FAST-PPR: Personalized PageRank Estimation for Large Graphs

Peter Lofgren (Stanford)

Joint work with Siddhartha Banerjee (Stanford),

Ashish Goel (Stanford), and C. Seshadhri (Sandia)

Motivation: Personalized Search

2

Motivation: Personalized Search

3

Re-ranked by PPR

Result Preview

4

2 sec

6 min1.2 hour

Fast-PPR Monte-Carlo

Local-Update

Personalized PageRank

5

Goal

6

Previous Algorithm: Monte-Carlo

7

Previous Algorithm: Monte-Carlo[Avrachenkov, et al 2007]

Previous Algorithm: Local Update

8

[Anderson, et al 2007]

Main Result

9

Analogy: Bidirectional Search

10

Bidirectional PageRank Algorithm

11

Reverse Work(Frontier Discovery)

Forward Work(Random Walks)

u

Main Idea

12

Experimental Setup

13

Empirical Running Time

14

Log Scale

Summary

15

Thank You

• Paper available on Arxiv

• Code available at cs.stanford.edu/~plofgren

16

17

Frontier is Important

18

FrontierAidedSignificanceThresholding

Algorithm (Simple Version)

19

Algorithm (Simple Version)

20

Average Running Time

21

Reverse Work (Local Update)

Forward Work (Monte-Carlo)

Correctness

22

Algorithm (Theoretical Version)

23

24

Algorithm (Theoretical Version)

v1

Local Update Algorithm

25

Uu Uv2

Uv3

Ut

Local Update Algorithm

26

Local Update Algorithm

27