Graph-based Feature Extraction for Online Advertising Targeting

Graph-based Feature Extraction for Online Advertising PredictionKyle NapierkowskiRadiumOne

2

Agenda• The online advertising prediction problem• The User-Domain matrix as a graph• PageRank• K-Core• Clustering with community detection

• Conclusions

3

Online Advertising Prediction

Browsing Visit Advertiser Site Buy Product

IN A NUTSHELL: Predict a user’s interest in buying products based on browsing behavior

4

The User-Domain MatrixRepresent m users’ visits to n web domains as a matrix

a.com

b.com

… n

1 0 0 … 12 0 0 … 0… … … … …m 0 1 … 1

m = 1 billion users

n = 500,000 domains How do we make sense of user interactions across sites?

How do we turn very sparse 500k domains into signals useful for modeling?

5

The User-Domain GraphCreate features from a User-Domain bipartite graph

Users

Domains

6

PageRank of DomainsCalculate PageRank using shared users between domains

7

K-Core of DomainsCategorize domains by their K-Core

8

Community detection (clustering)Project to domain-only unipartite graph using ItemSimilarityRecommender, and

detect communities using igraph

9

Combining Features

10

Conclusions• Graph-based approaches offer new opportunities

to extract information for online ad targeting

• Dato’s GraphLab makes graph analytics very easy

• GraphLab a good first step and can be extended with more specialized libraries

11

Thank you! Contact me@[email protected]

Learn moreRadiumOne.comLeanDataScience.com

Work with usradiumone.com/careers

Questions?