Social graph analysis for credit modeling

speaker: Ouyang Ruofei event: DataScienceSG [2018-11-07 Wed 19:57:50]

credit evaluation

• personal info, demographic
• gps, mobile device info
• social graph: good credit customers are linked to other good credit
  • use phonebook and other mobile phone data (call, sms) to link customers
  • 50% of customers are linked

use Neo4J to store the social graph

• use information value to select useful features
• select features from friends and friends-of-friends in credit modelling

label propagation to classify node as good or bad

• similar to page rank

labels may change in the future

graph is partial, there are missing edges

• use node2vec and vector distance to determine if 2 nodes are friends
• use vector distance for propagation