Skip to content
GitLab
Projects
Groups
Snippets
Help
Loading...
Help
Help
Support
Community forum
Keyboard shortcuts
?
Submit feedback
Contribute to GitLab
Sign in / Register
Toggle navigation
Open sidebar
Tiago Peixoto
graphtool
Commits
c0b1794f
Commit
c0b1794f
authored
May 29, 2020
by
Tiago Peixoto
Browse files
inference: update cookbook and docstrings
parent
100f01d4
Changes
11
Hide whitespace changes
Inline
Sidebyside
Showing
11 changed files
with
291 additions
and
212 deletions
+291
212
doc/demos/inference/_minimization.rst
doc/demos/inference/_minimization.rst
+1
1
doc/demos/inference/_model_class_selection.rst
doc/demos/inference/_model_class_selection.rst
+74
137
doc/demos/inference/_model_selection.rst
doc/demos/inference/_model_selection.rst
+8
8
doc/demos/inference/_reconstruction.rst
doc/demos/inference/_reconstruction.rst
+24
16
doc/demos/inference/_reconstruction_dynamics.rst
doc/demos/inference/_reconstruction_dynamics.rst
+8
5
doc/demos/inference/_sampling.rst
doc/demos/inference/_sampling.rst
+117
32
doc/demos/inference/inference.rst
doc/demos/inference/inference.rst
+3
0
src/graph/inference/partition_modes/graph_partition_mode.cc
src/graph/inference/partition_modes/graph_partition_mode.cc
+1
0
src/graph/inference/partition_modes/graph_partition_mode.hh
src/graph/inference/partition_modes/graph_partition_mode.hh
+5
0
src/graph_tool/inference/partition_centroid.py
src/graph_tool/inference/partition_centroid.py
+2
5
src/graph_tool/inference/partition_modes.py
src/graph_tool/inference/partition_modes.py
+48
8
No files found.
doc/demos/inference/_minimization.rst
View file @
c0b1794f
...
...
@@ 34,7 +34,7 @@ which yields
<Graph object, undirected, with 115 vertices and 613 edges, 4 internal vertex properties, 2 internal graph properties, at 0x...>
w
e then fit the degreecorrected model by calling
W
e then fit the degreecorrected model by calling
:
.. testcode:: football
...
...
doc/demos/inference/_model_class_selection.rst
View file @
c0b1794f
...
...
@@ 52,129 +52,65 @@ The computation of the posterior entropy :math:`\mathcal{S}`, however,
is
significantly
more
difficult
,
since
it
involves
measuring
the
precise
value
of
:
math
:`
q
(\
boldsymbol
b
)`.
A
direct
"brute force"
computation
of
:
math
:`\
mathcal
{
S
}`
is
implemented
via
:meth:`~graph_tool.inference.blockmodel.BlockState.collect_partition_histogram` and
:func:`~graph_tool.inference.blockmodel.microstate_entropy`, however this is only
feasible for very small networks. For larger networks, we are forced to
perform approximations. The simplest is a "mean field" one, where we
assume the posterior factorizes as
.. math::
q(\boldsymbol b) \approx \prod_i{q_i(b_i)}
where
.. math::
q_i(r) = P(b_i = r  \boldsymbol A)
is the marginal group membership distribution of node :math:`i`. This
yields an entropy value given by
.. math::
S \approx \sum_i\sum_rq_i(r)\ln q_i(r).
This approximation should be seen as an upper bound, since any existing
correlation between the nodes (which are ignored here) will yield
smaller entropy values.
A more accurate assumption is called the `Bethe approximation`
[mezardinformation2009]_, and takes into account the correlation
between adjacent nodes in the network,
.. math::
q(\boldsymbol b) \approx \prod_{i<j}q_{ij}(b_i,b_j)^{A_{ij}}\prod_iq_i(b_i)^{1k_i}
where :math:`A_{ij}` is the `adjacency matrix
<https://en.wikipedia.org/wiki/Adjacency_matrix>`_, :math:`k_i` is the
degree of node :math:`i`, and
.. math::
q_{ij}(r, s) = P(b_i = r, b_j = s\boldsymbol A)
is the joint group membership distribution of nodes :math:`i` and
:math:`j` (a.k.a. the `edge marginals`). This yields an entropy value
given by
.. math::
S \approx \sum_{i<j}A_{ij}\sum_{rs}q_{ij}(r,s)\ln q_{ij}(r,s)  \sum_i(1k_i)\sum_rq_i(r)\ln q_i(r).
Typically, this approximation yields smaller values than the mean field
one, and is generally considered to be superior. However, formally, it
depends on the graph being sufficiently locally "treelike", and the
posterior being indeed strongly correlated with the adjacency matrix
itself  two characteristics which do not hold in general. Although
the approximation often gives reasonable results even when these
conditions do not strictly hold, in some situations when they are
strongly violated this approach can yield meaningless values, such as a
negative entropy. Therefore, it is useful to compare both approaches
whenever possible.
With these approximations, it possible to estimate the full model
evidence efficiently, as we show below, using
:meth:`~graph_tool.inference.blockmodel.BlockState.collect_vertex_marginals`,
:meth:`~graph_tool.inference.blockmodel.BlockState.collect_edge_marginals`,
:meth:`~graph_tool.inference.blockmodel.mf_entropy` and
:meth:`~graph_tool.inference.blockmodel.bethe_entropy`.
:
meth
:`~
graph_tool
.
inference
.
blockmodel
.
BlockState
.
collect_partition_histogram
`
and
:
func
:`~
graph_tool
.
inference
.
blockmodel
.
microstate_entropy
`,
however
this
is
only
feasible
for
very
small
networks
.
For
larger
networks
,
we
are
forced
to
perform
approximations
.
One
possibility
is
to
employ
the
method
described
in
[
peixoto

revealing

2020
]
_
,
based
on
fitting
a
mixture
"random label"
model
to
the
posterior
distribution
,
which
allows
us
to
compute
its
entropy
.
In
graph

tool
this
is
done
by
using
:
class
:`~
graph_tool
.
inference
.
partition_modes
.
PartitionModeState
`,
as
we
show
in
the
example
below
.
..
testcode
::
model

evidence
from
scipy
.
special
import
gammaln
g
=
gt
.
collection
.
data
[
"lesmis"
]
for
deg_corr
in
[
True
,
False
]:
state
=
gt
.
minimize_blockmodel_dl
(
g
,
deg_corr
=
deg_corr
)
#
Initialize
the
Markov
#
chain
from
the
"ground
# state"
state = state.copy(B=g.num_vertices())
dls
=
[]
#
description
length
history
vm = None # vertex marginals
em = None # edge marginals
bs
=
[]
#
partitions
def collect_marginals(s):
global vm, em
vm = s.collect_vertex_marginals(vm)
em = s.collect_edge_marginals(em)
def
collect_partitions
(
s
):
global
bs
,
dls
bs
.
append
(
s
.
get_state
().
a
.
copy
())
dls
.
append
(
s
.
entropy
())
# Now we collect the marginal distributions for exactly 200,000 sweeps
gt.mcmc_equilibrate(state, force_niter=20000, mcmc_args=dict(niter=10),
callback=collect_marginals)
#
Now
we
collect
2000
partitions
;
but
the
larger
this
is
,
the
#
more
accurate
will
be
the
calculation
gt
.
mcmc_equilibrate
(
state
,
force_niter
=
2000
,
mcmc_args
=
dict
(
niter
=
10
),
callback
=
collect_partitions
)
S_mf = gt.mf_entropy(g, vm)
S_bethe = gt.bethe_entropy(g, em)[0]
L = mean(dls)
#
Infer
partition
modes
pmode
=
gt
.
ModeClusterState
(
bs
)
print("Model evidence for deg_corr = %s:" % deg_corr,
L + S_mf, "(mean field),", L + S_bethe, "(Bethe)"
)
#
Minimize
the
mode
state
itself
gt
.
mcmc_equilibrate
(
pmode
,
wait
=
1
,
mcmc_args
=
dict
(
niter
=
1
,
beta
=
np
.
inf
)
)
.. testoutput:: modelevidence
#
Posterior
entropy
H
=
pmode
.
posterior_entropy
()
Model evidence for deg_corr = True: 383.787042... (mean field), 1053.691486... (Bethe)
Model evidence for deg_corr = False: 363.454006... (mean field), 948.144381... (Bethe
)
#
log
(
B
!) term
logB
=
mean
(
gammaln
(
np
.
array
([
len
(
np
.
unique
(
b
))
for
b
in
bs
])
+
1
)
)
If we consider the more accurate approximation, the outcome shows a
preference for the nondegreecorrected model.
#
Evidence
L
=

mean
(
dls
)
+
logB
+
H
print
(
f
"Model logevidence for deg_corr = {deg_corr}: {L}"
)
When using the nested model, the approach is entirely analogous. The
only difference now is that we have a hierarchical partition
:math:`\{\boldsymbol b_l\}` in the equations above, instead of simply
:math:`\boldsymbol b`. In order to make the approach tractable, we
assume the factorization
..
testoutput
::
model

evidence
.. math::
Model
log

evidence
for
deg_corr
=
True
:

678.785176
...
Model
log

evidence
for
deg_corr
=
False
:

672.643870
...
q(\{\boldsymbol b_l\}) \approx \prod_lq_l(\boldsymbol b_l)
The
outcome
shows
a
preference
for
the
non

degree

corrected
model
.
where :math:`q_l(\boldsymbol b_l)` is the marginal posterior for the
partition at level :math:`l`. For :math:`q_0(\boldsymbol b_0)` we may
use again either the meanfield or Bethe approximations, however for
:math:`l>0` only the meanfield approximation is applicable, since the
adjacency matrix of the higher layers is not constant. We show below the
When
using
the
nested
model
,
the
approach
is
entirely
analogous
.
We
show
below
the
approach
for
the
same
network
,
using
the
nested
model
.
...
...
@@ 182,49 +118,50 @@ approach for the same network, using the nested model.
g
=
gt
.
collection
.
data
[
"lesmis"
]
nL = 10
for
deg_corr
in
[
True
,
False
]:
state = gt.minimize_nested_blockmodel_dl(g, deg_corr=deg_corr) # Initialize the Markov
# chain from the "ground
# state"
bs = state.get_bs() # Get hierarchical partition.
bs += [np.zeros(1)] * (nL  len(bs)) # Augment it to L = 10 with
# singlegroup levels.
state = state.copy(bs=bs, sampling=True)
dls = [] # description length history
vm = [None] * len(state.get_levels()) # vertex marginals
em = None # edge marginals
def collect_marginals(s):
global vm, em
levels = s.get_levels()
vm = [sl.collect_vertex_marginals(vm[l]) for l, sl in enumerate(levels)]
em = levels[0].collect_edge_marginals(em)
state
=
gt
.
NestedBlockState
(
g
,
deg_corr
=
deg_corr
)
#
Equilibrate
gt
.
mcmc_equilibrate
(
state
,
force_niter
=
1000
,
mcmc_args
=
dict
(
niter
=
10
))
dls
=
[]
#
description
length
history
bs
=
[]
#
partitions
def
collect_partitions
(
s
):
global
bs
,
dls
bs
.
append
(
s
.
get_state
())
dls
.
append
(
s
.
entropy
())
# Now we collect the marginal distributions for exactly 200,000 sweeps
gt.mcmc_equilibrate(state, force_niter=20000, mcmc_args=dict(niter=10),
callback=collect_marginals)
#
Now
we
collect
2000
partitions
;
but
the
larger
this
is
,
the
#
more
accurate
will
be
the
calculation
gt
.
mcmc_equilibrate
(
state
,
force_niter
=
2000
,
mcmc_args
=
dict
(
niter
=
10
),
callback
=
collect_partitions
)
#
Infer
partition
modes
pmode
=
gt
.
ModeClusterState
(
bs
,
nested
=
True
)
#
Minimize
the
mode
state
itself
gt
.
mcmc_equilibrate
(
pmode
,
wait
=
1
,
mcmc_args
=
dict
(
niter
=
1
,
beta
=
np
.
inf
))
S_mf = [gt.mf_entropy(sl.g, vm[l]) for l, sl in enumerate(state.get_levels())]
S_bethe = gt.bethe_entropy(g, em)[0]
L = mean(dls)
#
Posterior
entropy
H
=
pmode
.
posterior_entropy
()
print("Model evidence for deg_corr = %s:" % deg_corr,
L + sum(S_mf), "(mean field),", L + S_bethe + sum(S_mf[1:]), "(Bethe)"
)
#
log
(
B
!) term
logB
=
mean
([
sum
(
gammaln
(
len
(
np
.
unique
(
bl
))+
1
)
for
bl
in
b
)
for
b
in
bs
]
)
#
Evidence
L
=

mean
(
dls
)
+
logB
+
H
print
(
f
"Model logevidence for deg_corr = {deg_corr}: {L}"
)
..
testoutput
::
model

evidence
Model evidence for deg_corr = True: 
183.999497... (mean field), 840.108164... (Bethe)
Model evidence for deg_corr = False: 
187.344952... (mean field), 738.797625... (Bethe)
Model
log

evidence
for
deg_corr
=
True
:

660.331060
...
Model
log

evidence
for
deg_corr
=
False
:

657.839574
...
The results are similar: If we consider the most accurate approximation,
the nondegreecorrected model possesses the largest evidence. Note also
that we observe a better evidence for the nested models themselves, when
comparing to the evidences for the nonnested model  which is not
quite surprising, since the nonnested model is a special case of the
nested one.
The
results
are
similar
:
The
non

degree

corrected
model
possesses
the
largest
evidence
.
Note
also
that
we
observe
a
better
evidence
for
the
nested
models
themselves
,
when
comparing
to
the
evidences
for
the
non

nested
model

which
is
not
quite
surprising
,
since
the
non

nested
model
is
a
special
case
of
the
nested
one
.
doc/demos/inference/_model_selection.rst
View file @
c0b1794f
...
...
@@ 25,8 +25,8 @@ we have
.. testoutput:: modelselection
:options: +NORMALIZE_WHITESPACE
Nondegreecorrected DL:
85
20.825480
...
Degreecorrected DL:
82
27.987410
...
Nondegreecorrected DL: 85
53.474528
...
Degreecorrected DL: 82
66.554118
...
Since it yields the smallest description length, the degreecorrected
fit should be preferred. The statistical significance of the choice can
...
...
@@ 52,7 +52,7 @@ fits. In our particular case, we have
.. testoutput:: modelselection
:options: +NORMALIZE_WHITESPACE
ln Λ: 2
92.83807
0...
ln Λ: 2
86.92041
0...
The precise threshold that should be used to decide when to `reject a
hypothesis <https://en.wikipedia.org/wiki/Hypothesis_testing>`_ is
...
...
@@ 80,11 +80,11 @@ example, for the American football network above, we have:
.. testoutput:: modelselection
:options: +NORMALIZE_WHITESPACE
Nondegreecorrected DL: 17
57.843826
...
Degreecorrected DL: 180
9.86199
6...
ln Λ:
52.018170
...
Nondegreecorrected DL: 17
38.138494
...
Degreecorrected DL: 1
7
80
.57671
6...
ln Λ:
42.438221
...
Hence, with a posterior odds ratio of :math:`\Lambda \approx \mathrm{e}^{
5
2} \approx
10^{
22
}` in favor of the nondegreecorrected model, we conclude that the
Hence, with a posterior odds ratio of :math:`\Lambda \approx \mathrm{e}^{
4
2} \approx
10^{
19
}` in favor of the nondegreecorrected model, we conclude that the
degreecorrected variant is an unnecessarily complex description for
this network.
doc/demos/inference/_reconstruction.rst
View file @
c0b1794f
...
...
@@ 164,15 +164,14 @@ simple example, using
# intervals of 10 sweeps:
u = None # marginal posterior edge probabilities
pv
=
None
#
m
ar
ginal posterior group membership probabilitie
s
bs
=
[]
#
p
ar
tition
s
cs = [] # average local clustering coefficient
def collect_marginals(s):
global
pv, u
, cs
global
u, bs
, cs
u = s.collect_marginal(u)
bstate = s.get_block_state()
b = gt.perfect_prop_hash([bstate.levels[0].b])[0]
pv = bstate.levels[0].collect_vertex_marginals(pv, b=b)
bs.append(bstate.levels[0].b.a.copy())
cs.append(gt.local_clustering(s.get_graph()).fa.mean())
gt.mcmc_equilibrate(state, force_niter=10000, mcmc_args=dict(niter=10),
...
...
@@ 223,7 +222,10 @@ reconstructed network:
bstate = state.get_block_state()
bstate = bstate.levels[0].copy(g=u)
pv = u.own_property(pv)
# Disambiguate partitions and obtain marginals
pmode = gt.PartitionModeState(bs, converge=True)
pv = pmode.get_marginal(u)
edash = u.new_ep("vector<double>")
edash[u.edge(15, 73)] = [.1, .1, 0]
bstate.draw(pos=u.own_property(g.vp.pos), vertex_shape="pie", vertex_pie_fractions=pv,
...
...
@@ 293,7 +295,7 @@ with uniform error rates, as we see with the same example:
# intervals of 10 sweeps:
u = None # marginal posterior edge probabilities
pv
=
None
#
m
ar
ginal posterior group membership probabilitie
s
bs
=
[]
#
p
ar
tition
s
cs = [] # average local clustering coefficient
gt.mcmc_equilibrate(state, force_niter=10000, mcmc_args=dict(niter=10),
...
...
@@ 412,15 +414,14 @@ inference:
# intervals of 10 sweeps:
u = None # marginal posterior edge probabilities
pv
=
None
#
m
ar
ginal posterior group membership probabilitie
s
bs
=
[]
#
p
ar
tition
s
cs = [] # average local clustering coefficient
def collect_marginals(s):
global
pv
, u, cs
global
bs
, u, cs
u = s.collect_marginal(u)
bstate = s.get_block_state()
b = gt.perfect_prop_hash([bstate.levels[0].b])[0]
pv = bstate.levels[0].collect_vertex_marginals(pv, b=b)
bs.append(bstate.levels[0].b.a.copy())
cs.append(gt.local_clustering(s.get_graph()).fa.mean())
gt.mcmc_equilibrate(state, force_niter=10000, mcmc_args=dict(niter=10),
...
...
@@ 465,7 +466,11 @@ the same measurement probability. The reconstructed network is visualized below.
bstate = state.get_block_state()
bstate = bstate.levels[0].copy(g=u)
pv = u.own_property(pv)
# Disambiguate partitions and obtain marginals
pmode = gt.PartitionModeState(bs, converge=True)
pv = pmode.get_marginal(u)
bstate.draw(pos=u.own_property(g.vp.pos), vertex_shape="pie", vertex_pie_fractions=pv,
edge_color=ecolor, edge_dash_style=edash, edge_gradient=None,
output="lesmisuncertainreconstructionmarginals.svg")
...
...
@@ 516,14 +521,13 @@ latent multiedges of a network of political books:
# intervals of 10 sweeps:
u = None # marginal posterior multigraph
pv
=
None
#
m
ar
ginal posterior group membership probabilitie
s
bs
=
[]
#
p
ar
tition
s
def collect_marginals(s):
global
pv
, u
global
bs
, u
u = s.collect_marginal_multigraph(u)
bstate = state.get_block_state()
b = gt.perfect_prop_hash([bstate.levels[0].b])[0]
pv = bstate.levels[0].collect_vertex_marginals(pv, b=b)
bs.append(bstate.levels[0].b.a.copy())
gt.mcmc_equilibrate(state, force_niter=10000, mcmc_args=dict(niter=10),
callback=collect_marginals)
...
...
@@ 538,7 +542,11 @@ latent multiedges of a network of political books:
bstate = state.get_block_state()
bstate = bstate.levels[0].copy(g=u)
pv = u.own_property(pv)
# Disambiguate partitions and obtain marginals
pmode = gt.PartitionModeState(bs, converge=True)
pv = pmode.get_marginal(u)
bstate.draw(pos=u.own_property(g.vp.pos), vertex_shape="pie", vertex_pie_fractions=pv,
edge_pen_width=gt.prop_to_size(ew, .1, 8, power=1), edge_gradient=None,
output="polbookserasedpoisson.svg")
...
...
doc/demos/inference/_reconstruction_dynamics.rst
View file @
c0b1794f
...
...
@@ 117,14 +117,13 @@ epidemic process.
# intervals of 10 sweeps:
gm = None
b
m
=
None
b
s
=
[]
betas = []
def collect_marginals(s):
global gm, b
m
global gm, b
s
gm = s.collect_marginal(gm)
b = gt.perfect_prop_hash([s.bstate.b])[0]
bm = s.bstate.collect_vertex_marginals(bm, b=b)
bs.append(s.bstate.b.a.copy())
betas.append(s.params["global_beta"])
gt.mcmc_equilibrate(rstate, force_niter=10000, mcmc_args=dict(niter=10, xstep=0),
...
...
@@ 132,9 +131,13 @@ epidemic process.
print("Posterior similarity: ", gt.similarity(g, gm, g.new_ep("double", 1), gm.ep.eprob))
print("Inferred infection probability: %g ± %g" % (mean(betas), std(betas)))
# Disambiguate partitions and obtain marginals
pmode = gt.PartitionModeState(bs, converge=True)
pv = pmode.get_marginal(gm)
gt.graph_draw(gm, gm.own_property(g.vp.pos), vertex_shape="pie", vertex_color="black",
vertex_pie_fractions=
gm.own_property(bm)
, vertex_pen_width=1,
vertex_pie_fractions=
pv
, vertex_pen_width=1,
edge_pen_width=gt.prop_to_size(gm.ep.eprob, 0, 5),
eorder=gm.ep.eprob, output="dolphinsposterior.svg")
...
...
doc/demos/inference/_sampling.rst
View file @
c0b1794f
...
...
@@ 81,37 +81,41 @@ Note that the value of ``wait`` above was made purposefully low so that
the output would not be overly long. The most appropriate value requires
experimentation, but a typically good value is ``wait=1000``.
The function :func:`~graph_tool.inference.mcmc.mcmc_equilibrate` accepts
a
``callback`` argument that takes an optional function to be invoked
The function :func:`~graph_tool.inference.mcmc.mcmc_equilibrate` accepts
a
``callback`` argument that takes an optional function to be invoked
after each call to
:meth:`~graph_tool.inference.blockmodel.BlockState.multiflip_mcmc_sweep`. This function
should accept a single parameter which will contain the actual
:class:`~graph_tool.inference.blockmodel.BlockState` instance. We will use this in
the example below to collect the posterior vertex marginals (via
:class:`~graph_tool.inference.blockmodel.BlockState.collect_vertex_marginals`),
i.e. the posterior probability that a node belongs to a given group:
:meth:`~graph_tool.inference.blockmodel.BlockState.multiflip_mcmc_sweep`. This
function should accept a single parameter which will contain the actual
:class:`~graph_tool.inference.blockmodel.BlockState` instance. We will
use this in the example below to collect the posterior vertex marginals
(via :class:`~graph_tool.inference.partition_modes.PartitionModeState`,
which disambiguates group labels [peixotorevealing2020]_), i.e. the
posterior probability that a node belongs to a given group:
.. testcode:: modelaveraging
# We will first equilibrate the Markov chain
gt.mcmc_equilibrate(state, wait=1000, mcmc_args=dict(niter=10))
pv
=
None
bs
=
[] # collect some partitions
def collect_marginals(s):
global pv
b = gt.perfect_prop_hash([s.b])[0]
pv = s.collect_vertex_marginals(pv, b=b)
def collect_partitions(s):
global bs
bs.append(s.b.a.copy())
# Now we collect
the marginal
s for exactly 100,000 sweeps, at
#
intervals
of 10 sweeps:
# Now we collect
partition
s for exactly 100,000 sweeps, at
intervals
# of 10 sweeps:
gt.mcmc_equilibrate(state, force_niter=10000, mcmc_args=dict(niter=10),
callback=collect_
m
ar
ginal
s)
callback=collect_
p
ar
tition
s)
# Disambiguate partitions and obtain marginals
pmode = gt.PartitionModeState(bs, converge=True)
pv = pmode.get_marginal(g)
# Now the node marginals are stored in property map pv. We can
# visualize them as pie charts on the nodes:
state.draw(pos=g.vp.pos, vertex_shape="pie", vertex_pie_fractions=pv,
edge_gradient=None,
output="lesmissbmmarginals.svg")
output="lesmissbmmarginals.svg")
.. figure:: lesmissbmmarginals.*
:align: center
...
...
@@ 135,8 +139,8 @@ itself, as follows.
B = s.get_nonempty_B()
h[B] += 1
# Now we collect
the marginal
s for exactly 100,000 sweeps, at
#
intervals
of 10 sweeps:
# Now we collect
partition
s for exactly 100,000 sweeps, at
intervals
# of 10 sweeps:
gt.mcmc_equilibrate(state, force_niter=10000, mcmc_args=dict(niter=10),
callback=collect_num_groups)
...
...
@@ 194,7 +198,6 @@ network as above.
Change in description length: 73.716766...
Number of accepted vertex moves: 366160
.. warning::
When using
...
...
@@ 212,28 +215,34 @@ Similarly to the the nonnested case, we can use
:func:`~graph_tool.inference.mcmc.mcmc_equilibrate` to do most of the boring
work, and we can now obtain vertex marginals on all hierarchical levels:
.. testcode:: nestedmodelaveraging
# We will first equilibrate the Markov chain
gt.mcmc_equilibrate(state, wait=1000, mcmc_args=dict(niter=10))
pv = [None] * len(state.get_levels())
# collect nested partitions
bs = []
def collect_marginals(s):
global pv
bs = [gt.perfect_prop_hash([s.b])[0] for s in state.get_levels()]
pv = [s.collect_vertex_marginals(pv[l], b=bs[l]) for l, s in enumerate(s.get_levels())]
def collect_partitions(s):
global bs
bs.append(s.get_bs())
# Now we collect the marginals for exactly 100,000 sweeps
gt.mcmc_equilibrate(state, force_niter=10000, mcmc_args=dict(niter=10),
callback=collect_marginals)
callback=collect_partitions)
# Disambiguate partitions and obtain marginals
pmode = gt.PartitionModeState(bs, nested=True, converge=True)
pv = pmode.get_marginal(g)
# Get consensus estimate
bs = pmode.get_max_nested()
state = state.copy(bs=bs)
# Now the node marginals for all levels are stored in property map
# list pv. We can visualize the first level as pie charts on the nodes:
state_0 = state.get_levels()[0]
state_0.draw(pos=g.vp.pos, vertex_shape="pie", vertex_pie_fractions=pv[0],
edge_gradient=None, output="lesmisnestedsbmmarginals.svg")
# We can visualize the marginals as pie charts on the nodes:
state.draw(vertex_shape="pie", vertex_pie_fractions=pv,
output="lesmisnestedsbmmarginals.svg")
.. figure:: lesmisnestedsbmmarginals.*
:align: center
...
...
@@ 316,3 +325,79 @@ distribution.
:width: 200px
.. image:: lesmispartitionsample9.svg
:width: 200px
Characterizing the posterior distribution
+++++++++++++++++++++++++++++++++++++++++
The posterior distribution of partitions can have an elaborate
structure, containing multiple possible explanations for the data. In
order to summarize it, we can infer the modes of the distribution using
:class:`~graph_tool.inference.partition_modes.ModeClusterState`, as
described in [peixotorevealing2020]_. This amounts to identifying
clusters of partitions that are very similar to each other, but
sufficiently different from those that belong to other
clusters. Collective, such "modes" represent the different stories that
the data is telling us through the model. Here is an example using again
the Les Misérables network:
.. testcode:: partitionmodes
g = gt.collection.data["lesmis"]
state = gt.NestedBlockState(g)
# Equilibration
gt.mcmc_equilibrate(state, force_niter=1000, mcmc_args=dict(niter=10))
bs = []
def collect_partitions(s):
global bs
bs.append(s.get_bs())
# We will collect only partitions 1000 partitions. For more accurate
# results, this number should be increased.
gt.mcmc_equilibrate(state, force_niter=1000, mcmc_args=dict(niter=10),
callback=collect_partitions)
# Infer partition modes
pmode = gt.ModeClusterState(bs, nested=True)
# Minimize the mode state itself
gt.mcmc_equilibrate(pmode, wait=1, mcmc_args=dict(niter=1, beta=np.inf))
# Get inferred modes
modes = pmode.get_modes()
for i, mode in enumerate(modes):
b = mode.get_max_nested() # mode's maximum
pv = mode.get_marginal(g) # mode's marginal distribution
print(f"Mode {i} with size {mode.get_M()/len(bs)}")
state = state.copy(bs=b)
state.draw(vertex_shape="pie", vertex_pie_fractions=pv,
output="lesmispartitionmode%i.svg" % i)
Running the above code gives us the relative size of each mode,
corresponding to their collective posterior probability.
.. testoutput:: partitionmodes
Mode 0 with size 0.389389...
Mode 1 with size 0.352352...
Mode 2 with size 0.129129...