fast eigen-functions tracking on dynamic graphs
TRANSCRIPT
Arizona State University
FastEigen-FunctionsTrackingonDynamicGraphs
Chen Chen and Hanghang Tong
- 1 -
Arizona State University
Graphs are Ubiquitous!
- 2 -
Hospital NetworkCollaboration Network
Autonomous Network Transportation Network
Arizona State University
Key Graph Parameters§P1: Epidemic Threshold (Propagation network)§P2: Centrality of nodes (All networks)§P3: Clustering Coefficient (Social network)§P4: Graph Robustness (Router/Transportation)
- 3 -
Arizona State University
P1: Epidemic Threshold
§ Questions: How easy is it to spread disease?§ Intuition
§ Solution: Related to the leading eigenvalue of the adjacency matrix for ANY cascade model [ICDM 2011]
- 4 -
1 1 1 1
Arizona State University
P2: Node Centrality
§ Question: How important is a node?§ Intuition: Having more important friends
are considered influential§ Commonly used: Eigenvector CentralityThe eigenvector corresponding to the leading eigenvalue
- 5 -
Arizona State University
P3: Clustering Coefficient
§ Question: How the nodes in the graph cluster together?
§ Intuition:
§ Solution:
- 6 -
Arizona State University
P4: Graph Robustness
§ Question: How robust is a graph under external attack?
§ Intuition:
§ Solution:
- 7 -
Power Grid[wikipedia.com]
Sandy Aftermath[forbes.com]
[SDM2014]
Arizona State University
Challenge: Graphs are Dynamic!
- 8 -
Social Networks Propagation Netoworks
Router Netoworks[www.cisco.com]
Transportation Netoworks[www.mapofworld.com]
How to track key graph parameters?
Arizona State University
Eigen-Function Tracking
§ Q1. Track key graph parameters§ Q2. Estimate the error of tracking algorithms§ Q3. Analyze attribution for drastic changes
- 9 -
Arizona State University
Roadmap
§ Motivations§ Q1: Efficient tracking algorithms§ Q2: Error estimation methods§ Q3: Attribution analysis§ Conclusion
- 10 -
Arizona State University
Key Graph Parameters
§ Observations: P1-P4 are all eigen-functions
- 11 -
P1. Epidemic Threshold
P2. Eigenvector Centrality
P3. Clustering Coefficient (Triangles)
P4. Robustness Score
Arizona State University
Goal: Tracking Top Eigen-Pairs
§ Method 1.– Calculate from scratch whenever the
structure changes– Lanczos algorithm
- 12 -
Too costly for fast-changing large graphs!
Arizona State University
Key Idea
- 13 -
+ =
InitializeUpdate
Too Expensive
Arizona State University
Key Idea: Incrementally Update
§ Intuition:
§ Solution: Matrix Perturbation Theory
- 14 -
Time stamp omitted for brevity.
Arizona State University- 15 -
Details: Step 1
Challenge: two equation with four variables
Constraints Assumptions
Solution: Introduce additional constraints and assumptions
First order perturbation terms High order perturbation terms
Arizona State University
Details: Estimate
§ Discard high order term
- 16 -
Multiply on both side
,
,1
2
3
Arizona State University
Estimate (Option 1)
- 17 -
Multiply on both side
(Trip-Basic)
,
,
Time Complexity:
(Discard high order)
Lanczos:
1
2
3
4
Arizona State University
Estimate (Option 2)
§ Keep high order perturbation terms
- 18 -
Time Complexity:
(Trip-Basic)
(Trip)
Arizona State University
Evaluation
§ Data set:– Autonomous systems AS-733(https://snap.stanford.edu/data/as.html)– 100 days time spans
• (11/08/1997-02/16/1998)• (03/15/1998-06/26/1998)
– Maximum #nodes = 4,013– Maximum #edges = 14,399
- 19 -
Arizona State University
Trip-Basic vs. Trip: Effectiveness
- 20 -
Time Stamp
Arizona State University
Day15 Day30 Day45 Day60 Day75 Day900
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1Fi
rst E
igen
valu
e Er
ror R
ate
First Eigenvalue
Trip-BasicTripIterLow-RankNystrom
Ours
Effectiveness Comparison
- 21 -
Day15 Day30 Day45 Day60 Day75 Day90
Arizona State University
Effectiveness vs. Efficiency
- 22 -
0 5 10 15 20 25 30 0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
>=0.8
Speedup
Firs
t Eig
enve
ctor
Err
or R
ate
k=5
Trip-BasicTripIterLow-RankSVD deltaNystrom
Speed-up
Speed-up
Ours
Arizona State University
Roadmap
§ Motivations§ Q1: Efficient tracking algorithms§ Q2: Error estimation methods§ Q3: Attribution analysis§ Conclusion
- 23 -
Arizona State University
Q2.Error Estimation
§ Setting:
- 24 -
0 10 20 30 40 50 60 70 80 90 1000
0.1
0.2
0.3
0.4
0.5
0.6
0.7
Time Stamp
Estim
ated
Err
or (F
irst E
igen
valu
e)Error Estimate
Trip-BasicTripOption1Option2
Time Stamps
Erro
rRat
esTime Steps
True Errors
Estimated Errors
Arizona State University
Q3. Attribution Analysis
§ Precision (Edge Addition)
- 25 -
Number of Eigen-Pairs Number of Eigen-Pairs
Top
10 A
dded
Edg
es P
reci
sion
Top
10 A
dded
Edg
es P
reci
sion
First Eigenvalue Robustness Score
Arizona State University
Conclusion
§ Goal: Tracking key graph parameters§ Solutions:
– Key idea: • Fixed eigen-space, Matrix perturbation theory
– Algorithms: Trip-Basic, Trip
§ More Details:– Error Estimation– Attribution Analysis
- 26 -