nonparametric link prediction in dynamic graphs
DESCRIPTION
Nonparametric Link Prediction in Dynamic Graphs. Purnamrita Sarkar (UC Berkeley) Deepayan Chakrabarti (Facebook) Michael Jordan (UC Berkeley). Link Prediction. Who is most likely to be interact with a given node?. Should Facebook suggest Alice as a friend for Bob ?. Alice. Bob. - PowerPoint PPT PresentationTRANSCRIPT
![Page 1: Nonparametric Link Prediction in Dynamic Graphs](https://reader035.vdocuments.site/reader035/viewer/2022062521/5681667a550346895dda1a83/html5/thumbnails/1.jpg)
1
Nonparametric Link Prediction in Dynamic Graphs
Purnamrita Sarkar (UC Berkeley)Deepayan Chakrabarti (Facebook)Michael Jordan (UC Berkeley)
![Page 2: Nonparametric Link Prediction in Dynamic Graphs](https://reader035.vdocuments.site/reader035/viewer/2022062521/5681667a550346895dda1a83/html5/thumbnails/2.jpg)
2
Link Prediction Who is most likely to be interact with a given node?
Friend suggestion in Facebook
Should Facebook suggest Alice
as a friend for Bob?
Bob
Alice
![Page 3: Nonparametric Link Prediction in Dynamic Graphs](https://reader035.vdocuments.site/reader035/viewer/2022062521/5681667a550346895dda1a83/html5/thumbnails/3.jpg)
3
Link Prediction
Alice
Bob
Charlie
Movie recommendation in Netflix
Should Netflix suggest this
movie to Alice?
![Page 4: Nonparametric Link Prediction in Dynamic Graphs](https://reader035.vdocuments.site/reader035/viewer/2022062521/5681667a550346895dda1a83/html5/thumbnails/4.jpg)
4
Link Prediction Prediction using simple features
degree of a node number of common neighbors last time a link appeared
What if the graph is dynamic?
![Page 5: Nonparametric Link Prediction in Dynamic Graphs](https://reader035.vdocuments.site/reader035/viewer/2022062521/5681667a550346895dda1a83/html5/thumbnails/5.jpg)
5
Related Work
Generative models Exp. family random graph models [Hanneke+/’06] Dynamics in latent space [Sarkar+/’05] Extension of mixed membership block models
[Fu+/10] Other approaches
Autoregressive models for links [Huang+/09] Extensions of static features [Tylenda+/09]
![Page 6: Nonparametric Link Prediction in Dynamic Graphs](https://reader035.vdocuments.site/reader035/viewer/2022062521/5681667a550346895dda1a83/html5/thumbnails/6.jpg)
6
Goal
Link Prediction incorporating graph dynamics, requiring weak modeling assumptions, allowing fast predictions, and offering consistency guarantees.
![Page 7: Nonparametric Link Prediction in Dynamic Graphs](https://reader035.vdocuments.site/reader035/viewer/2022062521/5681667a550346895dda1a83/html5/thumbnails/7.jpg)
7
Outline
Model Estimator Consistency Scalability Experiments
![Page 8: Nonparametric Link Prediction in Dynamic Graphs](https://reader035.vdocuments.site/reader035/viewer/2022062521/5681667a550346895dda1a83/html5/thumbnails/8.jpg)
8
The Link Prediction Problem in Dynamic Graphs
G1 G2 GT+1……
Y1 (i,j)=1
Y2 (i,j)=0
YT+1 (i,j)=?
YT+1(i,j) | G1,G2, …,GT ~ Bernoulli (gG1,G2,…GT(i,j))
Edge in T+1 Features of previous graphsand this pair of nodes
![Page 9: Nonparametric Link Prediction in Dynamic Graphs](https://reader035.vdocuments.site/reader035/viewer/2022062521/5681667a550346895dda1a83/html5/thumbnails/9.jpg)
9
cn
ℓℓ
deg
Including graph-based features
Example set of features for pair (i,j): cn(i,j) (common neighbors) ℓℓ(i,j) (last time a link was formed) deg(j)
Represent dynamics using “datacubes” of these features. ≈ multi-dimensional histogram on binned feature values
ηt = #pairs in Gt with these features
1 ≤ cn ≤ 33 ≤ deg ≤ 61 ≤ ℓℓ ≤ 2
ηt+ = #pairs in Gt with these
features, which had an edge in Gt+1
high ηt+/ηt this feature
combination is more likely to create a new edge at time t+1
![Page 10: Nonparametric Link Prediction in Dynamic Graphs](https://reader035.vdocuments.site/reader035/viewer/2022062521/5681667a550346895dda1a83/html5/thumbnails/10.jpg)
10
G1 G2 GT……
Y1 (i,j)=1 Y2 (i,j)=0 YT+1 (i,j)=?
1 ≤ cn(i,j) ≤ 33 ≤ deg(i,j) ≤ 61 ≤ ℓℓ (i,j) ≤ 2
Including graph-based features
How do we form these datacubes? Vanilla idea: One datacube for Gt→Gt+1
aggregated over all pairs (i,j) Does not allow for differently evolving communities
![Page 11: Nonparametric Link Prediction in Dynamic Graphs](https://reader035.vdocuments.site/reader035/viewer/2022062521/5681667a550346895dda1a83/html5/thumbnails/11.jpg)
11
YT+1 (i,j)=?
1 ≤ cn(i,j) ≤ 33 ≤ deg(i,j) ≤ 61 ≤ ℓℓ (i,j) ≤ 2
Our Model
How do we form these datacubes? Our Model: One datacube for each neighborhood
Captures local evolution
G1 G2 GT……
Y1 (i,j)=1 Y2 (i,j)=0
![Page 12: Nonparametric Link Prediction in Dynamic Graphs](https://reader035.vdocuments.site/reader035/viewer/2022062521/5681667a550346895dda1a83/html5/thumbnails/12.jpg)
12
Our Model
Number of node pairs- with feature s- in the neighborhood of i- at time t
Number of node pairs- with feature s- in the neighborhood of i- at time t- which got connected at time t+1
Datacube
1 ≤ cn(i,j) ≤ 33 ≤ deg(i,j) ≤ 61 ≤ ℓℓ (i,j) ≤ 2
Neighborhood Nt(i)= nodes within 2 hops
Features extracted from (Nt-p,…Nt)
![Page 13: Nonparametric Link Prediction in Dynamic Graphs](https://reader035.vdocuments.site/reader035/viewer/2022062521/5681667a550346895dda1a83/html5/thumbnails/13.jpg)
13
Our Model
Datacube dt(i) captures graph evolution in the local neighborhood of a node in the recent past
Model:
What is g(.)?
YT+1(i,j) | G1,G2, …,GT ~ Bernoulli ( gG1,G2,…GT(i,j))g(dt(i), st(i,j))
Features of the pair
Local evolution patterns
![Page 14: Nonparametric Link Prediction in Dynamic Graphs](https://reader035.vdocuments.site/reader035/viewer/2022062521/5681667a550346895dda1a83/html5/thumbnails/14.jpg)
14
Outline
Model Estimator Consistency Scalability Experiments
![Page 15: Nonparametric Link Prediction in Dynamic Graphs](https://reader035.vdocuments.site/reader035/viewer/2022062521/5681667a550346895dda1a83/html5/thumbnails/15.jpg)
15
Kernel Estimator for g
G1 G2 …… GTGT-1GT-2
query data-cube at T-1 and feature vector at time T
compute similarities
datacube, feature pair
t=1
{{
{
{
{
{
{
{
…
datacube, feature pair
t=2
{{
{
{
{
{
{
{
…datacube,
feature pair t=3
{{
{
{
{
{
{
{
…{
{
![Page 16: Nonparametric Link Prediction in Dynamic Graphs](https://reader035.vdocuments.site/reader035/viewer/2022062521/5681667a550346895dda1a83/html5/thumbnails/16.jpg)
16
Factorize the similarity function Allows computation of g(.) via simple lookups
}} }
K( , )I{ == }
Kernel Estimator for g
![Page 17: Nonparametric Link Prediction in Dynamic Graphs](https://reader035.vdocuments.site/reader035/viewer/2022062521/5681667a550346895dda1a83/html5/thumbnails/17.jpg)
17
Kernel Estimator for g
G1 G2 …… GTGT-1GT-2
datacubes t=1
datacubes t=2
datacubes t=3
compute similarities only between data cubes
w1
w2
w3
w4
η1 , η1+
η2 , η2+
η3 , η3+
η4 , η4+
44332211
44332211
wwwwwwww
![Page 18: Nonparametric Link Prediction in Dynamic Graphs](https://reader035.vdocuments.site/reader035/viewer/2022062521/5681667a550346895dda1a83/html5/thumbnails/18.jpg)
18
Factorize the similarity function Allows computation of g(.) via simple lookups What is K( , )?
}}
}
K( , )I{ == }
Kernel Estimator for g
![Page 19: Nonparametric Link Prediction in Dynamic Graphs](https://reader035.vdocuments.site/reader035/viewer/2022062521/5681667a550346895dda1a83/html5/thumbnails/19.jpg)
19
Similarity between two datacubes
Idea 1 For each cell s, take
(η1+/η1 – η2
+/η2)2 and sum
Problem: Magnitude of η is ignored 5/10 and 50/100 are treated
equally
Consider the distribution
η1 , η1+
η2 , η2+
![Page 20: Nonparametric Link Prediction in Dynamic Graphs](https://reader035.vdocuments.site/reader035/viewer/2022062521/5681667a550346895dda1a83/html5/thumbnails/20.jpg)
20
Similarity between two datacubes
0 5 10 15 20 25 30 35 40 450
0.01
0.02
0.03
0.04
0.05
0.06
0.07
0 5 10 15 20 25 30 35 40 450
0.02
0.04
0.06
0.08
0.1
0.12
0.14
) , dist(b) , K( 0<b<1
As b0, K( , ) 0 unless dist( , ) =0
Idea 2 For each cell s, compute
posterior distribution of edge creation prob.
dist = total variation distance between distributions summed over all cells
η1 , η1+
η2 , η2+
![Page 21: Nonparametric Link Prediction in Dynamic Graphs](https://reader035.vdocuments.site/reader035/viewer/2022062521/5681667a550346895dda1a83/html5/thumbnails/21.jpg)
21
1tη) , K(#1f
) , (f) , (h) , (g
1tη) , K(
#1h
Want to show: gg
Kernel Estimator for g
![Page 22: Nonparametric Link Prediction in Dynamic Graphs](https://reader035.vdocuments.site/reader035/viewer/2022062521/5681667a550346895dda1a83/html5/thumbnails/22.jpg)
22
Outline
Model Estimator Consistency Scalability Experiments
![Page 23: Nonparametric Link Prediction in Dynamic Graphs](https://reader035.vdocuments.site/reader035/viewer/2022062521/5681667a550346895dda1a83/html5/thumbnails/23.jpg)
23
Consistency of Estimator
Lemma 1: As T→∞, for some R>0,
Proof using:
) , (f) , (h) , (g
As T→∞,
![Page 24: Nonparametric Link Prediction in Dynamic Graphs](https://reader035.vdocuments.site/reader035/viewer/2022062521/5681667a550346895dda1a83/html5/thumbnails/24.jpg)
24
Consistency of Estimator
Lemma 2: As T→∞,
) , (f) , (h) , (g
![Page 25: Nonparametric Link Prediction in Dynamic Graphs](https://reader035.vdocuments.site/reader035/viewer/2022062521/5681667a550346895dda1a83/html5/thumbnails/25.jpg)
25
Consistency of Estimator
Assumption: finite graph Proof sketch:
Dynamics are Markovian with finite state spacethe chain must eventually enter a closed, irreducible communication classgeometric ergodicity if class is aperiodic(if not, more complicated…)strong mixing with exponential decayvariances decay as o(1/T)
![Page 26: Nonparametric Link Prediction in Dynamic Graphs](https://reader035.vdocuments.site/reader035/viewer/2022062521/5681667a550346895dda1a83/html5/thumbnails/26.jpg)
26
Consistency of Estimator
Theorem:
Proof Sketch:
for some R>0
So
![Page 27: Nonparametric Link Prediction in Dynamic Graphs](https://reader035.vdocuments.site/reader035/viewer/2022062521/5681667a550346895dda1a83/html5/thumbnails/27.jpg)
27
Outline
Model Estimator Consistency Scalability Experiments
![Page 28: Nonparametric Link Prediction in Dynamic Graphs](https://reader035.vdocuments.site/reader035/viewer/2022062521/5681667a550346895dda1a83/html5/thumbnails/28.jpg)
28
Scalability Full solution:
Summing over all n datacubes for all T timesteps Infeasible
Approximate solution: Sum over nearest neighbors of query datacube
How do we find nearest neighbors? Locality Sensitive Hashing (LSH)
[Indyk+/98, Broder+/98]
![Page 29: Nonparametric Link Prediction in Dynamic Graphs](https://reader035.vdocuments.site/reader035/viewer/2022062521/5681667a550346895dda1a83/html5/thumbnails/29.jpg)
29
Using LSH
Devise a hashing function for datacubes such that “Similar” datacubes tend to be hashed to the
same bucket “Similar” = small total variation distance
between cells of datacubes
![Page 30: Nonparametric Link Prediction in Dynamic Graphs](https://reader035.vdocuments.site/reader035/viewer/2022062521/5681667a550346895dda1a83/html5/thumbnails/30.jpg)
30
0 5 10 15 20 25 30 35 40 450
0.01
0.02
0.03
0.04
0.05
0.06
0.07
Using LSH
Step 1: Map datacubes to bit vectors
Use B2 bits for each bucket For probability mass p the first bits are set to
1Use B1 buckets to discretize [0,1]
Total M*B1*B2 bits, where M = max number of occupied cells << total number of cells
![Page 31: Nonparametric Link Prediction in Dynamic Graphs](https://reader035.vdocuments.site/reader035/viewer/2022062521/5681667a550346895dda1a83/html5/thumbnails/31.jpg)
31
Using LSH
Step 1: Map datacubes to bit vectors Total variation distance
L1 distance between distributions Hamming distance between vectors
Step 2: Hash function = k out of MB1B2 bits
![Page 32: Nonparametric Link Prediction in Dynamic Graphs](https://reader035.vdocuments.site/reader035/viewer/2022062521/5681667a550346895dda1a83/html5/thumbnails/32.jpg)
32
Fast Search Using LSH
1111111111000000000111111111000
10000101000011100001101010000
10101010000011100001101010000
101010101110111111011010111110
1111111111000000000111111111001
00000001
1111
0011
.
.
.
.
1011
![Page 33: Nonparametric Link Prediction in Dynamic Graphs](https://reader035.vdocuments.site/reader035/viewer/2022062521/5681667a550346895dda1a83/html5/thumbnails/33.jpg)
33
Outline
Model Estimator Consistency Scalability Experiments
![Page 34: Nonparametric Link Prediction in Dynamic Graphs](https://reader035.vdocuments.site/reader035/viewer/2022062521/5681667a550346895dda1a83/html5/thumbnails/34.jpg)
34
Experiments
Baselines LL: last link (time of last occurrence of a pair)
CN: rank by number of common neighbors in AA: more weight to low-degree common neighbors Katz: accounts for longer paths
CN-all: apply CN to AA-all, Katz-all: similar
ss
![Page 35: Nonparametric Link Prediction in Dynamic Graphs](https://reader035.vdocuments.site/reader035/viewer/2022062521/5681667a550346895dda1a83/html5/thumbnails/35.jpg)
35
Setup
Pick random subset S from nodes with degree>0 in GT+1
, predict a ranked list of nodes likely to link to s Report mean AUC (higher is better)
G1 G2 GT
Training data Test dataGT+
1
![Page 36: Nonparametric Link Prediction in Dynamic Graphs](https://reader035.vdocuments.site/reader035/viewer/2022062521/5681667a550346895dda1a83/html5/thumbnails/36.jpg)
36
Simulations Social network model of Hoff et al.
Each node has an independently drawn feature vector
Edge(i,j) depends on features of i and j Seasonality effect
Feature importance varies with seasondifferent communities in each season
Feature vectors evolve smoothly over timeevolving community structures
![Page 37: Nonparametric Link Prediction in Dynamic Graphs](https://reader035.vdocuments.site/reader035/viewer/2022062521/5681667a550346895dda1a83/html5/thumbnails/37.jpg)
37
Simulations
NonParam is much better than others in the presence of seasonality
CN, AA, and Katz implicitly assume smooth evolution
![Page 38: Nonparametric Link Prediction in Dynamic Graphs](https://reader035.vdocuments.site/reader035/viewer/2022062521/5681667a550346895dda1a83/html5/thumbnails/38.jpg)
38
Sensor Network*
* www.select.cs.cmu.edu/data
![Page 39: Nonparametric Link Prediction in Dynamic Graphs](https://reader035.vdocuments.site/reader035/viewer/2022062521/5681667a550346895dda1a83/html5/thumbnails/39.jpg)
39
Summary
Link formation is assumed to depend on the neighborhood’s evolution over a time window
Admits a kernel-based estimator Consistency Scalability via LSH
Works particularly well for Seasonal effects differently evolving communities