Signal Processing on Graphs: Recent Results, Challenges

and Applications 1

Antonio Ortega

Signal and Image Processing Institute

Department of Electrical Engineering
University of Southern California

Sept. 2013

1
Supported in part by NSF (CCF-1018977) and NASA (AIST-05-0081).
A. Ortega (USC) Signal Processing on Graphs Sept. 2013 1 / 81
Acknowledgements

Collaborators
- Dr. Sunil Narang (Microsoft)
- Dr. Godwin Shen (Northrop-Grumman)
- Eduardo Martnez Enrquez (Univ. Carlos III, Madrid)
- Akshay Gadde, Jessie Chao, Yongzhe Wang (USC)
- Prof. Marco Levorato (UCI), Prof. Urbashi Mitra (USC)
- Prof. Fernando Daz de Mara (Univ. Carlos III, Madrid)
- Prof. Gene Cheung (NII)
- Prof. Pierre Vandergheynst (EPFL), Prof. Pascal Frossard (EPFL),
Dr. David Shuman (EPFL).
Funding
- NASA AIST-05-0081
- NSF CCF-1018977

A. Ortega (USC) Signal Processing on Graphs Sept. 2013 2 / 81

 Introduction

Next Section

1 Introduction

2 Wavelet Transforms on Arbitrary Graphs

3 Applications

4 Conclusions

A. Ortega (USC) Signal Processing on Graphs Sept. 2013 3 / 81

 Introduction

Motivation
Graphs provide a flexible model to represent many datasets:
Examples in Euclidean domains

1.5
1
(a) (b) (c)

(a) Computer graphics2 (b) Wireless sensor networks 3

(c) image - graphs

2
From [Sweldens, 1999]
3
From http://www.purelink.ca
A. Ortega (USC) Signal Processing on Graphs Sept. 2013 4 / 81
 Introduction

Motivation

Examples in non-Euclidean settings

Combined ARQ-Queue

Queue
2

0 1 2 3

ARQ

(a) (b)

(a) Social Networks 4 , (b) Finite State Machines(FSM)

Graphs can capture complex relational characteristics (e.g., spatial, topological).

4
Zacharay Karate Club [Zacahary, 1977]
A. Ortega (USC) Signal Processing on Graphs Sept. 2013 5 / 81
 Introduction

Graph Signal Processing?

Assume fixed graph structure: different graph signals on a given

graph

Define linear transforms for graph signals

Use these for compression, denoising, interpolation, etc

A. Ortega (USC) Signal Processing on Graphs Sept. 2013 6 / 81

 Introduction

What do we know about transforms for graph signals?

A. Ortega (USC) Signal Processing on Graphs Sept. 2013 7 / 81

 Introduction

What do we know about transforms for graph signals?

More than you think

A. Ortega (USC) Signal Processing on Graphs Sept. 2013 7 / 81

 Introduction

What do we know about transforms for graph signals?

More than you think

2 1 0 0 0 0 0 1
1 2 1 0 0 0 0 0
0 1 2 1 0 0 0 0
0 0 1 2 1 0 0 0
H=
0 0 0 1 2 1 0 0


0 0 0 0 1 2 1 0
0 0 0 0 0 1 2 1
1 0 0 0 0 0 1 2

A. Ortega (USC) Signal Processing on Graphs Sept. 2013 7 / 81

 Introduction

What do we know about transforms for graph signals?

More than you think

2 1 0 0 0 0 0 1
1 2 1 0 0 0 0 0
0 1 2 1 0 0 0 0
0 0 1 2 1 0 0 0
H=
0 0 0 1 2 1 0 0


0 0 0 0 1 2 1 0
0 0 0 0 0 1 2 1
1 0 0 0 0 0 1 2

Interpretation
Circulant matrix Circular convolution
Eigenvectors: DFT
High pass filter: each row adds to 0

Where is the graph?

A. Ortega (USC) Signal Processing on Graphs Sept. 2013 7 / 81

 Introduction

What do we know about transformations on Graphs?

Alternative representation
2 1 0 0 0 0 0 1
1 2 1 0 0 0 0 0
0 1 2 1 0 0 0 0
0 0 1 2 1 0 0 0
H=
0 0 0 1 2 1 0 0


0 0 0 0 1 2 1 0
0 0 0 0 0 1 2 1
1 0 0 0 0 0 1 2

A. Ortega (USC) Signal Processing on Graphs Sept. 2013 8 / 81

 Introduction

What do we know about transformations on Graphs?

Alternative representation
2 1 0 0 0 0 0 1
1 2 1 0 0 0 0 0
0 1 2 1 0 0 0 0
0 0 1 2 1 0 0 0
H=
0 0 0 1 2 1 0 0


0 0 0 0 1 2 1 0
0 0 0 0 0 1 2 1
1 0 0 0 0 0 1 2

2 0 0 0 0 0 0 0 0 1 0 0 0 0 0 1
0 2 0 0 0 0 0 0 1 0 1 0 0 0 0 0
0 0 2 0 0 0 0 0 0 1 0 1 0 0 0 0
0 0 0 2 0 0 0 0 0 0 1 0 1 0 0 0
H=
0 0 0 0 2 0 0 0 0 0 0 1 0 1 0 0


0 0 0 0 0 2 0 0 0 0 0 0 1 0 1 0
0 0 0 0 0 0 2 0 0 0 0 0 0 1 0 1
0 0 0 0 0 0 0 2 1 0 0 0 0 0 1 0

A. Ortega (USC) Signal Processing on Graphs Sept. 2013 8 / 81

 Introduction

What do we know about transformations on Graphs?

Alternative representation
2 1 0 0 0 0 0 1
1 2 1 0 0 0 0 0
0 1 2 1 0 0 0 0
0 0 1 2 1 0 0 0
H=
0 0 0 1 2 1 0 0


0 0 0 0 1 2 1 0
0 0 0 0 0 1 2 1
1 0 0 0 0 0 1 2

2 0 0 0 0 0 0 0 0 1 0 0 0 0 0 1
0 2 0 0 0 0 0 0 1 0 1 0 0 0 0 0
0 0 2 0 0 0 0 0 0 1 0 1 0 0 0 0
0 0 0 2 0 0 0 0 0 0 1 0 1 0 0 0
H=
0 0 0 0 2 0 0 0 0 0 0 1 0 1 0 0


0 0 0 0 0 2 0 0 0 0 0 0 1 0 1 0
0 0 0 0 0 0 2 0 0 0 0 0 0 1 0 1
0 0 0 0 0 0 0 2 1 0 0 0 0 0 1 0

H=DA
Interpretation?
A. Ortega (USC) Signal Processing on Graphs Sept. 2013 8 / 81
 Introduction

What do we know about transformations on Graphs?

2 0 0 0 0 0 0 0 0 1 0 0 0 0 0 1
0 2 0 0 0 0 0 0 1 0 1 0 0 0 0 0
0 0 2 0 0 0 0 0 0 1 0 1 0 0 0 0
0 0 0 2 0 0 0 0 0 0 1 0 1 0 0 0
H=
0 0 0 0 2 0 0 0 0 0 0 1 0 1 0 0


0 0 0 0 0 2 0 0 0 0 0 0 1 0 1 0
0 0 0 0 0 0 2 0 0 0 0 0 0 1 0 1
0 0 0 0 0 0 0 2 1 0 0 0 0 0 1 0

A. Ortega (USC) Signal Processing on Graphs Sept. 2013 9 / 81

 Introduction

What do we know about transformations on Graphs?

2 0 0 0 0 0 0 0 0 1 0 0 0 0 0 1
0 2 0 0 0 0 0 0 1 0 1 0 0 0 0 0
0 0 2 0 0 0 0 0 0 1 0 1 0 0 0 0
0 0 0 2 0 0 0 0 0 0 1 0 1 0 0 0
H=
0 0 0 0 2 0 0 0 0 0 0 1 0 1 0 0


0 0 0 0 0 2 0 0 0 0 0 0 1 0 1 0
0 0 0 0 0 0 2 0 0 0 0 0 0 1 0 1
0 0 0 0 0 0 0 2 1 0 0 0 0 0 1 0

H=DA
A and D: adjacency and degree matrices
H = L: graph Laplacian
H can be interpreted as a local operation on this graph
A. Ortega (USC) Signal Processing on Graphs Sept. 2013 9 / 81
 Introduction

Graphs

A. Ortega (USC) Signal Processing on Graphs Sept. 2013 10 / 81

 Introduction

Graphs

0 1 0 0 0 0 0 1
1 0 1 0 0 0 0 0
0 1 0 1 0 0 0 0
0 0 1 0 1 0 0 0
0 0 0 1 0 1 0 0
0 0 0 0 1 0 1 0
0 0 0 0 0 1 0 1
1 0 0 0 0 0 1 0

H is a simple polynomial of L = D A on the cycle graph

Can we do similar things on more complex graphs?

A. Ortega (USC) Signal Processing on Graphs Sept. 2013 10 / 81

 Introduction

Graphs

H is a simple polynomial of L = D A on the cycle graph

Can we do similar things on more complex graphs?

A. Ortega (USC) Signal Processing on Graphs Sept. 2013 11 / 81

 Introduction

Graphs

H is a simple polynomial of L = D A on the cycle graph

Can we do similar things on more complex graphs?
Yes! But things get more complicated
A. Ortega (USC) Signal Processing on Graphs Sept. 2013 11 / 81
 Introduction

Graphs

0 1 0 0 0 0 0 1
1 0 1 0 1 0 0 0
0 1 0 1 0 0 0 0
0 0 1 0 1 0 0 0
0 1 0 1 0 1 0 0
0 0 0 0 1 0 1 0
0 0 0 0 0 1 0 1
1 0 0 0 0 0 1 0

A is no longer circulant no DFT in general, but...

Polynomials of L = D A or A are local operators
There will be a frequency interpretation
A. Ortega (USC) Signal Processing on Graphs Sept. 2013 12 / 81
 Introduction

What makes these graph transforms?

A. Ortega (USC) Signal Processing on Graphs Sept. 2013 13 / 81

 Introduction

What makes these graph transforms?

Shift invariance: same filter at every sample

A. Ortega (USC) Signal Processing on Graphs Sept. 2013 13 / 81

 Introduction

What makes these graph transforms?

Shift invariance: same filter at every sample

Graph-based shift invariance Operator is the same, local variations
captured by A or L.
H=L=DA

A. Ortega (USC) Signal Processing on Graphs Sept. 2013 13 / 81

 Introduction

What makes these graph transforms?

Shift invariance: same filter at every sample

Graph-based shift invariance Operator is the same, local variations
captured by A or L.
H=L=DA
This can be generalized:
L1
X L1
X
k
H= k L or H = k Ak
k=0 k=0

Or alternatively, because based on Graph Fourier Transform

A. Ortega (USC) Signal Processing on Graphs Sept. 2013 13 / 81
 Introduction

Summary

Localized linear operations on graphs using polynomials of A or L.

Frequency interpretation is possible for eigenvectors of A or L.
A great deal depends on the topology of the graph

In what follows we consider mostly undirected graphs without self

loops and use L.
[Shuman, Narang, Frossard, Ortega, Vandergheysnt, SPM2013]
Other approaches are possible based on A
[Sandryhaila and Moura 2013]

A. Ortega (USC) Signal Processing on Graphs Sept. 2013 14 / 81

 Introduction

Research Goals
Extend signal processing methods to arbitrary graphs
Downsampling, graph-frequency localization, multiresolution, wavelets,
interpolation
Outcomes
Work with massive graph-datasets: localized frequency analysis
Novel insights about traditional applications (image/video processing)
New applications
This talk
Graph Signal Processing
Graph Filterbank design
Applications
Edge Aware Image Filtering
Depth image coding
Wireless network optimization
Recommendation System Example

A. Ortega (USC) Signal Processing on Graphs Sept. 2013 15 / 81

 Introduction Basic Theory

Graphs 101

Graph G = (V, E , w ).
2
3 Adjacency matrix A
4
6 16 Degree matrix D = diag {di }
15
1

5
8 14 19
17
Laplacian matrix L = D A.
7
10 13 18

9
Normalized Laplacian matrix
L = D1/2 LD1/2
12
11

21 22 23
20

24 25 26
Graph Signal
27 f = {f (1), f (2), ..., f (N)}

Assumptions:
1. Undirected graphs without self loops.
2. Scalar sample values

A. Ortega (USC) Signal Processing on Graphs Sept. 2013 16 / 81

 Introduction Basic Theory

Spectrum of Graphs

Graph Laplacian Matrix L = D A = UU0

Eigen-vectors of L : U = {uk }k=1:N

Eigen-values of L : diag {} = 1 2 ... N

Eigen-pair system {(k , uk )} provides Fourier-like interpretation

Graph Fourier Transform (GFT)

A. Ortega (USC) Signal Processing on Graphs Sept. 2013 17 / 81

 Introduction Basic Theory

Graph Frequencies

(a) = /4 0 (b) = /4 1 (c) = /4 4 (d) = /4 7

DCT basis for regular signals

(a) = 0.00 (b) = 0.04 (c) = 1.20 (d) = 1.55

Eigenvectors of an arbitrary graph

A. Ortega (USC) Signal Processing on Graphs Sept. 2013 18 / 81

 Introduction Basic Theory

Eigenvectors of graph Laplacian

(a) = 0.00 (b) = 0.04 (c) = 0.20

(d) = 0.40 (e) = 1.20 (f) = 1.49

A. Ortega (USC) Signal Processing on Graphs Sept. 2013 19 / 81

 Introduction Basic Theory

Graph Transforms

Processing/
Input Signal Transform Output Signal
Analysis

Desirable properties
Invertible
Critically sampled
Orthogonal
Localized in graph (space) and graph spectrum (frequency)
Local Linear Transform
Can we define Graph Wavelets?

A. Ortega (USC) Signal Processing on Graphs Sept. 2013 20 / 81

 Wavelet Transforms on Arbitrary Graphs

Next Section

1 Introduction

2 Wavelet Transforms on Arbitrary Graphs

3 Applications

4 Conclusions

A. Ortega (USC) Signal Processing on Graphs Sept. 2013 21 / 81

 Wavelet Transforms on Arbitrary Graphs

Discrete Wavelet Transforms in 2 slides 1

(a) 2 Channel Filterbank (b) Tree-structured Filterbank

From Vetterli and Kovacevic, Wavelets and Subband Coding, 95

A. Ortega (USC) Signal Processing on Graphs Sept. 2013 22 / 81

 Wavelet Transforms on Arbitrary Graphs

Discrete Wavelet Transforms in 2 slides 2

(a) Separable Transform (b) Example Image

Note: Filters have some frequency and space localization

From Vetterli and Kovacevic, [Ding07]

A. Ortega (USC) Signal Processing on Graphs Sept. 2013 23 / 81

 Wavelet Transforms on Arbitrary Graphs Graph Transforms Prior Work

Prior work Spatial Graph Transforms

Designed in the vertex domain of the graph. Examples:
Graph wavelets [Crovella03]
Approaches for WSN [Wang06], [Wagner05] [Shen-ICASSP08]

1-hop averaging transform

N
1 X
y [n] = A[n, m]x[m] y = D1 Ax = Prw x
dn
m=1

1-hop difference transform

N
1 X
y [n] = A[n, m](x[n] x[m]) y = Lrw x = x Prw x
dn
m=1

A. Ortega (USC) Signal Processing on Graphs Sept. 2013 24 / 81

 Wavelet Transforms on Arbitrary Graphs Graph Transforms Prior Work

Prior Work Spectral Graph Transforms

Designed in the spectral domain of the graph. Examples:
Diffusion Wavelets [Coifman and Maggioni 2006]
Spectral Wavelets on Graphs [Hammond et al. 2011]

Spectral Wavelet transforms [Hammond et al. 2011]:

Design spectral kernels: h() : (G ) R.

Th = h(L) = Uh()Ut

where
h() = diag {h(i )}

A. Ortega (USC) Signal Processing on Graphs Sept. 2013 25 / 81

 Wavelet Transforms on Arbitrary Graphs Graph Transforms Prior Work

Spectral Graph Transforms Contd

Output Coefficients:
X
w f = Th f = h().f()u
(G )

Polynomial kernel approximation:

K
X
h() ak k
k=0

K
X
Th ak Lk
k=0

K -hop localized: no spectral decomposition required.

A. Ortega (USC) Signal Processing on Graphs Sept. 2013 26 / 81
 Wavelet Transforms on Arbitrary Graphs Graph Transforms Prior Work

Vertex-Frequency Localization on Graphs

Wavelet Filters: provide simultaneous localization in spatial and

spectral domain:

A. Ortega (USC) Signal Processing on Graphs Sept. 2013 27 / 81

 Wavelet Transforms on Arbitrary Graphs Graph Transforms Prior Work

Vertex-Frequency Localization on Graphs

Wavelet Filters: provide simultaneous localization in spatial and

spectral domain:

A. Ortega (USC) Signal Processing on Graphs Sept. 2013 27 / 81

 Wavelet Transforms on Arbitrary Graphs Graph Transforms Prior Work

Vertex-Frequency Localization on Graphs

Wavelet Filters: provide simultaneous localization in spatial and

spectral domain:

A. Ortega (USC) Signal Processing on Graphs Sept. 2013 27 / 81

 Wavelet Transforms on Arbitrary Graphs Graph Transforms Prior Work

Vertex-Frequency Localization on Graphs

Wavelet Filters: provide simultaneous localization in spatial and

spectral domain:

A. Ortega (USC) Signal Processing on Graphs Sept. 2013 27 / 81

 Wavelet Transforms on Arbitrary Graphs Graph Transforms Prior Work

Vertex-Frequency Localization on Graphs

Wavelet Filters: provide simultaneous localization in spatial and

spectral domain:

A. Ortega (USC) Signal Processing on Graphs Sept. 2013 27 / 81

 Wavelet Transforms on Arbitrary Graphs Graph Transforms Prior Work

Vertex-Frequency Localization on Graphs

Wavelet Filters: provide simultaneous localization in spatial and

spectral domain:

A. Ortega (USC) Signal Processing on Graphs Sept. 2013 27 / 81

 Wavelet Transforms on Arbitrary Graphs Graph Transforms Prior Work

Vertex-Frequency Localization on Graphs

Wavelet Filters: provide simultaneous localization in spatial and

spectral domain:

A. Ortega (USC) Signal Processing on Graphs Sept. 2013 27 / 81

 Wavelet Transforms on Arbitrary Graphs Graph Transforms Prior Work

Vertex-Frequency Localization on Graphs

Wavelet Filters: provide simultaneous localization in spatial and

spectral domain:

A. Ortega (USC) Signal Processing on Graphs Sept. 2013 27 / 81

 Wavelet Transforms on Arbitrary Graphs Graph Transforms Prior Work

Vertex-Frequency Localization on Graphs

Wavelet Filters: provide simultaneous localization in spatial and

spectral domain:

Advantages:
Possible benefits of localized frequency analysis.
Fast approximate solutions to global optimization problems.

A. Ortega (USC) Signal Processing on Graphs Sept. 2013 27 / 81

 Wavelet Transforms on Arbitrary Graphs Graph Transforms Prior Work

Graph Filterbank Designs

Formulation of critically sampled graph filterbank design problem

Design filters using spectral techniques [Hammond et al. 2009].
Orthogonal (not compactly supported) [IEEE TSP June 2012]
Bi-Orthogonal (compactly supported) [IEEE TSP Oct 2013]
analysis side synthesis side

- -

lter downsample upsample lter

A. Ortega (USC) Signal Processing on Graphs Sept. 2013 28 / 81

 Wavelet Transforms on Arbitrary Graphs Downsampling in Graphs

Downsampling/Upsampling in Graphs

Downsampling-upsampling operation:
Regular Signals: (a) regular signal (b) regular signal after DU by 2


f (n) if n = 2m
fdu (n) =
0 if n = 2m + 1

Graph signals:
 (c) graph signal (d) graph signal after DU by 2

f (n) if n S
fdu (n) =
0 if n
/S

for some set S.

For regular signals DU by 2 operation is equivalent to
Fdu (e j ) = 1/2(F (e j ) + F (e j )) in the DFT domain.
What is the DU by 2 for graph signals in GFT domain?
A. Ortega (USC) Signal Processing on Graphs Sept. 2013 29 / 81
 Wavelet Transforms on Arbitrary Graphs Graph Transform Designs

Downsampling in Graphs

Downsampling function : define H : V {1} s.t.


1 if n H
H (n) = (1)
1 if n /H

Downsample-upsample (DU) operation given H :

1
fdu (n) = [f (n) + H (n)f (n)] (2)
2

A. Ortega (USC) Signal Processing on Graphs Sept. 2013 30 / 81

 Wavelet Transforms on Arbitrary Graphs Graph Transform Designs

Downsampling in Graphs

Define J = JH = diag {H (n)}.

In vector form:
1
fdu = (f + J f)
2
1
= (f + f)
2

A. Ortega (USC) Signal Processing on Graphs Sept. 2013 31 / 81

 Wavelet Transforms on Arbitrary Graphs Graph Transform Designs

Downsampling in Graphs

Define J = JH = diag {H (n)}.

In vector form:
1
fdu = (f + J f)
2
1
= (f + f)
2

Spectral Folding [4]: For a bipartite graph f() = f (2 ).

A. Ortega (USC) Signal Processing on Graphs Sept. 2013 31 / 81

 Wavelet Transforms on Arbitrary Graphs Graph Transform Designs

Downsampling in Graphs

Define J = JH = diag {H (n)}.

In vector form:
1
fdu = (f + J f)
2
1
= (f + f)
2

Spectral Folding [4]: For a bipartite graph f() = f (2 ).

A. Ortega (USC) Signal Processing on Graphs Sept. 2013 31 / 81

 Wavelet Transforms on Arbitrary Graphs Graph Transform Designs

DFT aliasing vs GFT aliasing

Property DU by 2 regular signal DU by 2 bipartite graph

signal
frequency k = 2kN ; uniformly spaced in eigenvalues of L; irregu-
[0 2], larly spaced in [0 2]

Fourier ba- WNk = exp{jk n}; complex eigenvectors u of L; real

sis

frequency Fdu (e j ) = 1/2(F (e j ) + fdu () = 1/2(f()+f(2))

folding F (e j ))

A. Ortega (USC) Signal Processing on Graphs Sept. 2013 32 / 81

 Wavelet Transforms on Arbitrary Graphs Graph Transform Designs

Graph filterbanks

Filters designed in spectral domain (as [Hammond et al, 2009])

Analysis:
hi () : R R for i = 0, 1
Hi = hi (L) = Uhi ()Ut
Synthesis:
gi () : R R analysis side synthesis side
Gi = gi (L)

- -

lter downsample upsample lter

A. Ortega (USC) Signal Processing on Graphs Sept. 2013 33 / 81

 Wavelet Transforms on Arbitrary Graphs Graph Transform Designs

Graph filterbanks
Downsampling functions H = and L = in two channels.
nodes in H (or L) store the output of H1 (or H0 ) critically
sampled output.

Equivalent transform f = Teq f, s.t.,

1 1
Teq = G1 (I + J )H1 + G0 (I J )H0
2 2
1 1
= (G1 H1 + G0 H0 ) + (G1 J H1 G0 J H0 ) (3)
2| {z } 2| {z }
A B

B term is due to downsampling. For perfect reconstruction A = cI

and B = 0.

A. Ortega (USC) Signal Processing on Graphs Sept. 2013 34 / 81

 Wavelet Transforms on Arbitrary Graphs Graph Transform Designs

Graph filterbanks
Downsampling functions H = and L = in two channels.
nodes in H (or L) store the output of H1 (or H0 ) critically
sampled output.

Equivalent transform f = Teq f, s.t.,

1 1
Teq = G1 (I + J )H1 + G0 (I J )H0
2 2
1 1
= (G1 H1 + G0 H0 ) + (G1 J H1 G0 J H0 ) (3)
2| {z } 2| {z }
A B

B term is due to downsampling. For perfect reconstruction A = cI

and B = 0.
Since we use spectral filtering: choosing Hi is equivalent to choosing
hi ()

A. Ortega (USC) Signal Processing on Graphs Sept. 2013 34 / 81

 Wavelet Transforms on Arbitrary Graphs Graph Transform Designs

Wavelet filterbanks on bipartite graphs

Aliasing Cancellation B = 0 if for all (G ):

B() = g1 ()h1 (2 ) g0 ()h0 (2 ) = 0

A. Ortega (USC) Signal Processing on Graphs Sept. 2013 35 / 81

 Wavelet Transforms on Arbitrary Graphs Graph Transform Designs

Wavelet filterbanks on bipartite graphs

Aliasing Cancellation B = 0 if for all (G ):

B() = g1 ()h1 (2 ) g0 ()h0 (2 ) = 0

Perfect Reconstruction A = cI if for all (G ):

A() = g1 ()h1 () + g0 ()h0 () = c

A. Ortega (USC) Signal Processing on Graphs Sept. 2013 35 / 81

 Wavelet Transforms on Arbitrary Graphs Graph Transform Designs

Graph-QMF design 1
Solution analogous to Quadrature Mirror Filters (QMF), choose:
h1 () = h0 (2 )
g0 () = h0 ()
g1 () = h1 ()

A. Ortega (USC) Signal Processing on Graphs Sept. 2013 36 / 81

 Wavelet Transforms on Arbitrary Graphs Graph Transform Designs

Graph-QMF design 1
Solution analogous to Quadrature Mirror Filters (QMF), choose:
h1 () = h0 (2 )
g0 () = h0 ()
g1 () = h1 ()

Design h0 () s.t. for all

h02 () + h02 (2 ) = c

no exact polynomial solutions, good polynomial approximations

A. Ortega (USC) Signal Processing on Graphs Sept. 2013 36 / 81
 Wavelet Transforms on Arbitrary Graphs Graph Transform Designs

Graph-QMF design 2

Polynomial kernel approximation:

Approximate Meyer kernels as m degree polynomial.
trade off between accuracy and complexity .

A. Ortega (USC) Signal Processing on Graphs Sept. 2013 37 / 81

 Wavelet Transforms on Arbitrary Graphs Graph Transform Designs

Graph-QMF design 2

Polynomial kernel approximation:

Approximate Meyer kernels as m degree polynomial.
trade off between accuracy and complexity .

A. Ortega (USC) Signal Processing on Graphs Sept. 2013 37 / 81

 Wavelet Transforms on Arbitrary Graphs Graph Transform Designs

Graph-QMF design 2

Polynomial kernel approximation:

Approximate Meyer kernels as m degree polynomial.
trade off between accuracy and complexity .

A. Ortega (USC) Signal Processing on Graphs Sept. 2013 37 / 81

 Wavelet Transforms on Arbitrary Graphs Graph Transform Designs

Graph-QMF design 2

Polynomial kernel approximation:

Approximate Meyer kernels as m degree polynomial.
trade off between accuracy and complexity .

A. Ortega (USC) Signal Processing on Graphs Sept. 2013 37 / 81

 Wavelet Transforms on Arbitrary Graphs Graph Transform Designs

Graph-QMF design 2

Polynomial kernel approximation:

Approximate Meyer kernels as m degree polynomial.
trade off between accuracy and complexity .

A. Ortega (USC) Signal Processing on Graphs Sept. 2013 37 / 81

 Wavelet Transforms on Arbitrary Graphs Graph Transform Designs

Graph-QMF design 2

Polynomial kernel approximation:

Approximate Meyer kernels as m degree polynomial.
trade off between accuracy and complexity .

A. Ortega (USC) Signal Processing on Graphs Sept. 2013 37 / 81

 Wavelet Transforms on Arbitrary Graphs Graph Transform Designs

GraphBior design 1

Analogous to CDF wavelet Filters [Narang and Ortega, IEEE TSP, 2013]

A. Ortega (USC) Signal Processing on Graphs Sept. 2013 38 / 81

 Wavelet Transforms on Arbitrary Graphs Graph Transform Designs

GraphBior design 1

Analogous to CDF wavelet Filters [Narang and Ortega, IEEE TSP, 2013]
Choose kernels, s.t.,

h0 () = g1 (2 )
g0 () = h1 (2 ),

for aliasing cancellation (B = 0).

A. Ortega (USC) Signal Processing on Graphs Sept. 2013 38 / 81

 Wavelet Transforms on Arbitrary Graphs Graph Transform Designs

GraphBior design 1

Analogous to CDF wavelet Filters [Narang and Ortega, IEEE TSP, 2013]
Choose kernels, s.t.,

h0 () = g1 (2 )
g0 () = h1 (2 ),

for aliasing cancellation (B = 0).

The PR condition (A = 0) becomes:

h1 ()g1 () + h1 (2 )g1 (2 ) = c
| {z } | {z }
p() p(2)

A. Ortega (USC) Signal Processing on Graphs Sept. 2013 38 / 81

 Wavelet Transforms on Arbitrary Graphs Graph Transform Designs

GraphBior design 1

Analogous to CDF wavelet Filters [Narang and Ortega, IEEE TSP, 2013]
Choose kernels, s.t.,

h0 () = g1 (2 )
g0 () = h1 (2 ),

for aliasing cancellation (B = 0).

The PR condition (A = 0) becomes:

h1 ()g1 () + h1 (2 )g1 (2 ) = c
| {z } | {z }
p() p(2)

Design p() as a maximally flat polynomial and factorize into

h1 (), g1 () terms. Exact reconstruction with polynomial filter
(compact support).
A. Ortega (USC) Signal Processing on Graphs Sept. 2013 38 / 81
 Wavelet Transforms on Arbitrary Graphs Graph Transform Designs

GraphBior design 3
Trade-off between spatial and spectral localization:
All solutions satisfy perfect reconstruction.
Spectral localization increases with longer filters.

A. Ortega (USC) Signal Processing on Graphs Sept. 2013 39 / 81

 Wavelet Transforms on Arbitrary Graphs Graph Transform Designs

GraphBior design 3
Trade-off between spatial and spectral localization:
All solutions satisfy perfect reconstruction.
Spectral localization increases with longer filters.

A. Ortega (USC) Signal Processing on Graphs Sept. 2013 39 / 81

 Wavelet Transforms on Arbitrary Graphs Graph Transform Designs

GraphBior design 3
Trade-off between spatial and spectral localization:
All solutions satisfy perfect reconstruction.
Spectral localization increases with longer filters.

A. Ortega (USC) Signal Processing on Graphs Sept. 2013 39 / 81

 Wavelet Transforms on Arbitrary Graphs Graph Transform Designs

GraphBior design 3
Trade-off between spatial and spectral localization:
All solutions satisfy perfect reconstruction.
Spectral localization increases with longer filters.

A. Ortega (USC) Signal Processing on Graphs Sept. 2013 39 / 81

 Wavelet Transforms on Arbitrary Graphs Graph Transform Designs

GraphBior design 3
Trade-off between spatial and spectral localization:
All solutions satisfy perfect reconstruction.
Spectral localization increases with longer filters.

A. Ortega (USC) Signal Processing on Graphs Sept. 2013 39 / 81

 Wavelet Transforms on Arbitrary Graphs Graph Transform Designs

GraphBior design 3
Trade-off between spatial and spectral localization:
All solutions satisfy perfect reconstruction.
Spectral localization increases with longer filters.

A. Ortega (USC) Signal Processing on Graphs Sept. 2013 39 / 81

 Wavelet Transforms on Arbitrary Graphs Graph Transform Designs

GraphBior design 3
Trade-off between spatial and spectral localization:
All solutions satisfy perfect reconstruction.
Spectral localization increases with longer filters.

A. Ortega (USC) Signal Processing on Graphs Sept. 2013 39 / 81

 Wavelet Transforms on Arbitrary Graphs Bipartite Subgraph Decomposition

Bipartite Subgraph Decomposition

But not all graphs are bipartite...

A. Ortega (USC) Signal Processing on Graphs Sept. 2013 40 / 81

 Wavelet Transforms on Arbitrary Graphs Bipartite Subgraph Decomposition

Bipartite Subgraph Decomposition

But not all graphs are bipartite...

Solution: Iteratively decompose non-bipartite graph G into K
bipartite subgraphs:
each subgraph covers the same vertex set.
each edge in G belongs to exactly one bipartite graph.

A. Ortega (USC) Signal Processing on Graphs Sept. 2013 40 / 81

 Wavelet Transforms on Arbitrary Graphs Bipartite Subgraph Decomposition

Bipartite Subgraph Decomposition

But not all graphs are bipartite...

Solution: Iteratively decompose non-bipartite graph G into K
bipartite subgraphs:
each subgraph covers the same vertex set.
each edge in G belongs to exactly one bipartite graph.
apply wavelet filterbanks in K stages (dimensions).

in the k th stage restrict filtering downsampling operations on k th

bipartite graph.

A. Ortega (USC) Signal Processing on Graphs Sept. 2013 40 / 81

 Wavelet Transforms on Arbitrary Graphs Bipartite Subgraph Decomposition

Bipartite Subgraph Decomposition

Example of a 2-dimensional (K = 2) decomposition:

A. Ortega (USC) Signal Processing on Graphs Sept. 2013 41 / 81

 Wavelet Transforms on Arbitrary Graphs Bipartite Subgraph Decomposition

Bipartite Subgraph Decomposition

Example of a 2-dimensional (K = 2) decomposition:

A. Ortega (USC) Signal Processing on Graphs Sept. 2013 41 / 81

 Wavelet Transforms on Arbitrary Graphs Bipartite Subgraph Decomposition

Bipartite Subgraph Decomposition

Example of a 2-dimensional (K = 2) decomposition:

A. Ortega (USC) Signal Processing on Graphs Sept. 2013 41 / 81

 Wavelet Transforms on Arbitrary Graphs Bipartite Subgraph Decomposition

Bipartite Subgraph Decomposition

Example of a 2-dimensional (K = 2) decomposition:

A. Ortega (USC) Signal Processing on Graphs Sept. 2013 41 / 81

 Wavelet Transforms on Arbitrary Graphs Bipartite Subgraph Decomposition

Bipartite Subgraph Decomposition

Example of a 2-dimensional (K = 2) decomposition:

A. Ortega (USC) Signal Processing on Graphs Sept. 2013 41 / 81

 Wavelet Transforms on Arbitrary Graphs Bipartite Subgraph Decomposition

Multi-dimensional Filterbanks on graphs

Two-dimensional two-channel filterbank on graphs:

2 2

2
2

2 2

2 2

Advantages:
Perfect reconstruction and orthogonal for any graph and any bpt
decomposition.
defined metrics to find good bipartite decompositions.
A. Ortega (USC) Signal Processing on Graphs Sept. 2013 42 / 81
 Wavelet Transforms on Arbitrary Graphs Example

Example

Minnesota traffic graph and graph signal

A. Ortega (USC) Signal Processing on Graphs Sept. 2013 43 / 81

 Wavelet Transforms on Arbitrary Graphs Example

Example

Bipartite decomposition

A. Ortega (USC) Signal Processing on Graphs Sept. 2013 44 / 81

 Wavelet Transforms on Arbitrary Graphs Example

Example

2 0 2 0.1 0 0.1

LL Channel LH Channel

0.1 0 0.1

Empty Channel

HL Channel HH Channel

Output coefficients of the proposed filterbanks with parameter m = 24.

A. Ortega (USC) Signal Processing on Graphs Sept. 2013 45 / 81

 Wavelet Transforms on Arbitrary Graphs Example

Example

1 0 1 0.1 0 0.1

LL Channel LH Channel

1 0 1 0.05 0 0.05

HL Channel HH Channel

Reconstructed graph-signals for each channel.

A. Ortega (USC) Signal Processing on Graphs Sept. 2013 46 / 81

 Applications

Next Section

1 Introduction

2 Wavelet Transforms on Arbitrary Graphs

3 Applications

4 Conclusions

A. Ortega (USC) Signal Processing on Graphs Sept. 2013 47 / 81

 Applications Edge Aware Image Processing

Depth Image Coding [Narang, Chao and Ortega, 2013]

Block Diagram

Edge Graph Edge

Detection Selection Encoding
(JBIG) Output
GraphBior Bit stream
Input Image Filterbanks

Wavelet Coefficients
Graph-based
Encoding
Wavelet Transform
(SPIHT)

A. Ortega (USC) Signal Processing on Graphs Sept. 2013 48 / 81

 Applications Edge Aware Image Processing

Depth Image Coding [Narang, Chao and Ortega, 2013]

CDF 9/7 Graph 9/7

A. Ortega (USC) Signal Processing on Graphs Sept. 2013 49 / 81

 Applications Edge Aware Image Processing

Depth Image Coding [Narang, Chao and Ortega, 2013]

Edge detection: Prewitt

Laplacian Normalization:
Random Walk Laplacian
Filterbanks: GraphBior 4/3 and
CDF 9/7
Unreliable Link Weight: 0.01
Transform level: 5
Encoder: SPIHT

A. Ortega (USC) Signal Processing on Graphs Sept. 2013 50 / 81

 Applications Bilateral Filtering as a Graph Operation

Bilateral Filtering (BF) [Tomasi and Manduchi, 98]

Weighted average of nearby similar pixels
X wij
xout [j] = P xin [i] (4)
i wij
i
with weights given by
kpi pj k2 (xin [i] xin [j])2
   
wij = exp . exp (5)
2s2 2x2

(a) Noisy data (b) Similarity weights (c) Filtered output (From Tomasi and
Manduchi, 1998)

A. Ortega (USC) Signal Processing on Graphs Sept. 2013 51 / 81

 Applications Bilateral Filtering as a Graph Operation

BF as a Graph Based Transform

Graph G = (V, E ) with

pixels as nodes
V = {1, 2, . . . , n}
edges E = {(i, j, wij )}
image xin as graph signal Bilateral Filter Graph

We can write bilateral filtering in (4) as

xout = D1 Wxin (6)

A. Ortega (USC) Signal Processing on Graphs Sept. 2013 52 / 81

 Applications Bilateral Filtering as a Graph Operation

Spectral Interpretation

Using the definition of graph Laplacian L = I D1/2 WD1/2

D1/2 xout = (I L)D1/2 xin (7)

Using L = UUt and x = D1/2 x

xout = U(I )Ut xin (8)

Iterated bilateral filter

xout = U(I )k Ut xin (9)

A. Ortega (USC) Signal Processing on Graphs Sept. 2013 53 / 81

 Applications Bilateral Filtering as a Graph Operation

Spectral Response of the BF

1

0.8 BF
Iterated BF k = 2
0.6 k=3
k=4
0.4
h()

0.2

0.2

0.4
0 0.2 0.4 0.6 0.8 1 1.2 1.4

Spectral responses of the BF and iterated BF. The graph is formed using the lena
image which has maximum eigenvalue equal to 1.28.
A. Ortega (USC) Signal Processing on Graphs Sept. 2013 54 / 81
 Applications Bilateral Filtering as a Graph Operation

Flexible Spectral Design [Gadde, Narang and Ortega, 2013]

Key idea: use graph derived from bilateral filter

U h() Ut xin = h(L)xin

xout = |{z} (10)
|{z} | {z }
Inverse Spectral GFT
GFT response

Design polynomial h() to have local implementation.

A. Ortega (USC) Signal Processing on Graphs Sept. 2013 55 / 81

 Applications Bilateral Filtering as a Graph Operation

Examples: Smoothing a noisy image

(d) (e) (f)

1 1.4
exact
1.2
approx.
0.5
1

0.8
0
0.6

0.4
0.5
0.2

1 0
0 0.5 1 1.5 2 0 0.5 1 1.5 2
(a) (b) (c)

(a) Original (d) Noisy SNR = 20 dB (b) Spectral response of the BF (c) Spectral
response obtained by the regularization (e) Output of the BF, SNR = 20.65 dB
(f) Output of h() filter, SNR = 22.64 dB

A. Ortega (USC) Signal Processing on Graphs Sept. 2013 56 / 81

 Applications Bilateral Filtering as a Graph Operation

Examples: Edge preserving coarsening

(a) (b) (c)

1 1.2

1
exact
0.8 approx.
0.8
0.6 0.6

0.4 0.4

0.2
0.2
0

0 0.2
0 0.5 1 1.5 2 0 0.5 1 1.5 2
(d) (e)

(a) Original image (b) 20 iterations of BF (d) Spectral response of the iterated
BF (c) output of the proposed spectral filter (e) Corresponding Spectral response
and its polynomial approximation

A. Ortega (USC) Signal Processing on Graphs Sept. 2013 57 / 81

 Applications Graph Filtering of Cost-to-Go Functions

Graph Filtering of Cost-to-Go Functions [Levorato, Narang, Mitra,

Ortega 2012]

Markov decision process:

0,a0 0,a1
S = {S(0), S(1), ...} sequence
of states 1,a0 1,a1

S(t) S state at time t

S state space 2,a 0 2,a1

3,a 0 3,a1

A = {A(1), A(2), ...} sequence

of actions
T,a 0 T,a 1
A(t) AS(t) action at time t
A action space.
Example of a FSM with T states and 2 actions

A. Ortega (USC) Signal Processing on Graphs Sept. 2013 58 / 81

 Applications Graph Filtering of Cost-to-Go Functions

Graph Filtering of Cost-to-Go Functions

Graph Formulation:
Nodes set : V = S A = {(s, a)}sS,aA .
Graph signal: expected long term discounted cost v (s, a) from state s
given action a conditioned upon the policy :
X X
X
V (s, a) = c(s, a) + p (s, a, s2 )(s2 , a2 )c(s2 , a2 )
=1 s2 S a2 A

An optimal policy exists in the set of randomized policies past

independent policies (s, a) : S A [0, 1] maps state s to the
probability that action a is selected.
Problem: Computation, compression and optimization of discounted cost
function v (s, a).

A. Ortega (USC) Signal Processing on Graphs Sept. 2013 59 / 81

 Applications Graph Filtering of Cost-to-Go Functions

Graph Filtering of Cost-to-Go Functions

Very large state space
Wavelet based approach:
Reduce the size of the problem by downsampling and filtering.
Operate upon the smooth approximation of cost function on
downsampled graph.
Example: Expected cost for secondary transmitter observing state
(which depends on unobserved primary transmitter)

A. Ortega (USC) Signal Processing on Graphs Sept. 2013 60 / 81

 Applications Graph Filtering of Cost-to-Go Functions

Graph Filtering of Cost-to-Go Functions

Results [Globecom, 2012]

0.35
= 0.9
2
= 0.8
2

0.3
Policy error

0.25

0.2

0.15

0.1
15 20 25 30 35 40
B = F

Error between policies computed on original and downsampled graph (as a

function of graph size.) 1 and 2 : transmission failure probabilities.
A. Ortega (USC) Signal Processing on Graphs Sept. 2013 61 / 81
 Applications Graph-based Prediction for Recommendation Systems

Graph based Prediction in Recommendation Systems [Gadde,

Narang, Ortega 2013]

Collaborative filtering problem: given known movie ratings for a large

set of users, identify recommendations for a specific user.
Graph representation of recommender systems:
movies (or users) as vertices and
edge-weights reflecting similarity between them.
Interpolation based methods for rating prediction:
find all movies that the specific user has rated and are neighbors in
weighted graph.
interpolate ratings of these movies to unknown movie.

A. Ortega (USC) Signal Processing on Graphs Sept. 2013 62 / 81

 Applications Graph-based Prediction for Recommendation Systems

Graph based Prediction in Recommendation Systems

kNN graph graph 2

Unrated Unrated
vertex Vertex

0.11789 0.75042 0.14664 1.1025

A typical instance of interpolation in MovieLens 100k dataset: (a) kNN method (err = 2.81 in
this example). (b) Interpolation based on local sub-graph (err = 0.78 in this case).

A. Ortega (USC) Signal Processing on Graphs Sept. 2013 63 / 81

 Applications Graph-based Prediction for Recommendation Systems

Preliminary results [ICASSP 2013]

1.4 LS projection (K=K*+10)

kNN (k=30)
1.2 PMF
Proposed method
Proposed method with bilateral weights
1

0.8
RMSE

0.6

0.4

0.2

0
140 4180 81120 121160 161200 201240 241280 281668
Number of training samples per movie

A. Ortega (USC) Signal Processing on Graphs Sept. 2013 64 / 81

 Conclusions

Next Section

1 Introduction

2 Wavelet Transforms on Arbitrary Graphs

3 Applications

4 Conclusions

A. Ortega (USC) Signal Processing on Graphs Sept. 2013 65 / 81

 Conclusions

What makes these graph transforms?

Graph-based shift invariance:

L1
X L1
X
H= k Lk or H = k Ak
k=0 k=0

Graph Fourier Transform

H = h(L) = Uh()U

A. Ortega (USC) Signal Processing on Graphs Sept. 2013 66 / 81

 Conclusions

Conclusions

Extending signal processing methods to arbitrary graphs:

Downsampling, Space-frequency, Multiresolution, Wavelets
Many open questions: very diverse types of graphs, results may apply
to special classes only
Outcomes
Work with massive graph-datasets: potential benefits of localized
frequency analysis
Novel insights about traditional applications (image/video processing)
To get started:
[Shuman, Narang, Frossard, Ortega, Vandergheysnt, SPM2013]
GlobalSIP Symposium on Graph Signal Processing

A. Ortega (USC) Signal Processing on Graphs Sept. 2013 67 / 81

 Conclusions References

References I

S.K. Narang and A. Ortega.

Lifting based wavelet transforms on graphs.
In Proc. of Asia Pacific Signal and Information Processing Association (APSIPA), October 2009.

S. Narang, G. Shen, and A. Ortega.

Unidirectional graph-based wavelet transforms for efficient data gathering in sensor networks.
In In Proc. of ICASSP10.
S. Narang and A. Ortega.
Downsampling Graphs using Spectral Theory
In In Proc. of ICASSP11.
G. Shen and A. Ortega.
Transform-based Distributed Data Gathering.
IEEE Transactions on Signal Processing.

G. Shen, S. Pattem, and A. Ortega.

Energy-efficient graph-based wavelets for distributed coding in wireless sensor networks.
In Proc. of ICASSP09, April 2009.

G. Shen, S. Narang, and A. Ortega.

Adaptive distributed transforms for irregularly sampled wireless sensor networks.
In Proc. of ICASSP09, April 2009.

G. Shen and A. Ortega.

Tree-based wavelets for image coding: Orthogonalization and tree selection.
In Proc. of PCS09, May 2009.

A. Ortega (USC) Signal Processing on Graphs Sept. 2013 68 / 81

 Conclusions References

References II

I.F. Akyildiz, W. Su, Y. Sankarasubramaniam, and E. Cayirci.

A survey on sensor networks.
IEEE Communication Magazine, 40(8):102114, August 2002.

R. Baraniuk, A. Cohen, and R. Wagner.

Approximation and compression of scattered data by meshless multiscale decompositions.
Applied Computational Harmonic Analysis, 25(2):133147, September 2008.

C.L. Chang and B. Girod.

Direction-adaptive discrete wavelet transform for image compression.
IEEE Transactions on Image Processing, 16(5):12891302, May 2007.

C. Chong and S. P. Kumar.

Sensor networks: Evolution, opportunities, and challenges.
Proceedings of the IEEE, 91(8):12471256, August 2003.

R.R. Coifman and M. Maggioni.

Diffusion wavelets.
Applied Computational Harmonic Analysis, 21(1):5394, 2006.

R. Cristescu, B. Beferull-Lozaon, and M. Vetterli.

Networked Slepian-Wolf: Theory, algorithms, and scaling laws.
IEEE Transactions on Information Theory, 51(12):40574073, December 2005.

M. Crovella and E. Kolaczyk.

Graph wavelets for spatial traffic analysis.
In IEEE INFOCOMM, 2003.

A. Ortega (USC) Signal Processing on Graphs Sept. 2013 69 / 81

 Conclusions References

References III

TinyOS-2.
Collection tree protocol.
http://www.tinyos.net/tinyos-2.x/doc/.

I. Daubechies, I. Guskov, P. Schroder, and W. Sweldens.

Wavelets on irregular point sets.
Phil. Trans. R. Soc. Lond. A, 357(1760):23972413, September 1999.

W. Ding, F. Wu, X. Wu, S. Li, and H. Li.

Adaptive directional lifting-based wavelet transform for image coding.
IEEE Transactions on Image Processing, 16(2):416427, February 2007.

M. Gastpar, P. Dragotti, and M. Vetterli.

The distributed Karhunen-Loeve transform.
IEEE Transactions on Information Theory, 52(12):51775196, December 2006.

B. Girod and S. Han.

Optimum update for motion-compensated lifting.
IEEE Signal Processing Letters, 12(2):150153, February 2005.

V.K. Goyal.
Theoretical foundations of transform coding.
IEEE Signal Processing Magazine, 18(5):921, September 2001.

S. Haykin.
Adaptive Filter Theory.
Prentice Hall, 4th edition, 2004.

A. Ortega (USC) Signal Processing on Graphs Sept. 2013 70 / 81

 Conclusions References

References IV

W. R. Heinzelman, A. Chandrakasan, and H. Balakrishnan.

Energy-efficient routing protocols for wireless microsensor networks.
In Proc. of Hawaii Intl. Conf. on Sys. Sciences, January 2000.

M. Jansen, G. Nason, and B. Silverman.

Scattered data smoothing by empirical Bayesian shrinkage of second generation wavelet coefficients.
In Wavelets: Applications in Signal and Image Processing IX, Proc. of SPIE, 2001.

D. Jungnickel.
Graphs, Networks and Algorithms.
Springer-Verlag Press, 2nd edition, 2004.

M. Maitre and M. N. Do,

Shape-adaptive wavelet encoding of depth maps,
In Proc. of PCS09, 2009.

K. Mechitov, W. Kim, G. Agha, and T. Nagayama.

High-frequency distributed sensing for structure monitoring.
In In Proc. First Intl. Workshop on Networked Sensing Systems (INSS), 2004.

Y. Morvan, P.H.N. de With, and D. Farin,

Platelet-based coding of depth maps for the transmission of multiview images,
2006, vol. 6055, SPIE.

S. Pattem, B. Krishnamachari, and R. Govindan.

The impact of spatial correlation on routing with compression in wireless sensor networks.
ACM Transactions on Sensor Networks, 4(4):6066, August 2008.

A. Ortega (USC) Signal Processing on Graphs Sept. 2013 71 / 81

 Conclusions References

References V

E. Le Pennec and S. Mallat.

Sparse geometric image representations with bandelets.
IEEE Transactions on Image Processing, 14(4):423 438, April 2005.

J.G. Proakis, E.M. Sozer, J.A. Rice, and M. Stojanovic.

Shallow water acoustic networks.
IEEE Communications Magazine, 39(11):114119, 2001.

P. Rickenbach and R. Wattenhofer.

Gathering correlated data in sensor networks.
In Proceedings of the 2004 Joint Workshop on Foundations of Mobile Computing, October 2004.

S. Lafon and A. B. Lee, Diffusion maps and coarse-graining: A unified framework for dimensionality reduction, graph
partitioning, and data set parameterization, IEEE Trans. Pattern Anal. Mach. Intell., vol. 28, no. 9, pp. 13931403, Sep.
2006.
D. Ron, I. Safro, and A. Brandt, Relaxation-based coarsening and multiscale graph organization, Multiscale Model.
Simul., vol. 9, no. 1, pp. 407423, Sep. 2011.

G. Karypis, and V. Kumar, Multilevel k-way Partitioning Scheme for Irregular Graphs, J. Parallel Distrib. Comput.
vol. 48(1), pp. 96-129, 1998.

A. Said and W.A. Pearlman.

A New, Fast, and Efficient Image Codec Based on Set Partitioning in Hierarchical Trees.
IEEE Transactions on Circuits and Systems for Video Technology, 6(3):243 250, June 1996.

A. Ortega (USC) Signal Processing on Graphs Sept. 2013 72 / 81

 Conclusions References

References VI

A. Sanchez, G. Shen, and A. Ortega,

Edge-preserving depth-map coding using graph-based wavelets,
In Proc. of Asilomar09, 2009.

G. Shen and A. Ortega.

Optimized distributed 2D transforms for irregularly sampled sensor network grids using wavelet lifting.
In Proc. of ICASSP08, April 2008.

G. Shen and A. Ortega.

Joint routing and 2D transform optimization for irregular sensor network grids using wavelet lifting.
In IPSN 08, April 2008.

G. Shen and A. Ortega.

Compact image representation using wavelet lifting along arbitrary trees.
In Proc. of ICIP08, October 2008.

G. Shen, S. Pattem, and A. Ortega.

Energy-efficient graph-based wavelets for distributed coding in wireless sensor networks.
In Proc. of ICASSP09, April 2009.

G. Shen, S. Narang, and A. Ortega.

Adaptive distributed transforms for irregularly sampled wireless sensor networks.
In Proc. of ICASSP09, April 2009.

G. Shen and A. Ortega.

Tree-based wavelets for image coding: Orthogonalization and tree selection.
In Proc. of PCS09, May 2009.

A. Ortega (USC) Signal Processing on Graphs Sept. 2013 73 / 81

 Conclusions References

References VII

G. Shen, W.-S. Kim, A. Ortega, J. Lee and H.C. Wey.

Edge-aware Intra Prediction for Depth Map Coding.
Submitted to Proc. of ICIP10.
G. Shen and A. Ortega.
Transform-based Distributed Data Gathering.
To Appear in IEEE Transactions on Signal Processing.

G. Strang.
Linear Algebra and its Applications.
Thomson Learning, 3rd edition, 1988.

W. Sweldens.
The lifting scheme: A construction of second generation wavelets.
Tech. report 1995:6, Industrial Mathematics Initiative, Department of Mathematics, University of South Carolina, 1995.

M. Tanimoto, T. Fujii, and K. Suzuki,

View synthesis algorithm in view synthesis reference software 2.0(VSRS2.0), ISO/IEC JTC1/SC29/WG11, Feb. 2009.

G. Valiente.
Algorithms on Trees and Graphs.
Springer, 1st edition, 2002.

V. Velisavljevic, B. Beferull-Lozano, M. Vetterli, and P.L. Dragotti.

Directionlets: Anisotropic multidirectional representation with separable filtering.
IEEE Transactions on Image Processing, 15(7):1916 1933, July 2006.

A. Ortega (USC) Signal Processing on Graphs Sept. 2013 74 / 81

 Conclusions References

References VIII
R. Wagner, H. Choi, R. Baraniuk, and V. Delouille.
Distributed wavelet transform for irregular sensor network grids.
In IEEE Stat. Sig. Proc. Workshop (SSP), July 2005.

R. Wagner, R. Baraniuk, S. Du, D.B. Johnson, and A. Cohen.

An architecture for distributed wavelet analysis and processing in sensor networks.
In IPSN 06, April 2006.

A. Wang and A. Chandraksan.

Energy-efficient DSPs for wireless sensor networks.
IEEE Signal Processing Magazine, 19(4):6878, July 2002.

Y. Zhu, K. Sundaresan, and R. Sivakumar.

Practical limits on achievable energy improvements and useable delay tolerance in correlation aware data gathering in
wireless sensor networks.
In IEEE SECON05, September 2005.

S.K. Narang, G. Shen and A. Ortega,

Unidirectional Graph-based Wavelet Transforms for Efficient Data Gathering in Sensor Networks.
pp.2902-2905, ICASSP10, Dallas, April 2010.

S.K. Narang and A. Ortega,

Local Two-Channel Critically Sampled Filter-Banks On Graphs,
Intl. Conf. on Image Proc. (2010),

R. R. Coifman and M. Maggioni,

Diffusion Wavelets,
Appl. Comp. Harm. Anal., vol. 21 no. 1 (2006), pp. 5394

A. Ortega (USC) Signal Processing on Graphs Sept. 2013 75 / 81

 Conclusions References

References IX
A. Sandryhaila and J. Moura,
Discrete Signal Processing on Graphs
IEEE Transactions on Signal Processing, 2013

D. K. Hammond, P. Vandergheynst, and R. Gribonval,

Wavelets on graphs via spectral graph theory,
Applied and Computational Harmonic Analysis, March 2011.

D. Shuman, S. K. Narang, P. Frossard, A. Ortega, P. Vandergheynst,

Signal Processing on Graphs: Extending High-Dimensional Data Analysis to Networks and Other Irregular Data
Domains
Signal Processing Magazine, May 2013

M. Crovella and E. Kolaczyk,

Graph wavelets for spatial traffic analysis,
in INFOCOM 2003, Mar 2003, vol. 3, pp. 18481857.

G. Shen and A. Ortega,

Optimized distributed 2D transforms for irregularly sampled sensor network grids using wavelet lifting,
in ICASSP08, April 2008, pp. 25132516.

W. Wang and K. Ramchandran,

Random multiresolution representations for arbitrary sensor network graphs,
in ICASSP, May 2006, vol. 4, pp. IVIV.

R. Wagner, H. Choi, R. Baraniuk, and V. Delouille.

Distributed wavelet transform for irregular sensor network grids.
In IEEE Stat. Sig. Proc. Workshop (SSP), July 2005.

A. Ortega (USC) Signal Processing on Graphs Sept. 2013 76 / 81

 Conclusions References

References X

S. K. Narang and A. Ortega,

Lifting based wavelet transforms on graphs,
(APSIPA ASC 09), 2009.

B. Zeng and J. Fu,

Directional discrete cosine transforms for image coding,
in Proc. of ICME 2006, 2006.

E. Le Pennec and S. Mallat,

Sparse geometric image representations with bandelets,
IEEE Trans. Image Proc., vol. 14, no. 4, pp. 423438, Apr. 2005.

M. Vetterli V. Velisavljevic, B. Beferull-Lozano and P.L. Dragotti,

Directionlets: Anisotropic multidirectional representation with separable filtering,
IEEE Trans. Image Proc., vol. 15, no. 7, pp. 19161933, Jul. 2006.

P.H.N. de With Y. Morvan and D. Farin,

Platelet-based coding of depth maps for the transmission of multiview images,
in In Proceedings of SPIE: Stereoscopic Displays and Applications, 2006, vol. 6055.

M. Tanimoto, T. Fujii, and K. Suzuki,

View synthesis algorithm in view synthesis reference software 2.0 (VSRS2.0),
Tech. Rep. Document M16090, ISO/IEC JTC1/SC29/WG11, Feb. 2009.

S. K. Narang and A. Ortega,

Local two-channel critically-sampled filter-banks on graphs,
In ICIP10, Sep. 2010.

A. Ortega (USC) Signal Processing on Graphs Sept. 2013 77 / 81

 Conclusions References

References XI

S.K. Narang and Ortega A.,

Perfect reconstruction two-channel wavelet filter-banks for graph structured data,
IEEE trans. on Sig. Proc., vol. 60, no. 6, June 2012.

J. P-Trufero, S.K. Narang and A. Ortega,

Distributed Transforms for Efficient Data Gathering in Arbitrary Networks,
In ICIP11., Sep 2011.

E. M-Enrquez, F. Daz-de-Mara and A. Ortega,

Video Encoder Based On Lifting Transforms On Graphs,
In ICIP11., Sep 2011.

Gilbert Strang,
The discrete cosine transform,
SIAM Review, vol. 41, no. 1, pp. 135147, 1999.

S.K. Narang and A. Ortega,

Downsampling graphs using spectral theory,
in ICASSP 11., May 2011.

Fattal, R. (2009).
Edge-avoiding wavelets and their applications.
In SIGGRAPH 09: ACM SIGGRAPH 2009 papers, pages 110, New York, NY, USA. ACM.

Hsu, C.-P. (1983).

Minimum-via topological routing.
Computer-Aided Design of Integrated Circuits and Systems, IEEE Transactions on, 2(4):235 246.

A. Ortega (USC) Signal Processing on Graphs Sept. 2013 78 / 81

 Conclusions References

References XII

Kim, B.-J. and Pearlman, W. A. (1997).

An embedded wavelet video coder using three-dimensional set partitioning in hierarchical trees (SPIHT).
In Data Compression Conference, 1997. DCC 97. Proceedings, pages 251 260.

Le Pennec, E. and Mallat, S. (2005).

Sparse geometric image representations with bandelets.
Image Processing, IEEE Transactions on, 14(4):423 438.

Martnez Enrquez, E. and Ortega, A. (2011).

Lifting transforms on graphs for video coding.
In Data Compression Conference (DCC), 2011, pages 73 82.

Pesquet-Popescu, B. and Bottreau, V. (2001).

Three-dimensional lifting schemes for motion compensated video compression.
In ICASSP 01: Proceedings of the Acoustics, Speech, and Signal Processing, 2001. on IEEE International Conference,
pages 17931796, Washington, DC, USA.

I. Pesenson,
Sampling in Paley-Wiener spaces on combinatorial graphs,
Trans. Amer. Math. Soc, vol. 360, no. 10, pp. 56035627, 2008.

Secker, A. and Taubman, D. (2003).

Lifting-based invertible motion adaptive transform (limat) framework for highly scalable video compression.
Image Processing, IEEE Transactions on, 12(12):1530 1542.

A. Ortega (USC) Signal Processing on Graphs Sept. 2013 79 / 81

 Conclusions References

References XIII
Shapiro, J. M. (1992).
An embedded wavelet hierarchical image coder.
In Acoustics, Speech, and Signal Processing, 1992. ICASSP-92., 1992 IEEE International Conference on, volume 4, pages
657 660 vol.4.
Shen, G. and Ortega, A. (2008).
Compact image representation using wavelet lifting along arbitrary trees.
In Image Processing, 2008. ICIP 2008. 15th IEEE International Conference on, pages 2808 2811.

Velisavljevic, V., Beferull-Lozano, B., Vetterli, M., and Dragotti, P. L. (2006).

Directionlets: anisotropic multidirectional representation with separable filtering.
Image Processing, IEEE Transactions on, 15(7):1916 1933.

M. Levorato, S.K. Narang, U. Mitra, A. Ortega.

Reduced Dimension Policy Iteration for Wireless Network Control via Multiscale Analysis.
Globecom, 2012.

A. Gjika, M. Levorato, A. Ortega, U. Mitra.

Online learning in wireless networks via directed graph lifting transform.
Allerton, 2012.

I. Daubechies and I. Guskov and P. Schrder and W. Sweldens.

Wavelets on Irregular Point Sets.
Phil. Trans. R. Soc. Lond. A, 1991.

W. W. Zachary.
An information flow model for conflict and fission in small groups.
Journal of Anthropological Research, 33, 452-473 (1977).

A. Ortega (USC) Signal Processing on Graphs Sept. 2013 80 / 81

 Conclusions References

References XIV

S. K. Narang, Y. H. Chao, and A. Ortega,

Graph-wavelet filterbanks for edge-aware image processing,
IEEE SSP Workshop, pp. 141144, Aug. 2012.

A. Sandryhaila and J.M.F. Moura (2013).

Discrete Signal Processing on Graphs.
Signal Processing, IEEE Transactions on, 61(7):1644-1656

S.K. Narang, A. Ortgea (2013).

Compact Support Biorthogonal Wavelet Filterbanks for Arbitrary Undirected Graphs.
Signal Processing, IEEE Transactions on

V. Ekambaram, G. Fanti, B. Ayazifar, K. Ramchandran.

Critically-Sampled Perfect-Reconstruction Spline-Wavelet Filterbanks for Graph Signals.
IEEE GlobalSIP 2013

A. Ortega (USC) Signal Processing on Graphs Sept. 2013 81 / 81