Neighbor-based recommendations on a live interaction graph
JetGraph is not a full recommender platform. It is an in-memory graph engine that can supply live affinity, novelty, and interaction-velocity signals to whatever ranks the final list.
What the graph holds
Users, items, and interaction edges (view, cart, purchase, skip) with time. Optional: session, device, category nodes if those relationships matter to novelty or fraud-adjacent abuse of the recommender.
Collaborative filtering (neighbor-based)
Classic “users who interacted with X also interacted with Y” is a short traversal plus aggregation. JetGraph can serve that pattern for hot, bounded neighborhoods. Unbounded “compute embeddings over 200M users nightly” still belongs in a batch ML stack. Use JetGraph when the product needs this session’s graph, not last week’s matrix factorization alone.
Novelty signals
Has this user ever seen this item (or category)? First-time edges are useful both for exploration and for fraud (new payee, new beneficiary). Same novelty primitive as the fraud path.
Item affinity and interaction velocity
Velocity rings describe how fast a user or item is accumulating interactions — useful for trending, for throttling scraping, and for detecting inorganic bursts. HyperLogLog cardinality describes how many distinct users touched an item without a full scan.
Honest fit
If you only need Redis sorted sets for “top 10 in category,” you do not need JetGraph. If you need relationship novelty + neighbor overlap + velocity in one place at millisecond latency, you do. Compare: JetGraph vs Redis.