blockmodel.py 88.7 KB
Newer Older
1
2
3
4
5
#! /usr/bin/env python
# -*- coding: utf-8 -*-
#
# graph_tool -- a general graph manipulation python module
#
Tiago Peixoto's avatar
Tiago Peixoto committed
6
# Copyright (C) 2006-2014 Tiago de Paula Peixoto <tiago@skewed.de>
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
#
# This program is free software: you can redistribute it and/or modify
# it under the terms of the GNU General Public License as published by
# the Free Software Foundation, either version 3 of the License, or
# (at your option) any later version.
#
# This program is distributed in the hope that it will be useful,
# but WITHOUT ANY WARRANTY; without even the implied warranty of
# MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
# GNU General Public License for more details.
#
# You should have received a copy of the GNU General Public License
# along with this program.  If not, see <http://www.gnu.org/licenses/>.

from __future__ import division, absolute_import, print_function
import sys
if sys.version_info < (3,):
    range = xrange

26
27
from .. import _degree, _prop, Graph, GraphView, libcore, _get_rng, PropertyMap
from .. stats import label_self_loops
28
29
import random
from numpy import *
30
import numpy
31
32
from scipy.optimize import fsolve, fminbound
import scipy.special
33
from collections import defaultdict
34
35
import copy
import heapq
36
37
38
39
40

from .. dl_import import dl_import
dl_import("from . import libgraph_tool_community as libcommunity")


41
42
43
44
45
46
47
48
49
50
51
52
def get_block_graph(g, B, b, vcount, ecount):
    cg, br, vcount, ecount = condensation_graph(g, b,
                                                vweight=vcount,
                                                eweight=ecount,
                                                self_loops=True)[:4]
    cg.vp["count"] = vcount
    cg.ep["count"] = ecount
    cg = Graph(cg, vorder=br)

    cg.add_vertex(B - cg.num_vertices())
    return cg

53
54
55
56
57
58
59
60
class BlockState(object):
    r"""This class encapsulates the block state of a given graph.

    This must be instantiated and used by functions such as :func:`mcmc_sweep`.

    Parameters
    ----------
    g : :class:`~graph_tool.Graph`
61
        Graph to be modelled.
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
    eweight : :class:`~graph_tool.PropertyMap` (optional, default: ``None``)
        Edge weights (i.e. multiplicity).
    vweight : :class:`~graph_tool.PropertyMap` (optional, default: ``None``)
        Vertex weights (i.e. multiplicity).
    b : :class:`~graph_tool.PropertyMap` (optional, default: ``None``)
        Initial block labels on the vertices. If not supplied, it will be
        randomly sampled.
    B : ``int`` (optional, default: ``None``)
        Number of blocks. If not supplied it will be either obtained from the
        parameter ``b``, or set to the maximum possible value according to the
        minimum description length.
    clabel : :class:`~graph_tool.PropertyMap` (optional, default: ``None``)
        This parameter provides a constraint label, such that vertices with
        different labels will not be allowed to belong to the same block. If not given,
        all labels will be assumed to be the same.
    deg_corr : ``bool`` (optional, default: ``True``)
        If ``True``, the degree-corrected version of the blockmodel ensemble will
        be assumed, otherwise the traditional variant will be used.
80
81
82
83
    max_BE : ``int`` (optional, default: ``1000``)
        If the number of blocks exceeds this number, a sparse representation of
        the block graph is used, which is slightly less efficient, but uses less
        memory,
84
85
    """

86
87
    def __init__(self, g, eweight=None, vweight=None, b=None,
                 B=None, clabel=None, deg_corr=True, max_BE=1000):
88
89
90
91
92
93
94
95
96
97
98
99
100
101
        self.g = g
        if eweight is None:
            eweight = g.new_edge_property("int")
            eweight.a = 1
        elif eweight.value_type() != "int32_t":
            eweight = eweight.copy(value_type="int32_t")
        if vweight is None:
            vweight = g.new_vertex_property("int")
            vweight.a = 1
        elif vweight.value_type() != "int32_t":
            vweight = vweight.copy(value_type="int32_t")
        self.eweight = eweight
        self.vweight = vweight

102
103
        self.E = int(self.eweight.fa.sum())
        self.N = int(self.vweight.fa.sum())
104
105
106
107
108
109

        self.deg_corr = deg_corr

        if b is None:
            if B is None:
                B = get_max_B(self.N, self.E, directed=g.is_directed())
110
111
112
            ba = random.randint(0, B, g.num_vertices())
            ba[:B] = arange(B)        # avoid empty blocks
            random.shuffle(ba)
113
            b = g.new_vertex_property("int")
114
            b.fa = ba
115
116
117
            self.b = b
        else:
            if B is None:
118
                B = int(b.fa.max()) + 1
119
120
            self.b = b = b.copy(value_type="int32_t")

121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
        if b.fa.max() >= B:
            raise ValueError("Maximum value of b is larger or equal to B!")

        # Construct block-graph
        self.bg = get_block_graph(g, B, b, vweight, eweight)
        self.bg.set_fast_edge_removal()

        self.mrs = self.bg.ep["count"]
        self.wr = self.bg.vp["count"]
        del self.bg.ep["count"]
        del self.bg.vp["count"]

        self.mrp = self.bg.degree_property_map("out", weight=self.mrs)

        if g.is_directed():
            self.mrm = self.bg.degree_property_map("in", weight=self.mrs)
        else:
            self.mrm = self.mrp
139
140
141

        self.vertices = libcommunity.get_vector(B)
        self.vertices.a = arange(B)
142
        self.B = B
143

144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
        self.clabel = clabel
        if self.clabel is None:
            self.clabel = self.g.new_vertex_property("int")

        self.bclabel = self.bg.new_vertex_property("int")
        libcommunity.vector_rmap(self.b.a, self.bclabel.a)
        libcommunity.vector_map(self.bclabel.a, self.clabel.a)

        self.emat = None
        if max_BE is None:
            max_BE = 1000
        self.max_BE = max_BE

        # used by mcmc_sweep()
        self.egroups = None
        self.nsampler = None
        self.sweep_vertices = None

        # used by merge_sweep()
        self.bnsampler = None
164
        self.bnnsampler = None
165

166
167
168
        libcommunity.init_safelog(int(2 * max(self.E, self.N)))
        libcommunity.init_xlogx(int(2 * max(self.E, self.N)))
        libcommunity.init_lgamma(int(3 * max(self.E, self.N)))
169

170
171
172
173
    def __get_emat(self):
        if self.emat is None:
            self.__regen_emat()
        return self.emat
174
175

    def __regen_emat(self):
176
177
178
179
        if self.B <= self.max_BE:
            self.emat = libcommunity.create_emat(self.bg._Graph__graph)
        else:
            self.emat = libcommunity.create_ehash(self.bg._Graph__graph)
180

181
    def __build_egroups(self, empty=False):
182
183
184
        self.esrcpos = self.g.new_edge_property("int")
        self.etgtpos = self.g.new_edge_property("int")
        self.is_weighted = True if self.eweight.fa.max() > 1 else False
185
        self.egroups = libcommunity.build_egroups(self.g._Graph__graph,
186
187
188
189
190
                                                  self.bg._Graph__graph,
                                                  _prop("v", self.g, self.b),
                                                  _prop("e", self.g, self.eweight),
                                                  _prop("e", self.g, self.esrcpos),
                                                  _prop("e", self.g, self.etgtpos),
191
                                                  self.is_weighted, empty)
192
193
194

    def __build_nsampler(self):
        self.nsampler = libcommunity.init_neighbour_sampler(self.g._Graph__graph,
195
196
                                                            _prop("e", self.g, self.eweight),
                                                            True)
197
198
    def __build_bnsampler(self):
        self.bnsampler = libcommunity.init_neighbour_sampler(self.bg._Graph__graph,
199
200
201
202
203
                                                             _prop("e", self.bg, self.mrs),
                                                             False)
        self.bnnsampler = libcommunity.init_neighbour_sampler(self.bg._Graph__graph,
                                                              _prop("e", self.bg, self.mrs),
                                                              True)
204
205
206
207
208
209

    def __cleanup_bg(self):
        emask = self.bg.new_edge_property("bool")
        emask.a = self.mrs.a[:len(emask.a)] > 0
        self.bg.set_edge_filter(emask)
        self.bg.purge_edges()
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228

    def get_blocks(self):
        r"""Returns the property map which contains the block labels for each vertex."""
        return self.b

    def get_bg(self):
        r"""Returns the block graph."""
        return self.bg

    def get_ers(self):
        r"""Returns the edge property map of the block graph which contains the :math:`e_{rs}` matrix entries."""
        return self.mrs

    def get_er(self):
        r"""Returns the vertex property map of the block graph which contains the number
        :math:`e_r` of half-edges incident on block :math:`r`. If the graph is
        directed, a pair of property maps is returned, with the number of
        out-edges :math:`e^+_r` and in-edges :math:`e^-_r`, respectively."""
        if self.bg.is_directed():
229
            return self.mrp, self.mrm
230
231
232
233
234
235
236
237
238
239
240
241
242
243
        else:
            return self.mrp

    def get_nr(self):
        r"""Returns the vertex property map of the block graph which contains the block sizes :math:`n_r`."""
        return self.wr

    def get_eweight(self):
        r"""Returns the block edge counts associated with the block matrix
        :math:`e_{rs}`. For directed graphs it is identical to :math:`e_{rs}`,
        but for undirected graphs it is identical except for the diagonal, which
        is :math:`e_{rr}/2`."""
        eweight = self.mrs.copy()
        if not self.g.is_directed():
244
245
            sl = label_self_loops(self.bg, mark_only=True)
            eweight.a[sl.a > 0] /= 2
246
247
        return eweight

248
249
    def entropy(self, complete=False, random=False, dl=False, dense=False,
                multigraph=False):
250
251
252
253
254
255
256
257
258
259
260
261
        r"""Calculate the entropy per edge associated with the current block partition.

        Parameters
        ----------
        complete : ``bool`` (optional, default: ``False``)
            If ``True``, the complete entropy will be returned, including constant
            terms not relevant to the block partition.
        random : ``bool`` (optional, default: ``False``)
            If ``True``, the entropy entropy corresponding to an equivalent random
            graph (i.e. no block partition) will be returned.
        dl : ``bool`` (optional, default: ``False``)
            If ``True``, the full description length will be returned.
262
263
264
265
266
        dense : ``bool`` (optional, default: ``False``)
            If ``True``, the "dense" variant of the entropy will be computed.
        multigraph : ``bool`` (optional, default: ``False``)
            If ``True``, the multigraph entropy will be used. Only has an effect
            if ``dense == True``.
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306

        Notes
        -----

        For the traditional blockmodel (``deg_corr == False``), the entropy is
        given by

        .. math::

          \mathcal{S}_t &\cong E - \frac{1}{2} \sum_{rs}e_{rs}\ln\left(\frac{e_{rs}}{n_rn_s}\right), \\
          \mathcal{S}^d_t &\cong E - \sum_{rs}e_{rs}\ln\left(\frac{e_{rs}}{n_rn_s}\right),

        for undirected and directed graphs, respectively, where :math:`e_{rs}`
        is the number of edges from block :math:`r` to :math:`s` (or the number
        of half-edges for the undirected case when :math:`r=s`), and :math:`n_r`
        is the number of vertices in block :math:`r` .


        For the degree-corrected variant with "hard" degree constraints the
        equivalent expressions are

        .. math::

            \mathcal{S}_c &\cong -E -\sum_kN_k\ln k! - \frac{1}{2} \sum_{rs}e_{rs}\ln\left(\frac{e_{rs}}{e_re_s}\right), \\
            \mathcal{S}^d_c &\cong -E -\sum_{k^+}N_{k^+}\ln k^+!  -\sum_{k^-}N_{k^-}\ln k^-! - \sum_{rs}e_{rs}\ln\left(\frac{e_{rs}}{e^+_re^-_s}\right),

        where :math:`e_r = \sum_se_{rs}` is the number of half-edges incident on
        block :math:`r`, and :math:`e^+_r = \sum_se_{rs}` and :math:`e^-_r =
        \sum_se_{sr}` are the number of out- and in-edges adjacent to block
        :math:`r`, respectively.

        If ``complete == False`` only the last term of the equations above will
        be returned. If ``random == True`` it will be assumed that :math:`B=1`
        despite the actual :math:`e_{rs}` matrix.  If ``dl == True``, the
        description length :math:`\mathcal{L}_t` of the model will be returned
        as well, as described in :func:`model_entropy`. Note that for the
        degree-corrected version the description length is

        .. math::

307
            \mathcal{L}_c = \mathcal{L}_t - \sum_rn_r\sum_kp^r_k\ln p^r_k,
308

309
310
        where :math:`p^r_k` is the fraction of nodes in block $r$ with degree :math:`k`. For directed
        graphs we have instead :math:`k \to (k^-, k^+)`.
311

312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
        If the "dense" entropies are requested, they will be computed as

        .. math::

            \mathcal{S}_t  &= \sum_{r>s} \ln{\textstyle {n_rn_s \choose e_{rs}}} + \sum_r \ln{\textstyle {{n_r\choose 2}\choose e_{rr}/2}}\\
            \mathcal{S}^d_t  &= \sum_{rs} \ln{\textstyle {n_rn_s \choose e_{rs}}},

        for simple graphs, and

        .. math::

            \mathcal{S}_m  &= \sum_{r>s} \ln{\textstyle \left(\!\!{n_rn_s \choose e_{rs}}\!\!\right)} + \sum_r \ln{\textstyle \left(\!\!{\left(\!{n_r\choose 2}\!\right)\choose e_{rr}/2}\!\!\right)}\\
            \mathcal{S}^d_m  &= \sum_{rs} \ln{\textstyle \left(\!\!{n_rn_s \choose e_{rs}}\!\!\right)},

        for multigraphs (i.e. ``multigraph == True``).

        Note that in all cases the value returned corresponds to the entropy `per edge`,
329
330
331
332
333
334
335
        i.e. :math:`(\mathcal{S}_{t/c}\; [\,+\, \mathcal{L}_{t/c}])/ E`.

        """

        E = self.E
        N = self.N

336
337
338
339
340
341
342
343
344
        if dense:
            if random:
                bg = get_block_graph(self.bg, 1,
                                     self.bg.new_vertex_property("int"),
                                     self.wr, self.mrs)
                S = libcommunity.entropy_dense(bg._Graph__graph,
                                               _prop("e", bg, bg.ep["count"]),
                                               _prop("v", bg, bg.vp["count"]),
                                               multigraph)
345
            else:
346
347
348
349
                S = libcommunity.entropy_dense(self.bg._Graph__graph,
                                               _prop("e", self.bg, self.mrs),
                                               _prop("v", self.bg, self.wr),
                                               multigraph)
350
        else:
351
352
353
354
355
356
357
358
            if self.deg_corr:
                if self.g.is_directed():
                    S_rand = E * log(E)
                else:
                    S_rand = E * log(2 * E)
            else:
                ak = E / float(N) if self.g.is_directed() else  2 * E / float(N)
                S_rand = E * log (N / ak)
359

360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
            if random:
                S = S_rand
            else:
                S = libcommunity.entropy(self.bg._Graph__graph,
                                         _prop("e", self.bg, self.mrs),
                                         _prop("v", self.bg, self.mrp),
                                         _prop("v", self.bg, self.mrm),
                                         _prop("v", self.bg, self.wr),
                                         self.deg_corr)

            if complete:
                if self.deg_corr:
                    S -= E
                    for v in self.g.vertices():
                        S -= scipy.special.gammaln(v.out_degree() + 1)
                        if self.g.is_directed():
                            S -= scipy.special.gammaln(v.in_degree() + 1)
                else:
                    S += E
            else:
                S -= S_rand
381

382
383
384
385
386
387
        if dl:
            if random:
                S += model_entropy(1, N, E, directed=self.g.is_directed()) * E
            else:
                S += model_entropy(self.B, N, E, directed=self.g.is_directed(), nr=self.wr.a) * E

388
            if self.deg_corr:
389
390
391
                S_seq = libcommunity.deg_entropy(self.g._Graph__graph,
                                                 _prop("v", self.g, self.b),
                                                 self.B)
392
393
394
395
396
                S += S_seq

        return S / E

    def remove_vertex(self, v):
397
        r"""Remove vertex ``v`` from its current block."""
398
399
400
401
402
403
404
405
        libcommunity.remove_vertex(self.g._Graph__graph,
                                   self.bg._Graph__graph,
                                   int(v),
                                   _prop("e", self.bg, self.mrs),
                                   _prop("v", self.bg, self.mrp),
                                   _prop("v", self.bg, self.mrm),
                                   _prop("v", self.bg, self.wr),
                                   _prop("v", self.g, self.b))
406
407
408
        self.egroups = None
        self.nb_list = None
        self.nb_count = None
409
410
411


    def add_vertex(self, v, r):
412
        r"""Add vertex ``v`` to block ``r``."""
413
414
415
416
417
418
419
420
        libcommunity.add_vertex(v.get_graph()._Graph__graph,
                                self.bg._Graph__graph,
                                int(v), int(r),
                                _prop("e", self.bg, self.mrs),
                                _prop("v", self.bg, self.mrp),
                                _prop("v", self.bg, self.mrm),
                                _prop("v", self.bg, self.wr),
                                _prop("v", self.g, self.b))
421
422
423
        self.egroups = None
        self.nb_list = None
        self.nb_count = None
424
425

    def move_vertex(self, v, nr):
426
        r"""Move vertex ``v`` to block ``nr``, and return the entropy difference."""
427
428
        dS = libcommunity.move_vertex(self.g._Graph__graph,
                                      self.bg._Graph__graph,
429
                                      self.__get_emat(),
430
431
432
433
434
435
436
437
438
                                      int(v), int(nr),
                                      _prop("e", self.bg, self.mrs),
                                      _prop("v", self.bg, self.mrp),
                                      _prop("v", self.bg, self.mrm),
                                      _prop("v", self.bg, self.wr),
                                      _prop("v", self.g, self.b),
                                      self.deg_corr,
                                      _prop("e", self.bg, self.eweight),
                                      _prop("v", self.bg, self.vweight))
439
440
441
        self.egroups = None
        self.nb_list = None
        self.nb_count = None
442
443
        return dS / float(self.E)

444
    def get_matrix(self, reorder=False, niter=0, ret_order=False):
445
446
        r"""Returns the block matrix, which contains the number of edges between
        each block pair.
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492

        Parameters
        ----------
        reorder : ``bool`` (optional, default: ``False``)
            If ``True``, the matrix is reordered so that blocks which are
            'similar' are close together.
        niter : ``int`` (optional, default: `0`)
            Number of iterations performed to obtain the best ordering. If
            ``niter == 0`` it will automatically determined. Only has effect
            if ``reorder == True``.
        ret_order : ``bool`` (optional, default: ``False``)
            If ``True``, the vertex ordering is returned. Only has effect if
            ``reorder == True``.

        Examples
        --------

        .. testsetup:: get_matrix

           gt.seed_rng(42)
           np.random.seed(42)
           from pylab import *

        .. doctest:: get_matrix

           >>> g = gt.collection.data["polbooks"]
           >>> state = gt.BlockState(g, B=5, deg_corr=True)
           >>> for i in range(1000):
           ...     ds, nmoves = gt.mcmc_sweep(state)
           >>> m = state.get_matrix(reorder=True)
           >>> figure()
           <...>
           >>> matshow(m)
           <...>
           >>> savefig("bloc_mat.pdf")

        .. testcleanup:: get_matrix

           savefig("bloc_mat.png")

        .. figure:: bloc_mat.*
           :align: center

           A  5x5 block matrix.

       """
493
        B = self.B
494
495
496
497
498
499
        vmap = {}
        for r in range(len(self.vertices)):
            vmap[self.vertices[r]] = r

        if reorder:
            if niter == 0:
500
                niter = 10
501
502
503
504
505

            states = []

            label = None
            states = [self]
506
            Bi = self.B // 2
507
508

            while Bi > 1:
509
510
511
512
513

                state = BlockState(states[-1].bg,
                                   b=states[-1].bg.vertex_index.copy("int"),
                                   B=states[-1].bg.num_vertices(),
                                   clabel=states[-1].bclabel,
514
                                   vweight=states[-1].wr,
515
516
517
518
519
520
                                   eweight=states[-1].mrs,
                                   deg_corr=states[-1].deg_corr,
                                   max_BE=states[-1].max_BE)

                state = greedy_shrink(state, B=Bi, nsweeps=niter,
                                      epsilon=1e-3, c=0,
521
                                      nmerge_sweeps=niter)
522
523

                for i in range(niter):
524
                    mcmc_sweep(state, c=0, beta=float("inf"))
525
526
527

                states.append(state)

528
                Bi //= 2
529

530
                if Bi < self.bclabel.a.max() + 1:
531
532
                    break

533
            vorder = list(range(len(states[-1].vertices)))
534
            for state in reversed(states[1:]):
535
                norder = [[] for i in range(state.B)]
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
                for v in state.g.vertices():
                    pos = vorder.index(state.b[v])
                    norder[pos].append(int(v))
                vorder = [item for sublist in norder for item in sublist]
        else:
            vorder = self.vertices

        order_map = zeros(B, dtype="int")
        for i, v in enumerate(vorder):
            order_map[vmap[v]] = i

        m = zeros((B, B))
        rmap = {}
        for e in self.bg.edges():
            r, s = vmap[int(e.source())], vmap[int(e.target())]
            r = order_map[r]
            s = order_map[s]
            rmap[r] = int(e.source())
            rmap[s] = int(e.target())
            m[r, s] = self.mrs[e]
            if not self.bg.is_directed():
                m[s, r] = m[r, s]

559
560
561
562
        for r in range(B):
            if r not in rmap:
                rmap[r] = r

563
564
565
566
567
568
        if ret_order:
            return m, rmap
        else:
            return m


569
def model_entropy(B, N, E, directed=False, nr=None):
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
    r"""Computes the amount of information necessary for the parameters of the traditional blockmodel ensemble, for ``B`` blocks, ``N`` vertices, ``E`` edges, and either a directed or undirected graph.

    A traditional blockmodel is defined as a set of :math:`N` vertices which can
    belong to one of :math:`B` blocks, and the matrix :math:`e_{rs}` describes
    the number of edges from block :math:`r` to :math:`s` (or twice that number
    if :math:`r=s` and the graph is undirected).

    For an undirected graph, the number of distinct :math:`e_{rs}` matrices is given by,

    .. math::

       \Omega_m = \left(\!\!{\left(\!{B \choose 2}\!\right) \choose E}\!\!\right)

    and for a directed graph,

    .. math::
       \Omega_m = \left(\!\!{B^2 \choose E}\!\!\right)


    where :math:`\left(\!{n \choose k}\!\right) = {n+k-1\choose k}` is the
    number of :math:`k` combinations with repetitions from a set of size :math:`n`.

    The total information necessary to describe the model is then,

    .. math::

596
597
       \mathcal{L}_t = \ln\Omega_m + \ln\left(\!\!{B \choose N}\!\!\right) + \ln N! - \sum_r \ln n_r!,

598

599
600
    where the remaining term is the information necessary to describe the
    block partition, where :math:`n_r` is the number of nodes in block :math:`r`.
601

602
603
604
605
    If ``nr`` is ``None``, it is assumed :math:`n_r=N/B`.

    The value returned corresponds to the information per edge, i.e.
    :math:`\mathcal{L}_t/E`.
606
607
608
609

    References
    ----------

Tiago Peixoto's avatar
Tiago Peixoto committed
610
611
612
    .. [peixoto-parsimonious-2013] Tiago P. Peixoto, "Parsimonious module inference in large networks",
       Phys. Rev. Lett. 110, 148701 (2013), :doi:`10.1103/PhysRevLett.110.148701`, :arxiv:`1212.4794`.
    .. [peixoto-hierarchical-2013] Tiago P. Peixoto, "Hierarchical block structures and high-resolution
613
       model selection in large networks ", :arxiv:`1310.4377`.
614
615
616

    """

617
618
619
620
621
622
    if directed:
        x = (B * B);
    else:
        x = (B * (B + 1)) / 2;
    L = lbinom(x + E - 1, E) + partition_entropy(B, N, nr)
    return L / E
623
624
625
626

def Sdl(B, S, N, E, directed=False):
    return S + model_entropy(B, N, E, directed)

627
628
629
630
631
632
633
def lbinom(n, k):
    return scipy.special.gammaln(n + 1) - scipy.special.gammaln(n - k + 1) - scipy.special.gammaln(k + 1)

def partition_entropy(B, N, nr=None):
    if nr is None:
        S = N * log(B) + log1p(-(1 - 1./B) ** N)
    else:
634
        S = lbinom(B + N - 1, N) + scipy.special.gammaln(N + 1) - scipy.special.gammaln(nr + 1).sum()
635
    return S
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655

def get_max_B(N, E, directed=False):
    r"""Return the maximum detectable number of blocks, obtained by minimizing:

    .. math::

        \mathcal{L}_t(B, N, E) - E\ln B

    where :math:`\mathcal{L}_t(B, N, E)` is the information necessary to
    describe a traditional blockmodel with `B` blocks, `N` nodes and `E`
    edges (see :func:`model_entropy`).

    Examples
    --------

    >>> gt.get_max_B(N=1e6, E=5e6)
    1572

    References
    ----------
Tiago Peixoto's avatar
Tiago Peixoto committed
656
657
    .. [peixoto-parsimonious-2013] Tiago P. Peixoto, "Parsimonious module inference in large networks",
       Phys. Rev. Lett. 110, 148701 (2013), :doi:`10.1103/PhysRevLett.110.148701`, :arxiv:`1212.4794`.
658
659
660
661
662
663
664
665
666
667
668


    """

    B = fminbound(lambda B: Sdl(B, -log(B), N, E, directed), 1, E,
                  xtol=1e-6, maxfun=1500, disp=0)
    if isnan(B):
        B = 1
    return max(int(ceil(B)), 2)

def get_akc(B, I, N=float("inf"), directed=False):
Tiago Peixoto's avatar
Tiago Peixoto committed
669
    r"""Return the minimum value of the average degree of the network, so that some block structure with :math:`B` blocks can be detected, according to the minimum description length criterion.
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702

    This is obtained by solving

    .. math::

       \Sigma_b = \mathcal{L}_t(B, N, E) - E\mathcal{I}_{t/c} = 0,

    where :math:`\mathcal{L}_{t}` is the necessary information to describe the
    blockmodel parameters (see :func:`model_entropy`), and
    :math:`\mathcal{I}_{t/c}` characterizes the planted block structure, and is
    given by

    .. math::

        \mathcal{I}_t &= \sum_{rs}m_{rs}\ln\left(\frac{m_{rs}}{w_rw_s}\right),\\
        \mathcal{I}_c &= \sum_{rs}m_{rs}\ln\left(\frac{m_{rs}}{m_rm_s}\right),

    where :math:`m_{rs} = e_{rs}/2E` (or :math:`m_{rs} = e_{rs}/E` for directed
    graphs) and :math:`w_r=n_r/N`. We note that :math:`\mathcal{I}_{t/c}\in[0,
    \ln B]`. In the case where :math:`E \gg B^2`, this simplifies to

    .. math::

       \left<k\right>_c &= \frac{2\ln B}{\mathcal{I}_{t/c}},\\
       \left<k^{-/+}\right>_c &= \frac{\ln B}{\mathcal{I}_{t/c}},

    for undirected and directed graphs, respectively. This limit is assumed if
    ``N == inf``.

    Examples
    --------

    >>> gt.get_akc(10, log(10) / 100, N=100)
703
    2.414413200430159
704
705
706

    References
    ----------
Tiago Peixoto's avatar
Tiago Peixoto committed
707
708
    .. [peixoto-parsimonious-2013] Tiago P. Peixoto, "Parsimonious module inference in large networks",
       Phys. Rev. Lett. 110, 148701 (2013), :doi:`10.1103/PhysRevLett.110.148701`, :arxiv:`1212.4794`.
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723

    """
    if N != float("inf"):
        if directed:
            get_dl = lambda ak: model_entropy(B, N, N * ak, directed) - N * ak * I
        else:
            get_dl = lambda ak: model_entropy(B, N, N * ak / 2., directed) - N * ak * I / 2.
        ak = fsolve(lambda ak: get_dl(ak), 10)
        ak = float(ak)
    else:
        ak = 2 * log(B) / S
        if directed:
            ak /= 2
    return ak

724
725
def mcmc_sweep(state, beta=1., c=1., dense=False, multigraph=False,
               sequential=True, vertices=None, verbose=False):
726
    r"""Performs a Markov chain Monte Carlo sweep on the network, to sample the block partition according to a probability :math:`\propto e^{-\beta \mathcal{S}_{t/c}}`, where :math:`\mathcal{S}_{t/c}` is the blockmodel entropy.
727
728
729
730
731

    Parameters
    ----------
    state : :class:`~graph_tool.community.BlockState`
        The block state.
732
    beta : ``float`` (optional, default: `1.0`)
733
        The inverse temperature parameter :math:`\beta`.
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
    c : ``float`` (optional, default: ``1.0``)
        This parameter specifies how often fully random moves are attempted,
        instead of more likely moves based on the inferred block partition.
        For ``c == 0``, no fully random moves are attempted, and for ``c == inf``
        they are always attempted.
    dense : ``bool`` (optional, default: ``False``)
        If ``True``, the "dense" variant of the entropy will be computed.
    multigraph : ``bool`` (optional, default: ``False``)
        If ``True``, the multigraph entropy will be used. Only has an effect
        if ``dense == True``.
    sequential : ``bool`` (optional, default: ``True``)
        If ``True``, the move attempts on the vertices are done in sequential
        random order. Otherwise a total of `N` moves attempts are made, where
        `N` is the number of vertices, where each vertex can be selected with
        equal probability.
Tiago Peixoto's avatar
Tiago Peixoto committed
749
    vertices : ``list of ints`` (optional, default: ``None``)
750
751
        A list of vertices which will be attempted to be moved. If ``None``, all
        vertices will be attempted.
752
753
754
755
756
757
    verbose : ``bool`` (optional, default: ``False``)
        If ``True``, verbose information is displayed.

    Returns
    -------

758
    dS : ``float``
759
760
761
762
763
764
765
766
       The entropy difference (per edge) after a full sweep.
    nmoves : ``int``
       The number of accepted block membership moves.


    Notes
    -----

767
    This algorithm performs a Markov chain Monte Carlo sweep on the network,
768
769
    where the block memberships are randomly moved, and either accepted or
    rejected, so that after sufficiently many sweeps the partitions are sampled
770
    with probability proportional to :math:`e^{-\beta\mathcal{S}_{t/c}}`, where
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
    :math:`\mathcal{S}_{t/c}` is the blockmodel entropy, given by

    .. math::

      \mathcal{S}_t &\cong - \frac{1}{2} \sum_{rs}e_{rs}\ln\left(\frac{e_{rs}}{n_rn_s}\right), \\
      \mathcal{S}^d_t &\cong - \sum_{rs}e_{rs}\ln\left(\frac{e_{rs}}{n_rn_s}\right),

    for undirected and directed traditional blockmodels (``deg_corr == False``),
    respectively, where :math:`e_{rs}` is the number of edges from block
    :math:`r` to :math:`s` (or the number of half-edges for the undirected case
    when :math:`r=s`), and :math:`n_r` is the number of vertices in block
    :math:`r`, and constant terms which are independent of the block partition
    were dropped (see :meth:`BlockState.entropy` for the complete entropy). For
    the degree-corrected variant with "hard" degree constraints the equivalent
    expressions are

    .. math::

       \mathcal{S}_c &\cong  - \frac{1}{2} \sum_{rs}e_{rs}\ln\left(\frac{e_{rs}}{e_re_s}\right), \\
       \mathcal{S}^d_c &\cong - \sum_{rs}e_{rs}\ln\left(\frac{e_{rs}}{e^+_re^-_s}\right),

    where :math:`e_r = \sum_se_{rs}` is the number of half-edges incident on
    block :math:`r`, and :math:`e^+_r = \sum_se_{rs}` and :math:`e^-_r =
    \sum_se_{sr}` are the number of out- and in-edges adjacent to block
    :math:`r`, respectively.

    The Monte Carlo algorithm employed attempts to improve the mixing time of
798
799
    the Markov chain by proposing membership moves :math:`r\to s` with
    probability :math:`p(r\to s|t) \propto e_{ts} + c`, where :math:`t` is the
800
    block label of a random neighbour of the vertex being moved. See
Tiago Peixoto's avatar
Tiago Peixoto committed
801
    [peixoto-efficient-2013]_ for more details.
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842

    This algorithm has a complexity of :math:`O(E)`, where :math:`E` is the
    number of edges in the network.

    Examples
    --------
    .. testsetup:: mcmc

       gt.seed_rng(42)
       np.random.seed(42)

    .. doctest:: mcmc

       >>> g = gt.collection.data["polbooks"]
       >>> state = gt.BlockState(g, B=3, deg_corr=True)
       >>> pv = None
       >>> for i in range(1000):        # remove part of the transient
       ...     ds, nmoves = gt.mcmc_sweep(state)
       >>> for i in range(1000):
       ...     ds, nmoves = gt.mcmc_sweep(state)
       ...     pv = gt.collect_vertex_marginals(state, pv)
       >>> gt.graph_draw(g, pos=g.vp["pos"], vertex_shape="pie", vertex_pie_fractions=pv, output="polbooks_blocks_soft.pdf")
       <...>

    .. testcleanup:: mcmc

       gt.graph_draw(g, pos=g.vp["pos"], vertex_shape="pie", vertex_pie_fractions=pv, output="polbooks_blocks_soft.png")

    .. figure:: polbooks_blocks_soft.*
       :align: center

       "Soft" block partition of a political books network with :math:`B=3`.

     References
    ----------

    .. [holland-stochastic-1983] Paul W. Holland, Kathryn Blackmond Laskey,
       Samuel Leinhardt, "Stochastic blockmodels: First steps",
       Carnegie-Mellon University, Pittsburgh, PA 15213, U.S.A., :doi:`10.1016/0378-8733(83)90021-7`
    .. [faust-blockmodels-1992] Katherine Faust, and Stanley
       Wasserman. "Blockmodels: Interpretation and Evaluation." Social Networks
843
       14, no. 1-2 (1992): 5-61. :doi:`10.1016/0378-8733(92)90013-W`
844
845
846
847
848
849
    .. [karrer-stochastic-2011] Brian Karrer, and M. E. J. Newman. "Stochastic
       Blockmodels and Community Structure in Networks." Physical Review E 83,
       no. 1 (2011): 016107. :doi:`10.1103/PhysRevE.83.016107`.
    .. [peixoto-entropy-2012] Tiago P. Peixoto "Entropy of Stochastic Blockmodel
       Ensembles." Physical Review E 85, no. 5 (2012): 056122. :doi:`10.1103/PhysRevE.85.056122`,
       :arxiv:`1112.6028`.
Tiago Peixoto's avatar
Tiago Peixoto committed
850
851
852
    .. [peixoto-parsimonious-2013] Tiago P. Peixoto, "Parsimonious module inference in large networks",
       Phys. Rev. Lett. 110, 148701 (2013), :doi:`10.1103/PhysRevLett.110.148701`, :arxiv:`1212.4794`.
    .. [peixoto-efficient-2013] Tiago P. Peixoto, "Efficient Monte Carlo and greedy
853
       heuristic for the inference of stochastic block models", :arxiv:`1310.4378`.
854
855
    """

856
    if state.B == 1:
857
858
        return 0., 0

859
    if vertices is not None:
860
861
862
        vlist = libcommunity.get_vector(len(vertices))
        vlist.a = vertices
        vertices = vlist
863
        state.sweep_vertices = vertices
864

865
866
867
868
869
    if state.sweep_vertices is None:
        vertices = libcommunity.get_vector(state.g.num_vertices())
        vertices.a = state.g.vertex_index.copy("int").fa
        state.sweep_vertices = vertices

870
871
    random_move = c == float("inf")

872
873
874
875
876
877
878
879
880
    if random_move:
        state._BlockState__build_egroups(empty=True)
    elif state.egroups is None:
        state._BlockState__build_egroups(empty=False)

    if state.nsampler is None:
        state._BlockState__build_nsampler()

    state.bnsampler = None
881
    state.bnnsampler = None
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901

    try:
        dS, nmoves = libcommunity.move_sweep(state.g._Graph__graph,
                                             state.bg._Graph__graph,
                                             state._BlockState__get_emat(),
                                             state.nsampler,
                                             _prop("e", state.bg, state.mrs),
                                             _prop("v", state.bg, state.mrp),
                                             _prop("v", state.bg, state.mrm),
                                             _prop("v", state.bg, state.wr),
                                             _prop("v", state.g, state.b),
                                             _prop("v", state.bg, state.bclabel),
                                             state.sweep_vertices,
                                             state.deg_corr, dense, multigraph,
                                             _prop("e", state.g, state.eweight),
                                             _prop("v", state.g, state.vweight),
                                             state.egroups,
                                             _prop("e", state.g, state.esrcpos),
                                             _prop("e", state.g, state.etgtpos),
                                             float(beta), sequential, random_move,
902
903
                                             c, state.is_weighted, verbose,
                                             _get_rng())
904
905
906
    finally:
        if random_move:
            state.egroups = None
907
908
909
    return dS / state.E, nmoves


910
911
def merge_sweep(state, bm, nmerges, nsweeps=10, dense=False, multigraph=False,
                random_moves=False, verbose=False):
912

913
914
    if state.B == 1:
        return 0., 0
915

916
    if state.bnsampler is None or state.bnnsampler is None:
917
918
919
920
921
922
923
924
        state._BlockState__build_bnsampler()

    state.egroups = None
    state.nsampler = None

    dS, nmoves = libcommunity.merge_sweep(state.bg._Graph__graph,
                                          state._BlockState__get_emat(),
                                          state.bnsampler,
925
                                          state.bnnsampler,
926
927
928
929
930
931
932
933
934
                                          _prop("e", state.bg, state.mrs),
                                          _prop("v", state.bg, state.mrp),
                                          _prop("v", state.bg, state.mrm),
                                          _prop("v", state.bg, state.wr),
                                          _prop("v", state.bg, bm),
                                          _prop("v", state.bg, state.bclabel),
                                          state.deg_corr, dense, multigraph,
                                          nsweeps, nmerges, random_moves,
                                          verbose, _get_rng())
935

936
    return dS / state.E, nmoves
937

938

939
940
941
942
943
944
def greedy_shrink(state, B, nsweeps=10, adaptive_sweeps=True, nmerge_sweeps=None,
                  epsilon=0, r=2, greedy=True, anneal=(1, 1), c=1, dense=False,
                  multigraph=False, random_move=False, verbose=False,
                  sequential=True):
    if B > state.B:
        raise ValueError("Cannot shrink to a larger size!")
945

946
947
    if nmerge_sweeps is None:
        nmerge_sweeps = max((2 * state.g.num_edges()) // state.g.num_vertices(), 1)
948

949
    nmerged = 0
950

951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
    state = BlockState(state.g, b=state.b, B=state.B,
                       clabel=state.clabel, vweight=state.vweight,
                       eweight=state.eweight, deg_corr=state.deg_corr,
                       max_BE=state.max_BE)

    cg = state.bg.copy()
    cg_vweight = cg.own_property(state.wr.copy())
    cg_eweight = cg.own_property(state.mrs.copy())
    cg_clabel = cg.own_property(state.bclabel.copy())

    # merge according to indirect neighbourhood
    bm = state.bg.vertex_index.copy("int")
    random = random_move
    while nmerged < state.B - B:
        dS, nmoves = merge_sweep(state, bm, nmerges=state.B - B - nmerged,
                                 nsweeps=nmerge_sweeps, random_moves=random)
        nmerged += nmoves
968
        if verbose:
969
970
971
972
973
974
975
976
977
978
979
980
981
982
983
984
985
986
987
988
989
990
991
            print("merging", dS, nmoves, nmerged)
        if nmoves == 0:
            random = True
            if verbose:
                print("can't merge... switching to random")

    # Merged block-level state
    bmap = -ones(len(bm.a), dtype=bm.a.dtype)
    libcommunity.vector_map(bm.a, bmap)

    bm = cg.own_property(bm)
    bg_state = BlockState(cg, b=bm, B=B, clabel=cg_clabel,
                          vweight=cg_vweight, eweight=cg_eweight,
                          deg_corr=state.deg_corr, max_BE=state.max_BE)

    if bg_state.g.num_vertices() != state.g.num_vertices() and nsweeps > 0:
        # Perform block-level moves
        if verbose:
            print("Performing block-level moves...")
        multilevel_minimize(bg_state, B=B, nsweeps=nsweeps,
                            adaptive_sweeps=adaptive_sweeps,
                            epsilon=epsilon, r=r, greedy=greedy,
                            anneal=anneal, c=c, dense=dense,
992
993
                            multigraph=multigraph, sequential=sequential,
                            verbose=verbose)
994
995
996
997
998
999
1000

    bm = bg_state.b
    libcommunity.vector_map(state.b.a, bm.a)

    state = BlockState(state.g, b=state.b, B=B, clabel=state.clabel,
                       vweight=state.vweight, eweight=state.eweight,
                       deg_corr=state.deg_corr, max_BE=state.max_BE)