Upload
others
View
6
Download
0
Embed Size (px)
Citation preview
Simon Streicher and Johan du PreezStellenbosch University
Graph Coloring: Comparing
Cluster Graphs to Factor Graphs
A B
C D EF
G
A B
C D EF
G
A,D E,DA,E,D
A,E
A,B,E B,E
A,C,D,F
B,E,G
E,D,G
G
A,E,DA,B,EA,C,D,F B,E,G E,D,G
A B F GC D E
A B
C
D
E
FGA
B
DE G
A
DE
A B
C D EF
G
Graph Coloring
Practical example is a four coloring map problem:
You only need four colors to color-in a map with noneighboring countries having the same color.
Noted by Francis Guthrie in 1852Theorem proven by Appel and Haken in 1976
A B
C D EF
G
Graph Coloring
Practical example is a four coloring map problem:
You only need four colors to color-in a map with noneighboring countries having the same color.
Noted by Francis Guthrie in 1852Theorem proven by Appel and Haken in 1976
A B
C D EF
G
A B
C D EF
G
A B
C D EF
G
Graph Coloring
Practical example is a four coloring map problem:
You only need four colors to color-in a map with noneighboring countries having the same color.
Noted by Francis Guthrie in 1852Theorem proven by Appel and Haken in 1976
A B
C D EF
G
A B
C D EF
G
A B
C D EF
G
DF
A
C
B
G
E
A B
C D EF
G
Graph Coloring
DF
A
C
B
G
E
Undirected graph
Graph Coloring
DF
A
C
B
G
E
DF
A
C
Maximal cliques
Graph Coloring
DF
A
C
B
G
E
DF
A
C
Maximal cliquesD
A
E
DF
A
C
B
G
E
Graph Coloring
DF
A
C
B
G
E
DF
A
C
Maximal cliquesD
A
E
DF
A
C
B
G
E
A B
E
DF
A
C
B
G
E
Graph Coloring
DF
A
C
B
G
E
DF
A
C
Maximal cliquesD
A
E
DF
A
C
B
G
E
A B
E
DF
A
C
B
G
EDF
A
C
B
G
E
B
G
E
Graph Coloring
DF
A
C
B
G
E
DF
A
C
Maximal cliquesD
A
E
DF
A
C
B
G
E
A B
E
DF
A
C
B
G
EDF
A
C
B
G
E
B
G
EDF
A
C
B
G
E
D
G
E
Graph Coloring
DF
A
C
B
G
E
DF
A
C
Maximal cliquesD
A
E
DF
A
C
B
G
E
A B
E
DF
A
C
B
G
EDF
A
C
B
G
E
B
G
EDF
A
C
B
G
E
D
G
E
DF
A
C
B
G
E
Graph Coloring
DF
A
C
B
G
E
DF
A
C
Maximal cliquesD
A
E
DF
A
C
B
G
E
A B
E
DF
A
C
B
G
EDF
A
C
B
G
E
B
G
EDF
A
C
B
G
E
D
G
E
DF
A
C
B
G
EDF
A
C
B
G
E
DF
A
C
D
A
E
A B
EB
G
E
D
G
E
Graph Coloring
DF
A
C
B
G
E
DF
A
C
Maximal cliquesD
A
E
DF
A
C
B
G
E
A B
E
DF
A
C
B
G
EDF
A
C
B
G
E
B
G
EDF
A
C
B
G
E
D
G
E
DF
A
C
B
G
EDF
A
C
B
G
E
DF
A
C
D
A
E
A B
EB
G
E
D
G
E
DF
A
C
B
G
E
Graph Coloring
DF
A
C
B
G
E
DF
A
C
Maximal cliquesD
A
E
DF
A
C
B
G
E
A B
E
DF
A
C
B
G
EDF
A
C
B
G
E
B
G
EDF
A
C
B
G
E
D
G
E
DF
A
C
B
G
EDF
A
C
B
G
E
DF
A
C
D
A
E
A B
EB
G
E
D
G
E
DF
A
C
B
G
E
A B
C D EF
G
Graph Coloring
Sudoku is also a graph coloring problem
Graph Coloring
Sudoku is also a graph coloring problem
4x4 Sudoku example:
A B C D
HGFE
I J K L
PONM
Graph Coloring
Sudoku is also a graph coloring problem
4x4 Sudoku example:
A B C D
HGFE
I J K L
PONM
A B C D
HGFE
I J K L
PONM
rows
Maximal cliques:
Graph Coloring
Sudoku is also a graph coloring problem
4x4 Sudoku example:
A B C D
HGFE
I J K L
PONM
A B C D
HGFE
I J K L
PONM
rows
Maximal cliques:
A B
FE
C D
HG
I J
NM
K L
PO
blocks
Graph Coloring
Sudoku is also a graph coloring problem
4x4 Sudoku example:
A B C D
HGFE
I J K L
PONM
A B C D
HGFE
I J K L
PONM
rows
Maximal cliques:
A B
FE
C D
HG
I J
NM
K L
PO
blocks
A
E
I
M
B
F
J
N
C
G
K
O
D
H
L
P
columns
Graph Coloring
Sudoku is also a graph coloring problem
4x4 Sudoku example:
A B C D
HGFE
I J K L
PONM
A B C D
HGFE
I J K L
PONM
rows
Maximal cliques:
A B
FE
C D
HG
I J
NM
K L
PO
blocks
A
E
I
M
B
F
J
N
C
G
K
O
D
H
L
P
columnsA B C D
HGFE
I J K L
PONM
C D
HG
A B
FE
K L
PO
I J
NM
D
H
L
P
C
G
K
O
B
F
J
N
A
E
I
M
Graph Coloring
Sudoku is also a graph coloring problem
4x4 Sudoku example:
A B C D
HGFE
I J K L
PONM
A B C D
HGFE
I J K L
PONM
rows
Maximal cliques:
A B
FE
C D
HG
I J
NM
K L
PO
blocks
A
E
I
M
B
F
J
N
C
G
K
O
D
H
L
P
columnsA B C D
HGFE
I J K L
PONM
C D
HG
A B
FE
K L
PO
I J
NM
D
H
L
P
C
G
K
O
B
F
J
N
A
E
I
M
1 2 3 4
3 4 1 2
2 3 4 1
4 1 2 3
1 2
3 4
3 4
1 2
2 3
4 1
4 1
2 3
1
3
2
4
2
4
3
1
3
1
4
2
4
2
1
3
Graph Coloring
1 2 3 4
3 4 1 2
2 3 4 1
4 1 2 3
A B C D
HGFE
I J K L
PONM
Probabilistic Graphical ModelsGraph Coloring
1 2 3 4
3 4 1 2
2 3 4 1
4 1 2 3
A B C D
HGFE
I J K L
PONM
1 2 3 4
2143
2 3 4 1
3214
Probabilistic Graphical ModelsGraph Coloring
1 2 3 4
3 4 1 2
2 3 4 1
4 1 2 3
A B C D
HGFE
I J K L
PONM
1 2 3 4
2143
2 3 4 1
3214
But how dowe take allconstraints
into account?
Probabilistic Graphical ModelsGraph Coloring
Probabilistic Graphical Models
PROBABILISTIC GRAPHICAL MODELSIn a general sense, a PGM of a system• clusters information into local sections, and• let the sections communicate about their combined outcome
PROBABILISTIC GRAPHICAL MODELSIn a general sense, a PGM of a system• clusters information into local sections, and• let the sections communicate about their combined outcome
In a probabilistic sense, these "local sections" are• prior distributions,• marginal distributions, and/or• conditional distributions;together, a compact representation of a larger space
PROBABILISTIC GRAPHICAL MODELSIn a general sense, a PGM of a system• clusters information into local sections, and• let the sections communicate about their combined outcome
PROBABILISTIC GRAPHICAL MODELSIn a general sense, a PGM of a system• clusters information into local sections, and• let the sections communicate about their combined outcome
A B
C D EF
G
PROBABILISTIC GRAPHICAL MODELSIn a general sense, a PGM of a system• clusters information into local sections, and• let the sections communicate about their combined outcome
A B
C D EF
G
DF
A
C D
A
E
A B
E
D
G
EB
G
E
PROBABILISTIC GRAPHICAL MODELSIn a general sense, a PGM of a system• clusters information into local sections, and• let the sections communicate about their combined outcome
A B
C D EF
G
DF
A
C D
A
E
A B
E
D
G
EB
G
E
P(A,B,C,D,E,F,G) = f (A,C,D,F) · f (A,D,E) · f (A,B,E) · f (B,E,G) · f (D,E,G)1 2 3 4 5
PROBABILISTIC GRAPHICAL MODELSIn a general sense, a PGM of a system• clusters information into local sections, and• let the sections communicate about their combined outcome
A B
C D EF
G
DF
A
C D
A
E
A B
E
D
G
EB
G
E
P(A,B,C,D,E,F,G) = f (A,C,D,F) · f (A,D,E) · f (A,B,E) · f (B,E,G) · f (D,E,G)1 2 3 4 5
1 2 3 4 11 2 4 3 11 3 2 4 1
4 3 2 1 1non normalized...
P(A,B,C,D)
elsewhere 0
A,C,D,F
P(A,B,C,D,E,F,G) = f (A,C,D,F) · f (A,D,E) · f (A,B,E) · f (B,E,G) · f (D,E,G)1 2 3 4 5
PROBABILISTIC GRAPHICAL MODELSIn a general sense, a PGM of a system• clusters information into local sections, and• let the sections communicate about their combined outcome
A B
C D EF
G
DF
A
C D
A
E
A B
E
D
G
EB
G
E
P(A,B,C,D,E,F,G) = f (A,C,D,F) · f (A,D,E) · f (A,B,E) · f (B,E,G) · f (D,E,G)1 2 3 4 5
1 2 3 4 11 2 4 3 11 3 2 4 1
4 3 2 1 1non normalized...
P(A,B,C,D)
elsewhere 0
A,C,D,F
P(A,B,C,D,E,F,G) = f (A,C,D,F) · f (A,D,E) · f (A,B,E) · f (B,E,G) · f (D,E,G)1 2 3 4 5
1 2 3 11 2 4 11 3 2 1
4 3 2 1non norm...
P(B,E,G)
elsewhere 0
1 2 3 4 11 2 4 3 11 3 2 4 1
4 3 2 1 1non normalized...
P(A,B,C,D)
elsewhere 0
B,E,G
P(A,B,C,D,E,F,G) = f (A,C,D,F) · f (A,D,E) · f (A,B,E) · f (B,E,G) · f (D,E,G)1 2 3 4 5
PROBABILISTIC GRAPHICAL MODELSIn a general sense, a PGM of a system• clusters information into local sections, and• let the sections communicate about their combined outcome
A B
C D EF
G
DF
A
C D
A
E
A B
E
D
G
EB
G
E
P(A,B,C,D,E,F,G) = f (A,C,D,F) · f (A,D,E) · f (A,B,E) · f (B,E,G) · f (D,E,G)1 2 3 4 5
1 2 3 4 11 2 4 3 11 3 2 4 1
4 3 2 1 1non normalized...
P(A,B,C,D)
elsewhere 0
A,C,D,F
P(A,B,C,D,E,F,G) = f (A,C,D,F) · f (A,D,E) · f (A,B,E) · f (B,E,G) · f (D,E,G)1 2 3 4 5
1 2 3 11 2 4 11 3 2 1
4 3 2 1non norm...
P(B,E,G)
elsewhere 0
1 2 3 4 11 2 4 3 11 3 2 4 1
4 3 2 1 1non normalized...
P(A,B,C,D)
elsewhere 0
B,E,G
P(A,B,C,D,E,F,G) = f (A,C,D,F) · f (A,D,E) · f (A,B,E) · f (B,E,G) · f (D,E,G)1 2 3 4 5
A,B,E D,E,GA,D,E
P(A,B,C,D,E,F,G) = f (A,C,D,F) · f (A,D,E) · f (A,B,E) · f (B,E,G) · f (D,E,G)1 2 3 4 5
In a general sense, a PGM of a system• clusters information into local sections, and• let the sections communicate about their combined outcome
A B
C D EF
G
A,C,D,F B,E,GA,B,E D,E,GA,D,E
PROBABILISTIC GRAPHICAL MODELS
In a general sense, a PGM of a system• clusters information into local sections, and• let the sections communicate about their combined outcome
A B
C D EF
G
A,C,D,F B,E,GA,B,E D,E,GA,D,E
Factor/Bethé graph
A,D,EA,B,EA,C,D,F B,E,G E,D,G
A B F GC D E
A B
C
D
E
FGA
B
DE G
A
DE
PROBABILISTIC GRAPHICAL MODELS
In a general sense, a PGM of a system• clusters information into local sections, and• let the sections communicate about their combined outcome
A B
C D EF
G
A,C,D,F B,E,GA,B,E D,E,GA,D,E
Factor/Bethé graph
A,D,EA,B,EA,C,D,F B,E,G E,D,G
A B F GC D E
A B
C
D
E
FGA
B
DE G
A
DE
Cluster graph
A,D E,DA,D,E
A,E
A,B,E B,E
A,C,D,F
B,E,G
E,D,G
G
PROBABILISTIC GRAPHICAL MODELS
In a general sense, a PGM of a system• clusters information into local sections, and• let the sections communicate about their combined outcome
A B
C D EF
G
A,C,D,F B,E,GA,B,E D,E,GA,D,E
Factor/Bethé graph
A,D,EA,B,EA,C,D,F B,E,G E,D,G
A B F GC D E
A B
C
D
E
FGA
B
DE G
A
DE
Cluster graph
A,D E,DA,D,E
A,E
A,B,E B,E
A,C,D,F
B,E,G
E,D,G
G Factor/Bethé graph
A,D,EA,B,EA,C,D,F B,E,G E,D,G
A B F GC D E
A B
C
D
E
FGA
B
DE G
A
DE
Cluster graph
A,D E,DA,D,E
A,E
A,B,E B,E
A,C,D,F
B,E,G
E,D,G
G
PROBABILISTIC GRAPHICAL MODELS
In a general sense, a PGM of a system• clusters information into local sections, and• let the sections communicate about their combined outcome
A B
C D EF
G
A,C,D,F B,E,GA,B,E D,E,GA,D,E
Factor/Bethé graph
A,D,EA,B,EA,C,D,F B,E,G E,D,G
A B F GC D E
A B
C
D
E
FGA
B
DE G
A
DE
Cluster graph
A,D E,DA,D,E
A,E
A,B,E B,E
A,C,D,F
B,E,G
E,D,G
G Factor/Bethé graph
A,D,EA,B,EA,C,D,F B,E,G E,D,G
A B F GC D E
A B
C
D
E
FGA
B
DE G
A
DE
Cluster graph
A,D E,DA,D,E
A,E
A,B,E B,E
A,C,D,F
B,E,G
E,D,G
G
Cluster graph
A,D E,DA,D,E
A,E
A,B,E B,E
A,C,D,F
B,E,G
E,D,G
G
PROBABILISTIC GRAPHICAL MODELS
In a general sense, a PGM of a system• clusters information into local sections, and• let the sections communicate about their combined outcome
A B
C D EF
G
A,C,D,F B,E,GA,B,E D,E,GA,D,E
Factor/Bethé graph
A,D,EA,B,EA,C,D,F B,E,G E,D,G
A B F GC D E
A B
C
D
E
FGA
B
DE G
A
DE
Cluster graph
A,D E,DA,D,E
A,E
A,B,E B,E
A,C,D,F
B,E,G
E,D,G
G Factor/Bethé graph
A,D,EA,B,EA,C,D,F B,E,G E,D,G
A B F GC D E
A B
C
D
E
FGA
B
DE G
A
DE
Cluster graph
A,D E,DA,D,E
A,E
A,B,E B,E
A,C,D,F
B,E,G
E,D,G
G
Cluster graph
A,D E,DA,D,E
A,E
A,B,E B,E
A,C,D,F
B,E,G
E,D,G
G
PROBABILISTIC GRAPHICAL MODELS
In a general sense, a PGM of a system• clusters information into local sections, and• let the sections communicate about their combined outcome
A B
C D EF
G
A,C,D,F B,E,GA,B,E D,E,GA,D,E
Factor/Bethé graph
A,D,EA,B,EA,C,D,F B,E,G E,D,G
A B F GC D E
A B
C
D
E
FGA
B
DE G
A
DE
Cluster graph
A,D E,DA,D,E
A,E
A,B,E B,E
A,C,D,F
B,E,G
E,D,G
G Factor/Bethé graph
A,D,EA,B,EA,C,D,F B,E,G E,D,G
A B F GC D E
A B
C
D
E
FGA
B
DE G
A
DE
Cluster graph
A,D E,DA,D,E
A,E
A,B,E B,E
A,C,D,F
B,E,G
E,D,G
G
Cluster graph
A,D E,DA,D,E
A,E
A,B,E B,E
A,C,D,F
B,E,G
E,D,G
G
PROBABILISTIC GRAPHICAL MODELS
In a general sense, a PGM of a system• clusters information into local sections, and• let the sections communicate about their combined outcome
A B
C D EF
G
A,C,D,F B,E,GA,B,E D,E,GA,D,E
Factor/Bethé graph
A,D,EA,B,EA,C,D,F B,E,G E,D,G
A B F GC D E
A B
C
D
E
FGA
B
DE G
A
DE
Cluster graph
A,D E,DA,D,E
A,E
A,B,E B,E
A,C,D,F
B,E,G
E,D,G
G Factor/Bethé graph
A,D,EA,B,EA,C,D,F B,E,G E,D,G
A B F GC D E
A B
C
D
E
FGA
B
DE G
A
DE
Cluster graph
A,D E,DA,D,E
A,E
A,B,E B,E
A,C,D,F
B,E,G
E,D,G
G
Cluster graph
A,D E,DA,D,E
A,E
A,B,E B,E
A,C,D,F
B,E,G
E,D,G
G
PROBABILISTIC GRAPHICAL MODELS
In a general sense, a PGM of a system• clusters information into local sections, and• let the sections communicate about their combined outcome
A B
C D EF
G
A,C,D,F B,E,GA,B,E D,E,GA,D,E
Factor/Bethé graph
A,D,EA,B,EA,C,D,F B,E,G E,D,G
A B F GC D E
A B
C
D
E
FGA
B
DE G
A
DE
Cluster graph
A,D E,DA,D,E
A,E
A,B,E B,E
A,C,D,F
B,E,G
E,D,G
G Factor/Bethé graph
A,D,EA,B,EA,C,D,F B,E,G E,D,G
A B F GC D E
A B
C
D
E
FGA
B
DE G
A
DE
Cluster graph
A,D E,DA,D,E
A,E
A,B,E B,E
A,C,D,F
B,E,G
E,D,G
G
Cluster graph
A,D E,DA,D,E
A,E
A,B,E B,E
A,C,D,F
B,E,G
E,D,G
G
PROBABILISTIC GRAPHICAL MODELS
A B
C D EF
G
PROBABILISTIC GRAPHICAL MODELS
A B
C D EF
G
PROBABILISTIC GRAPHICAL MODELS
A B
C D EF
G Factor/Bethé graph
A,D,EA,B,EA,C,D,F B,E,G E,D,G
A B F GC D E
A B
C
D
E
FGA
B
DE G
A
DE
A B
C D EF
G
A B
C D EF
G
PROBABILISTIC GRAPHICAL MODELS
A B
C D EF
G Factor/Bethé graph
A,D,EA,B,EA,C,D,F B,E,G E,D,G
A B F GC D E
A B
C
D
E
FGA
B
DE G
A
DE
A B
C D EF
G
A B
C D EF
G
Cluster graph
A,D E,DA,D,E
A,E
A,B,E B,E
A,C,D,F
B,E,G
E,D,G
G
A B
C D EF
G
A B
C D EF
G
PROBABILISTIC GRAPHICAL MODELS
CLUSTER GRAPHS
CLUSTER GRAPHSWe found that• Graph structure influence convergence speed and accuracy• Factor graphs are predominant in PGM literature• Cluster graphs outperform factor graphs
CLUSTER GRAPHSWe found that• Graph structure influence convergence speed and accuracy• Factor graphs are predominant in PGM literature• Cluster graphs outperform factor graphs
Why is cluster graphs the underdog?• Multiple solutions for the same clusters• Absence of a generic construction procedure
CLUSTER GRAPHSWe found that• Graph structure influence convergence speed and accuracy• Factor graphs are predominant in PGM literature• Cluster graphs outperform factor graphs
Why is cluster graphs the underdog?• Multiple solutions for the same clusters• Absence of a generic construction procedure
We propose the LTRIP procedure as a solution
B,C,D,E,FA,B,C,DA,B,G B,C,GB,E,F
Factor/Bethé graph:
A B C D E F G variables
clusters
CLUSTER GRAPHS
B,C,D,E,FA,B,C,DA,B,G B,C,GB,E,F
Factor/Bethé graph:
A B C D E F G variables
clusters
connection layer GA B C D E F
CLUSTER GRAPHS
B,C,D,E,FA,B,C,DA,B,G B,C,GB,E,F
Factor/Bethé graph:
A B C D E F G variables
clusters
connection layer GA B C D E F
*running intersection property
sepsets
A
A A
CLUSTER GRAPHS
B,C,D,E,FA,B,C,DA,B,G B,C,GB,E,F
Factor/Bethé graph:
A B C D E F G variables
clusters
connection layer GA B C D E F
*running intersection property
sepsets
B
B BB B B
CLUSTER GRAPHS
B,C,D,E,FA,B,C,DA,B,G B,C,GB,E,F
Factor/Bethé graph:
A B C D E F G variables
clusters
connection layer GA B C D E F
*running intersection property
sepsets C C C
C
CLUSTER GRAPHS
B,C,D,E,FA,B,C,DA,B,G B,C,GB,E,F
Factor/Bethé graph:
A B C D E F G variables
clusters
connection layer GA B C D E F
*running intersection property
sepsets
A B C D E F G
G
AB
C
E FA
B
B
BB C CD
D EF G
CLUSTER GRAPHS
B,C,D,E,FA,B,C,DA,B,G B,C,GB,E,F
Cluster graph:
A B C D E F G variables
clusters
connection layer GA B C D E F
*running intersection property
CLUSTER GRAPHS
B,C,D,E,FA,B,C,DA,B,G B,C,GB,E,F
Cluster graph:
A B C D E F G variables
clusters
connection layer GA B C D E F
*running intersection property
CLUSTER GRAPHS
*running intersection property
multivar. sepsets
LTRIP procedure:
A B C D E F G variables
clusters
B,E,F
B,C,D,E,F
A,B,C,D
A,B,GB,C,G
CLUSTER GRAPHS
*running intersection property
multivar. sepsets
LTRIP procedure:
A B C D E F G variables
clusters
A
B,E,F
B,C,D,E,F
A,B,C,D
A,B,GB,C,GA
CLUSTER GRAPHS
*running intersection property
multivar. sepsets
LTRIP procedure:
A B C D E F G variables
clusters
B,E,F
B,C,D,E,F
A,B,C,D
A,B,GB,C,G
B
BBA
B
B
CLUSTER GRAPHS
*running intersection property
multivar. sepsets
LTRIP procedure:
A B C D E F G variables
clusters
C
B,E,F
B,C,D,E,F
A,B,C,D
A,B,GB,C,G
B,C
BB,CA
B
CLUSTER GRAPHS
*running intersection property
multivar. sepsets
LTRIP procedure:
A B C D E F G variables
clusters
D
B,E,F
B,C,D,E,F
A,B,C,D
A,B,GB,C,G
B,C,D
BB,CA
B
CLUSTER GRAPHS
*running intersection property
multivar. sepsets
LTRIP procedure:
A B C D E F G variables
clusters
E
B,E,F
B,C,D,E,F
A,B,C,D
A,B,GB,C,G
B,C,D
BB,CA
B,E
CLUSTER GRAPHS
*running intersection property
multivar. sepsets
LTRIP procedure:
A B C D E F G variables
clusters
F
B,E,F
B,C,D,E,F
A,B,C,D
A,B,GB,C,G
B,C,D
BB,CA
B,E,F
CLUSTER GRAPHS
*running intersection property
multivar. sepsets
LTRIP procedure:
A B C D E F G variables
clusters
G
B,E,F
B,C,D,E,F
A,B,C,D
A,B,GB,C,G
B,C,D
BB,C
GA
B,E,F
CLUSTER GRAPHS
*running intersection property
multivar. sepsets
LTRIP procedure:
A B C D E F G variables
clusters
B,E,F
B,C,D,E,F
A,B,C,D
A,B,GB,C,G
B,C,D
BB,C
GA
B,E,F
CLUSTER GRAPHS
A B C D
HGFE
I J K L
PONM
Sudoku cluster graph:
CLUSTER GRAPHS
A B C D
HGFE
I J K L
PONM
Sudoku cluster graph:
A B C D
HGFE
I J K L
PONM
A B
FE
C D
HG
I J
NM
K L
PO
A
E
I
M
B
F
J
N
C
G
K
O
D
H
L
P
CLUSTER GRAPHS
A B C D
HGFE
I J K L
PONM
Sudoku cluster graph:
A,B,C,D
E,F,G,H
A,E,I,M
B,F,J,N
A,B
E,F
A,E
B,F
C,D
G,H
I,M
J,N
A,B,E,F
C,D,G,H
I,J,M,N
C,G
D,H
I,J
M,N
K,L,O,P
K,O
L,P
K,L
O,P
I,J,K,L
C,G,K,O
D,H,L,P
M,N,O,P
CLUSTER GRAPHS
A
A,B,C,D E,F,G,H A,E,I,M B,F,J,N A,B,E,F C,D,G,H K,L,O,PI,J,K,L C,G,K,O D,H,L,PM,N,O,P I,J,M,N
A E I MB F J NC G K OD H L PE I
M
FJ
N
CG
K
OD
H
L
P
A
B
E
I
M
F
J
N
C G K
O
DH
L
PA
B
E
I
MF
J
N
CG
K
O
DH
LP
B
Sudoku factor/Bethé graph:
CLUSTER GRAPHS
CLUSTER GRAPHSRESULTS
A comparison of cluster graphs vs. factor graphs on PGMs build from Sudoku puzzles
Datasets used:• Project Euler @ projecteuler.net/problem=96• Sterten's 95 hardest Sudokus @ magictour.free.fr/top95
Tested with different cluster sizes by splitting-up clusters
CLUSTER GRAPHSRESULTS
A B C D E F G H I
9x9 Sudoku puzzle
...J K L M N O P Q R
CLUSTER GRAPHSRESULTS
A B C D E F G H I
9x9 Sudoku puzzle
...J K L M N O P Q R
A B C D E F G H I
...
J K L M N O P Q R
CLUSTER GRAPHSRESULTS
A B C D E F G H I
Split each 9 variable clique
C B A
D E A
F G A
H I A
H I B
F B G
D E B
C B D
F E C
H C G
D I C
D G H
D E F
F E G
H E I
F I H
F G H
H I G
into cliques of 3 variables
CLUSTER GRAPHSRESULTS
A B C D E F G H I
Split each 9 variable clique into cliques of 5 variables
E D B A C
F I G A H
F B I G H
E D F B C
D I G H C
E D F G H
E F I G H
CLUSTER GRAPHSRESULTS
A B C D E F G H I
Split each 9 variable clique
E D F B G A C
E F B I G A H
E D F B G H C
E D F I G H C
into cliques of 7 variables
CLUSTER GRAPHSRESULTS
A B C D E F G H I
Split each 9 variable clique
A B C D E F G H I
into cliques of 9 variables
CLUSTER GRAPHSRESULTS
A B C D E F G H I
Split each 9 variable clique
A B C D E F G H I
into cliques of 9 variables
A,B,C,D
E,F,G,H
A,E,I,M
B,F,J,N
A,B
E,F
A,E
B,F
C,D
G,H
I,M
J,N
A,B,E,F
C,D,G,H
I,J,M,N
C,G
D,H
I,J
M,N
K,L,O,P
K,O
L,P
K,L
O,P
I,J,K,L
C,G,K,O
D,H,L,P
M,N,O,P
A
A,B,C,D E,F,G,H A,E,I,M B,F,J,N A,B,E,F C,D,G,H K,L,O,PI,J,K,L C,G,K,O D,H,L,PM,N,O,P I,J,M,N
A E I MB F J NC G K OD H L PE I
M
FJ
N
CG
K
OD
H
L
P
A
B
E
I
M
F
J
N
C G K
O
DH
L
PA
B
E
I
MF
J
N
CG
K
O
DH
LP
B
Build these clusters into both a factor graph and a cluster graph
CLUSTER GRAPHSRESULTS
A B C D E F G H I
Split each 9 variable clique
A B C D E F G H I
into cliques of 9 variables
A,B,C,D
E,F,G,H
A,E,I,M
B,F,J,N
A,B
E,F
A,E
B,F
C,D
G,H
I,M
J,N
A,B,E,F
C,D,G,H
I,J,M,N
C,G
D,H
I,J
M,N
K,L,O,P
K,O
L,P
K,L
O,P
I,J,K,L
C,G,K,O
D,H,L,P
M,N,O,P
A
A,B,C,D E,F,G,H A,E,I,M B,F,J,N A,B,E,F C,D,G,H K,L,O,PI,J,K,L C,G,K,O D,H,L,PM,N,O,P I,J,M,N
A E I MB F J NC G K OD H L PE I
M
FJ
N
CG
K
OD
H
L
P
A
B
E
I
M
F
J
N
C G K
O
DH
L
PA
B
E
I
MF
J
N
CG
K
O
DH
LP
B
Build these clusters into both a factor graph and a cluster graph
Run message passing on both graphs(note, for a solved Suduko all clusters is reduced to a single entry)
P(A,B,C,D,E,F,G,H,I)A1
B3
C2
D6
E4
F5
G7
H9
I8
elsewhere1
0
CLUSTER GRAPHSRESULTS
Project Euler: solution count
36 0
131
36
14
1337
factor graph
clus
ter g
raph
Size 3
CLUSTER GRAPHSRESULTS
Project Euler: solution count
36 0
131
36
14
1337
factor graph
clus
ter g
raph
Size 3
0
CLUSTER GRAPHSRESULTS
Project Euler: solution count
36 0
131
36
14
1337
factor graph
clus
ter g
raph
Size 3
0
cluster graph slower
CLUSTER GRAPHSRESULTS
Project Euler: solution count
36 0
131
36
14
1337
factor graph
clus
ter g
raph
Size 3
0
cluster graph slower
25 0
232
25
25
2327
factor graphcl
uste
r gra
phSize 5
CLUSTER GRAPHSRESULTS
Project Euler: solution count
36 0
131
36
14
1337
factor graph
clus
ter g
raph
Size 3
0
cluster graph slower
25 0
232
25
25
2327
factor graphcl
uste
r gra
phSize 5
0
CLUSTER GRAPHSRESULTS
Project Euler: solution count
36 0
131
36
14
1337
factor graph
clus
ter g
raph
Size 3
0
cluster graph slower
25 0
232
25
25
2327
factor graphcl
uste
r gra
phSize 5
0
cluster graph slower
CLUSTER GRAPHSRESULTS
Project Euler: solution count
36 0
131
36
14
1337
factor graph
clus
ter g
raph
Size 3
0
cluster graph slower
25 0
232
25
25
2327
factor graphcl
uste
r gra
phSize 5
0
cluster graph slower
5 0
414
5
45
419
0 0
464
0
50
464
cluster graph faster cluster graph faster
factor graph factor graph
clus
ter g
raph
clus
ter g
raph
Size 7 Size 9
0 0
CLUSTER GRAPHSRESULTS
Project Euler: solution count
36 0
131
36
14
1337
factor graph
clus
ter g
raph
Size 3
0
cluster graph slower
25 0
232
25
25
2327
factor graphcl
uste
r gra
phSize 5
0
cluster graph slower
5 0
414
5
45
419
0 0
464
0
50
464
cluster graph faster cluster graph faster
factor graph factor graph
clus
ter g
raph
clus
ter g
raph
Size 7 Size 9
0 0
Cluster graphs more successful than factor graphsNaive solver to test graph structures - can improve!
CLUSTER GRAPHSRESULTS
CONCLUSION• Main contribution is LTRIP for constructing cluster graphs• These cluster graphs show great promise over factor graphs• We hope LTRIP will enhance the popularity of cluster graphs
CONCLUSION• Main contribution is LTRIP for constructing cluster graphs• These cluster graphs show great promise over factor graphs• We hope LTRIP will enhance the popularity of cluster graphs
FUTURE WORK• Investigate more advance techniques for graph coloring PGMs• Mutual information approach for LTRIP's max spanning trees• Investigate cluster graphs on wider set of problems