Skip to content

Personalized PageRank

Personalized PageRank (PPR) is the first stage of every retrieval and Odin's structural signal. It answers one question before any traversal happens: relative to my seed entities, which nodes in the graph actually matter?

Getting that answer first is what makes the rest of the pipeline affordable. Instead of exploring the whole graph, beam search only ever expands outward from the handful of nodes PPR marks as important.

Why personalized

Classic PageRank ranks every node by global importance, the graph's celebrities. That is the wrong question here. You do not care that a node is famous across the whole graph; you care that it is important to the things you are investigating.

Personalized PageRank fixes this by biasing the random walk toward your seed set. The result is a per-query importance distribution: nodes well-connected to your seeds rise to the top, and everything distant or irrelevant fades toward zero. Change the seeds and the entire ranking shifts, because importance is always measured relative to what you asked about.

How it behaves

The intuition is a crowd of walkers. Release many of them at your seed nodes; each one follows edges at random but, every so often, teleports back to a seed. After enough steps, the share of time spent at each node is its PPR score. Three things fall out of this:

  • Nodes tightly connected to your seeds accumulate high scores.
  • Bottleneck nodes that many paths funnel through also score highly, and are often the most interesting connectors in an investigation.
  • Nodes far from every seed collect almost no walker time and are effectively skipped.

Using it directly

Inside retrieve() these scores become the anchors the pipeline explores from, but you can also ask for them on their own, a fast way to orient yourself in an unfamiliar neighborhood:

anchors = engine.find_anchors(seeds=["community/insurance_claims"], topn=20)
for node_id, ppr_score in anchors:
    print(f"{ppr_score:.4f}  {node_id}")

find_anchors() returns (node_id, ppr_score) tuples sorted by importance. The two levers that shape the result are your seeds (the personalization vector that decides what "important" is measured against) and topn, how many anchors you keep. Broad seeds give a wide sense of importance; a few specific seeds produce sharp, targeted anchors. If you are working within a partitioned graph, community_id and community_mode confine the walk to that scope.

Because PPR runs over the cached graph accessor, repeated retrievals against a warm cache stay fast even though a walk touches many nodes.

Once PPR has named the anchors, the pipeline hands them to Beam Search to grow multi-hop paths. For the practical, task-oriented version of everything above, see Finding Anchors.