From Closing Triangles to Closing Higher-Order Motifs Ryan A. Rossi 1 - - PowerPoint PPT Presentation

SLIDE 1

From Closing Triangles to Closing Higher-Order Motifs

Ryan A. Rossi1 | Anup Rao1, Sungchul Kim1, Eunyee Koh1, Nesreen K. Ahmed2

1 Adobe Research 2 Intel Labs

SLIDE 2

Higher-Order Motif Closures

§ Proposed general notion of a motif closure that goes beyond simple triangle closures § Introduce higher-order ranking and link prediction methods based on closing higher-order

network motifs

§ Demonstrate that these new motif closures often outperform triangle-based methods

SLIDE 3

Higher-Order Motif Closure Frequency

SLIDE 4

Higher-Order Motif Closure Frequency

SLIDE 5

Experiments

Th The experiments s invest stigate the fo following key quest stions: s:

§ Q1

• Q1. Do other motif closures perform better than triangle closure & its variants for some graphs?

§ Q2.

• Q2. Does the “best” motif closure depend highly on the underlying network and its structural

properties or is there one motif closure that always outperforms the others? Ex Expe perimenta tal Setu tup

§ Hold-out 10% of the observed node pairs uniformly at random § Randomly sample the same number of negative node pairs § Use the methods to obtain a ranking of the node pairs § Repeat this 10 times and average the result

SLIDE 6

Results

Result 1. Higher-order motif closures can outperform triangle closure (common neighbors) and other methods based on it Mean average precision (MAP) results for ranking methods based on closing higher-order motifs.

SLIDE 7

Results

Result 2. The best motif closure depends highly on the structural characteristics of the graph and its domain (biological vs. social network) as shown in the above Table. Mean average precision (MAP) results for ranking methods based on closing higher-order motifs.

SLIDE 8

Results

§ Average runtime in

milliseconds to compute all {3, 4}-node motif closures for each node pair.

§ The runtime includes the

baselines since they require 3-node motifs.

b n

• m
• u

s e b i

• D

M

• L

C b i

• C

E

• H

T b i

• D

M

• H

T i a

• r

e a l i t y w e b

• p
• l

b l

• g

s b i

• g

r i d

• w
• r

m b i

• g

r i d

• p

l a n t b i

• g

r i d

• y

e a s t e m a i l

• d

n c

• c
• r

e c . s

• c
• a

d v

• g

a t

• e

c

• n
• w

m 1 b n

• m

a c a q u e

• r

h e . r

• a

d

• m

i n n e s

• t

a s

• c
• f

b

• m

e s s a g e s e m a i l

• E

U e m a i l

• u

n i v 0.5 1 1.5 2 2.5 3 3.5

runtime (ms.)

Result 3. For any 4-node motif H, counting the number of motif closures Wij that would arise if an edge between i and j was added to G is fast taking less than a millisecond on average across all graphs

SLIDE 9

Summary of Contributions

§ Proposed General Notion of ”Motif Closure”

§ Moves beyond simple triangle closures

§ Motif closures are often more predictive than triangle closures § Important Findings & Implications of Results

§ Need to consider other motif closures (besides triangle closures) § Best motif closure depends highly on the network structure and processes governing it § Existing supervised methods can benefit new motif closures (by leveraging full spectrum, most dense to least)

SLIDE 10

Thanks for listening!

SLIDE 11

Appendix