SYSTEM Cited by 1 source
GraphSage¶
Definition¶
GraphSage is an inductive graph-embedding method (Hamilton, Ying, Leskovec, Inductive Representation Learning on Large Graphs, arXiv:1706.02216, 2017). Unlike transductive methods (DeepWalk, node2vec) that require the full graph at training, GraphSage generates embeddings for previously-unseen nodes by sampling + aggregating features from a node's neighborhood — a property load-bearing for production recommender systems where new items arrive continuously.
Role in Pinterest's Home Feed Blender¶
GraphSage is Pinterest's primary mechanism for defining Pin-to-Pin similarity in the diversification layer (Source: sources/2026-04-07-pinterest-evolution-of-multi-objective-optimization-at-pinterest-home):
- DPP era (2021) — GraphSage + categorical taxonomy were the sole similarity signals.
- SSD era (Early 2025) — GraphSage remains in the SSD signal mix alongside visual + text embeddings, capturing "relatedness in the Pin graph, including co-engagement patterns and neighborhood similarity."
GraphSage at Pinterest predates this post (Pinterest published its foundational PinSage paper at KDD 2018, extending GraphSage with random-walk-based sampling for web-scale recommenders — not ingested here).
Caveats¶
- Stub — this page covers GraphSage's role as a Pin-similarity signal in the 2026-04-07 MOO post; the academic paper's full method + the PinSage Pinterest-specific extension are not summarised here.
- Canonical Pinterest variant (PinSage) is referenced but not a separate wiki page at this time.
- Exact GraphSage version used in Pinterest Home Feed Blender is not specified in the post.
- Embedding dimension, training cadence, refresh latency not disclosed.
Seen in¶
- 2026-04-07 Pinterest — Evolution of Multi-Objective Optimization at Pinterest Home Feed (sources/2026-04-07-pinterest-evolution-of-multi-objective-optimization-at-pinterest-home) — canonical; GraphSage as Pin-similarity signal in both DPP and SSD diversification.
Related¶
- systems/pinterest-home-feed-blender
- concepts/vector-embedding
- patterns/multi-signal-pairwise-similarity — graph + visual + text + Semantic-ID similarity composition