System-Design-Notes

[System Design] PYMK System

High-Level Design

Pointwise Learing to Rank (LTR)

Take 2 users as input, out the probability of forming a connection

not consider social context

Graph-Based Prediction

• Graph-level prediction

  predict if a chemical compound is an enzyme

• Node-level-level prediction

  predict if a specific user (node) is a spammer

• Edge-level prediction

  predict if two users are likely to connect

Edge Prediction

Features

• User: demographic, number of connections
• Interactions: connection req, accept, follow, search, view, like…
• User-User: connections, schools, major, Companies in common
• Social: profile visits, number of connections, time discounted mutual connections

GNN

e.g., GCN, GraphSAGE, GAT, GIT

Nodes store info like age, gender…

Edges store user-user info like common schools, connections…

GNN produces node embeddings for each node. Then perform similarity measures like dot product to predict if there is an edge between two users

GNN Training Labels

GNN Predicting

Scoring Service

Summary