Design People You May Know: Graph-Based Friend Recommendations

The question: design "People You May Know" (PYMK) for a social or professional network: suggest people a user is likely to know and connect with. It's a recommendation problem on a graph, where the strongest signals are mutual connections, shared schools and workplaces, and contacts, and where the candidate space (every member) is too large to score directly.

1. Clarify requirements

2. Frame it as an ML problem

Link prediction: for a (viewer, candidate) pair, predict p(connection request sent and accepted) within some window. Optimising only sends would reward spammy suggestions; acceptance captures real relationships.

3. Candidate generation

Scoring a billion people per user is impossible, so generate a few thousand candidates from several sources:

• Friends of friends (FoF): with 300 connections each, a user has up to ~90,000 two-hop neighbours. Keep those with the most mutual connections (cap per hop to avoid celebrities exploding the walk).
• Shared affiliations: same school and graduation years, same employer and team, same city.
• Contacts: people whose phone number or email is in the user's uploaded address book (and vice versa), with consent.
• Graph embeddings: node embeddings (node2vec-style random walks, or a graph neural network) placed in an ANN index to find people close in the graph who aren't two hops away.

4. Features and model

A gradient-boosted tree or deep ranker predicts p(accept). Mutual connections are typically the strongest feature, but weighting them (a mutual who has 5,000 connections means less than one who has 50) improves results. See embeddings and choosing a model.

5. Batch plus real-time serving

Daily batch : 1B users × ~3,000 candidates = 3T pairs to score → too many;
              prune FoF by mutual count first (keep ~500), score 500B pairs on a big
              Spark/GPU cluster, store top 200 per user: 1B × 200 × 8 B ≈ 1.6 TB
Online      : read top 200, drop already connected/pending/dismissed/blocked,
              light rerank with fresh signals, return 20 in < 50 ms

6. Product and privacy rules

• Never suggest people the viewer blocked or who blocked them; respect "don't suggest me" settings.
• Be careful with sensitive inference: suggesting a therapist's other patients, or revealing that two people share a contact, is a real harm. Exclude sensitive contact categories and some signals entirely.
• Remove a suggestion after dismissal and don't bring it back for a long time.
• Limit how many times one person is suggested to others per day, so popular users aren't flooded with requests.

7. Evaluation

• Offline: recall@k of connections formed in the following week (time-based split); AUC of p(accept).
• Online: accepted connections per user, acceptance rate of sent requests, new-user connections in the first week, plus guardrails: request ignore and report rates.
• Watch the graph effect: a test arm changes who connects to whom, which affects control users too. Cluster-based randomisation (by community) reduces this leakage. See A/B testing.