Glossary · E-commerce ML
Graph Neural Networks
also: GNN · GraphSAGE · graph convolution
Definition
Graph Neural Networks learn representations over graph-structured data by message-passing between nodes. In e-commerce cross-sell, GNNs ingest the user-item-category graph and produce recommendations that respect product hierarchy, co-purchase relationships, and session structure — outperforming flat collaborative filtering by 15–25% on business metrics.
Standard recommendation models treat the user-item matrix as a flat bipartite graph. GNNs — GraphSAGE, GCN, PinSAGE — explicitly model multi-hop relationships and aggregate information from graph neighborhoods. The architecture's strength in cross-sell is its ability to exploit complementarity structure (products in the same cart) and category hierarchy simultaneously. Pinterest's production PinSAGE serves billions of recommendations daily using a GraphSAGE variant tuned for their pin-board-user graph.
Essays on this concept
- Business Analytics
Causal Discovery in Business Data: Applying PC Algorithm and FCI to Find Revenue Drivers Without Experiments
Correlation tells you that feature usage and retention move together. It doesn't tell you which causes which — or whether a third factor drives both. Causal discovery algorithms can untangle this from observational data alone.
- E-commerce ML
Cold-Start Problem Solved: Few-Shot Learning for New Product Recommendations Using Meta-Learning
New products get no recommendations. No recommendations means no clicks. No clicks means no data. No data means no recommendations. Meta-learning breaks this loop by transferring knowledge from products that came before.
- E-commerce ML
Graph Neural Networks for Cross-Sell: Modeling the Product Co-Purchase Network at Scale
Association rules find that beer and diapers are co-purchased. Graph neural networks understand why — the underlying structure of complementary needs, occasion-based shopping, and brand affinity networks that connect products across categories.
- E-commerce ML
Real-Time Fraud Detection at Checkout: A Streaming ML Pipeline Architecture with Sub-100ms Latency
You have 100 milliseconds to decide whether a transaction is fraudulent. In that window, you need to compute 200+ features from streaming data, run inference on a model trained on 1:1000 class imbalance, and return a score that balances revenue loss against customer friction.
- E-commerce ML
Search Ranking as a Revenue Optimization Problem: Learning-to-Rank with Business Objective Regularization
E-commerce search is not Google search. When a user types 'running shoes,' the goal isn't to find the most relevant document — it's to surface the product most likely to be purchased at the highest margin. This reframes ranking as a constrained revenue optimization problem.
- E-commerce ML
Transformer-Based Product Embeddings: Outperforming Collaborative Filtering with Multimodal Representations
Collaborative filtering needs a user to buy before it can recommend. Transformer-based embeddings understand products from their descriptions, images, and the behavioral context of browsing sessions — no purchase history required.
Related concepts
Authoritative references