yourself fighting the data structure, switch to a dict-of-dicts (graph[u][v] = weight) for O(1) lookup of a specific edge, at the cost of losing parallel edges.

Recursive DFS is the most natural traversal once you have an adjacency list: visit the node, mark it seen, recurse into every unseen neighbor. The visited set is mandatory; without it the function loops forever on any cycle. To handle disconnected graphs, restart DFS from every unvisited node and record the components separately. Recursion is fine up to a few thousand nodes; for deep graphs use the iterative-stack version (covered in the DFS-templates entry) to avoid RecursionError.