graph_blockmodel_partition.hh 17.8 KB
Newer Older
1
2
// graph-tool -- a general graph modification and manipulation thingy
//
Tiago Peixoto's avatar
Tiago Peixoto committed
3
// Copyright (C) 2006-2020 Tiago de Paula Peixoto <tiago@skewed.de>
4
//
5
6
7
8
// This program is free software; you can redistribute it and/or modify it under
// the terms of the GNU Lesser General Public License as published by the Free
// Software Foundation; either version 3 of the License, or (at your option) any
// later version.
9
//
10
11
12
13
// 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 Lesser General Public License for more
// details.
14
//
15
// You should have received a copy of the GNU Lesser General Public License
16
17
18
19
20
21
22
23
// along with this program. If not, see <http://www.gnu.org/licenses/>.

#ifndef GRAPH_BLOCKMODEL_PARTITION_HH
#define GRAPH_BLOCKMODEL_PARTITION_HH

#include "../support/util.hh"
#include "../support/int_part.hh"

24
25
#include "hash_map_wrap.hh"

Alex Henrie's avatar
Alex Henrie committed
26
27
#include <boost/range/counting_range.hpp>

28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
namespace graph_tool
{

// ===============
// Partition stats
// ===============

constexpr size_t null_group = std::numeric_limits<size_t>::max();

typedef vprop_map_t<std::vector<std::tuple<size_t, size_t, size_t>>>::type
    degs_map_t;

struct simple_degs_t {};

template <class Graph, class Vprop, class Eprop, class F>
43
[[gnu::always_inline]] [[gnu::flatten]] inline
44
45
46
47
48
49
50
void degs_op(size_t v, Vprop& vweight, Eprop& eweight, const simple_degs_t&,
             Graph& g, F&& f)
{
    f(in_degreeS()(v, g, eweight), out_degreeS()(v, g, eweight), vweight[v]);
}

template <class Graph, class Vprop, class Eprop, class F>
51
[[gnu::always_inline]] [[gnu::flatten]] inline
52
void degs_op(size_t v, Vprop& vweight, Eprop& eweight,
53
             const typename degs_map_t::unchecked_t& degs, Graph& g, F&& f)
54
55
56
57
58
59
60
61
62
63
64
65
66
{
    auto& ks = degs[v];
    if (ks.empty())
    {
        degs_op(v, vweight, eweight, simple_degs_t(), g, std::forward<F>(f));
    }
    else
    {
        for (auto& k : ks)
            f(get<0>(k), get<1>(k), get<2>(k));
    }
}

67
68
template <bool use_rmap>
class partition_stats
69
70
71
72
73
74
{
public:

    typedef gt_hash_map<pair<size_t,size_t>, int> map_t;

    template <class Graph, class Vprop, class VWprop, class Eprop, class Degs,
75
              class Vlist>
76
    partition_stats(Graph& g, Vprop& b, Vlist&& vlist, size_t E, size_t B,
77
                    VWprop& vweight, Eprop& eweight, Degs& degs,
78
79
                    std::vector<size_t>& bmap)
        : _bmap(bmap), _N(0), _E(E), _total_B(B)
80
    {
Tiago Peixoto's avatar
Tiago Peixoto committed
81
        if constexpr (!use_rmap)
82
83
84
85
86
87
88
        {
            _hist.resize(num_vertices(g));
            _total.resize(num_vertices(g));
            _ep.resize(num_vertices(g));
            _em.resize(num_vertices(g));
        }

89
90
91
92
93
94
95
96
97
98
        for (auto v : vlist)
        {
            if (vweight[v] == 0)
                continue;

            auto r = get_r(b[v]);

            degs_op(v, vweight, eweight, degs, g,
                    [&](auto kin, auto kout, auto n)
                    {
99
100
101
                        _hist[r][make_pair(kin, kout)] += n;
                        _em[r] += kin * n;
                        _ep[r] += kout * n;
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
                        _total[r] += n;
                        _N += n;
                    });
        }

        _actual_B = 0;
        for (auto n : _total)
        {
            if (n > 0)
                _actual_B++;
        }
    }

    size_t get_r(size_t r)
    {
Tiago Peixoto's avatar
Tiago Peixoto committed
117
        if constexpr (use_rmap)
118
119
120
121
122
123
124
125
126
127
        {
            constexpr size_t null =
                std::numeric_limits<size_t>::max();
            if (r >= _bmap.size())
                _bmap.resize(r + 1, null);
            size_t nr = _bmap[r];
            if (nr == null)
                nr = _bmap[r] = _hist.size();
            r = nr;
        }
128
        if (r >= _hist.size())
129
        {
130
131
132
133
            _hist.resize(r + 1);
            _total.resize(r + 1);
            _ep.resize(r + 1);
            _em.resize(r + 1);
134
        }
135
        return r;
136
137
138
139
140
    }

    double get_partition_dl()
    {
        double S = 0;
141
        S += lbinom(_N - 1, _actual_B - 1);
142
143
144
        S += lgamma_fast(_N + 1);
        for (auto nr : _total)
            S -= lgamma_fast(nr + 1);
145
        S += safelog_fast(_N);
146
147
148
        return S;
    }

149
150
    template <class Rs, class Ks>
    double get_deg_dl_ent(Rs&& rs, Ks&& ks)
151
152
    {
        double S = 0;
153
        for (auto r : rs)
154
        {
155
            r = get_r(r);
156
            size_t total = 0;
157
            if (ks.empty())
158
            {
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
                for (auto& k_c : _hist[r])
                {
                    S -= xlogx_fast(k_c.second);
                    total += k_c.second;
                }
            }
            else
            {
                auto& h = _hist[r];
                for (auto& k : ks)
                {
                    auto iter = h.find(k);
                    auto k_c = (iter != h.end()) ? iter->second : 0;
                    S -= xlogx(k_c);
                }
                total = _total[r];
175
            }
176
            S += xlogx_fast(total);
177
178
179
180
        }
        return S;
    }

181
182
    template <class Rs, class Ks>
    double get_deg_dl_uniform(Rs&& rs, Ks&&)
183
184
    {
        double S = 0;
185
        for (auto r : rs)
186
        {
187
            r = get_r(r);
188
189
190
191
192
193
            S += lbinom(_total[r] + _ep[r] - 1, _ep[r]);
            S += lbinom(_total[r] + _em[r] - 1, _em[r]);
        }
        return S;
    }

194
195
    template <class Rs, class Ks>
    double get_deg_dl_dist(Rs&& rs, Ks&& ks)
196
197
    {
        double S = 0;
198
        for (auto r : rs)
199
        {
200
            r = get_r(r);
201
202
203
204
            S += log_q(_ep[r], _total[r]);
            S += log_q(_em[r], _total[r]);

            size_t total = 0;
205
            if (ks.empty())
206
            {
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
                for (auto& k_c : _hist[r])
                {
                    S -= lgamma_fast(k_c.second + 1);
                    total += k_c.second;
                }
            }
            else
            {
                auto& h = _hist[r];
                for (auto& k : ks)
                {
                    auto iter = h.find(k);
                    auto k_c = (iter != h.end()) ? iter->second : 0;
                    S -= lgamma_fast(k_c + 1);
                }
                total = _total[r];
223
224
225
226
227
228
            }
            S += lgamma_fast(total + 1);
        }
        return S;
    }

229
230
    template <class Rs, class Ks>
    double get_deg_dl(int kind, Rs&& rs, Ks&& ks)
231
232
233
234
    {
        switch (kind)
        {
        case deg_dl_kind::ENT:
235
            return get_deg_dl_ent(rs, ks);
236
        case deg_dl_kind::UNIFORM:
237
            return get_deg_dl_uniform(rs, ks);
238
        case deg_dl_kind::DIST:
239
            return get_deg_dl_dist(rs, ks);
240
241
242
243
244
        default:
            return numeric_limits<double>::quiet_NaN();
        }
    }

245
246
247
248
249
250
251
252
253
254
255
256
257
    double get_deg_dl(int kind)
    {
        return get_deg_dl(kind, boost::counting_range(size_t(0), _total_B),
                          std::array<std::pair<size_t,size_t>,0>());
    }

    template <class Graph>
    double get_edges_dl(size_t B, Graph& g)
    {
        size_t BB = (graph_tool::is_directed(g)) ? B * B : (B * (B + 1)) / 2;
        return lbinom(BB + _E - 1, _E);
    }

258
259
260
261
262
263
264
265
266
267
    template <class VProp>
    double get_delta_partition_dl(size_t v, size_t r, size_t nr, VProp& vweight)
    {
        if (r == nr)
            return 0;

        if (r != null_group)
            r = get_r(r);

        if (nr != null_group)
268
            nr = get_r(nr);
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
307

        int n = vweight[v];
        if (n == 0)
        {
            if (r == null_group)
                n = 1;
            else
                return 0;
        }

        double S_b = 0, S_a = 0;

        if (r != null_group)
        {
            S_b += -lgamma_fast(_total[r] + 1);
            S_a += -lgamma_fast(_total[r] - n + 1);
        }

        if (nr != null_group)
        {
            S_b += -lgamma_fast(_total[nr] + 1);
            S_a += -lgamma_fast(_total[nr] + n + 1);
        }

        int dN = 0;
        if (r == null_group)
            dN += n;
        if (nr == null_group)
            dN -= n;

        S_b += lgamma_fast(_N + 1);
        S_a += lgamma_fast(_N + dN + 1);

        int dB = 0;
        if (r != null_group && _total[r] == n)
            dB--;
        if (nr != null_group && _total[nr] == 0)
            dB++;

308
        if ((dN != 0 || dB != 0))
309
310
311
312
313
314
315
        {
            S_b += lbinom_fast(_N - 1, _actual_B - 1);
            S_a += lbinom_fast(_N - 1 + dN, _actual_B + dB - 1);
        }

        if (dN != 0)
        {
316
317
            S_b += safelog_fast(_N);
            S_a += safelog_fast(_N + dN);
318
319
320
321
322
323
324
        }

        return S_a - S_b;
    }

    template <class VProp, class Graph>
    double get_delta_edges_dl(size_t v, size_t r, size_t nr, VProp& vweight,
325
                              size_t actual_B, Graph& g)
326
    {
327
        if (r == nr)
328
329
330
331
332
            return 0;

        if (r != null_group)
            r = get_r(r);
        if (nr != null_group)
333
            nr = get_r(nr);
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354

        double S_b = 0, S_a = 0;

        int n = vweight[v];

        if (n == 0)
        {
            if (r == null_group)
                n = 1;
            else
                return 0;
        }

        int dB = 0;
        if (r != null_group && _total[r] == n)
            dB--;
        if (nr != null_group && _total[nr] == 0)
            dB++;

        if (dB != 0)
        {
355
356
            S_b += get_edges_dl(actual_B, g);
            S_a += get_edges_dl(actual_B + dB, g);
357
358
359
360
361
362
363
364
365
        }

        return S_a - S_b;
    }

    template <class Graph, class VProp, class EProp, class Degs>
    double get_delta_deg_dl(size_t v, size_t r, size_t nr, VProp& vweight,
                            EProp& eweight, Degs& degs, Graph& g, int kind)
    {
366
        if (r == nr || vweight[v] == 0)
367
368
369
370
            return 0;
        if (r != null_group)
            r = get_r(r);
        if (nr != null_group)
371
            nr = get_r(nr);
372
373
374
375

        auto dop =
            [&](auto&& f)
            {
376
377
                degs_op(v, vweight, eweight, degs, g,
                        [&](auto... k) { f(k...); });
378
379
380
381
382
383
384
385
386
387
388
389
390
            };

        double dS = 0;
        switch (kind)
        {
        case deg_dl_kind::ENT:
            if (r != null_group)
                dS += get_delta_deg_dl_ent_change(r,  dop, -1);
            if (nr != null_group)
                dS += get_delta_deg_dl_ent_change(nr, dop, +1);
            break;
        case deg_dl_kind::UNIFORM:
            if (r != null_group)
391
                dS += get_delta_deg_dl_uniform_change(r,  dop, -1);
392
            if (nr != null_group)
393
                dS += get_delta_deg_dl_uniform_change(nr, dop, +1);
394
395
396
            break;
        case deg_dl_kind::DIST:
            if (r != null_group)
397
                dS += get_delta_deg_dl_dist_change(r,  dop, -1);
398
            if (nr != null_group)
399
                dS += get_delta_deg_dl_dist_change(nr, dop, +1);
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
            break;
        default:
            dS = numeric_limits<double>::quiet_NaN();
        }
        return dS;
    }

    template <class DegOP>
    double get_delta_deg_dl_ent_change(size_t r, DegOP&& dop, int diff)
    {
        int nr = _total[r];
        auto get_Sk = [&](size_t s, pair<size_t, size_t>& deg, int delta)
            {
                int nd = 0;
                auto iter = _hist[s].find(deg);
                if (iter != _hist[s].end())
                    nd = iter->second;
                assert(nd + delta >= 0);
418
                return -xlogx_fast(nd + delta);
419
420
421
422
423
424
425
426
427
428
429
430
431
            };

        double S_b = 0, S_a = 0;
        int dn = 0;

        dop([&](size_t kin, size_t kout, int nk)
            {
                dn += diff * nk;
                auto deg = make_pair(kin, kout);
                S_b += get_Sk(r, deg,         0);
                S_a += get_Sk(r, deg, diff * nk);
            });

432
433
        S_b += xlogx_fast(nr);
        S_a += xlogx_fast(nr + dn);
434
435
436
437
438

        return S_a - S_b;
    }

    template <class DegOP>
439
    double get_delta_deg_dl_uniform_change(size_t r, DegOP&& dop, int diff)
440
441
442
443
    {
        auto get_Se = [&](int dn, int dkin, int dkout)
            {
                double S = 0;
444
445
                S += lbinom_fast(_total[r] + dn + _ep[r] - 1 + dkout, _ep[r] + dkout);
                S += lbinom_fast(_total[r] + dn + _em[r] - 1 + dkin,  _em[r] + dkin);
446
447
448
449
450
451
452
453
                return S;
            };

        double S_b = 0, S_a = 0;
        int tkin = 0, tkout = 0, n = 0;
        dop([&](auto kin, auto kout, int nk)
            {
                tkin += kin * nk;
454
                tkout += kout * nk;
455
456
457
458
459
460
461
462
463
                n += nk;
            });

        S_b += get_Se(       0,           0,            0);
        S_a += get_Se(diff * n, diff * tkin, diff * tkout);
        return S_a - S_b;
    }

    template <class DegOP>
464
    double get_delta_deg_dl_dist_change(size_t r, DegOP&& dop, int diff)
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
493
494
495
496
497
    {
        auto get_Se = [&](int delta, int kin, int kout)
            {
                double S = 0;
                assert(_total[r] + delta >= 0);
                assert(_em[r] + kin >= 0);
                assert(_ep[r] + kout >= 0);
                S += log_q(_em[r] + kin, _total[r] + delta);
                S += log_q(_ep[r] + kout, _total[r] + delta);
                return S;
            };

        auto get_Sr = [&](int delta)
            {
                assert(_total[r] + delta + 1 >= 0);
                return lgamma_fast(_total[r] + delta + 1);
            };

        auto get_Sk = [&](pair<size_t, size_t>& deg, int delta)
            {
                int nd = 0;
                auto iter = _hist[r].find(deg);
                if (iter != _hist[r].end())
                    nd = iter->second;
                assert(nd + delta >= 0);
                return -lgamma_fast(nd + delta + 1);
            };

        double S_b = 0, S_a = 0;
        int tkin = 0, tkout = 0, n = 0;
        dop([&](size_t kin, size_t kout, int nk)
            {
                tkin += kin * nk;
498
                tkout += kout * nk;
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
                n += nk;

                auto deg = make_pair(kin, kout);
                S_b += get_Sk(deg,         0);
                S_a += get_Sk(deg, diff * nk);
            });

        S_b += get_Se(       0,           0,            0);
        S_a += get_Se(diff * n, diff * tkin, diff * tkout);

        S_b += get_Sr(       0);
        S_a += get_Sr(diff * n);

        return S_a - S_b;
    }

515
516
    template <class VWeight>
    void change_vertex(size_t v, size_t r, VWeight& vweight, int diff)
517
    {
518
519
520
521
522
523
524
525
526
527
528
529
530
        int vw = vweight[v];
        int dv = vw * diff;

        if (_total[r] == 0 && dv > 0)
            _actual_B++;

        if (_total[r] == vw && dv < 0)
            _actual_B--;

        _total[r] += dv;
        _N += dv;

        assert(_total[r] >= 0);
531
    }
532

533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
    template <class Graph, class VWeight, class EWeight, class Degs>
    void change_vertex_degs(size_t v, size_t r, Graph& g, VWeight& vweight,
                            EWeight& eweight, Degs& degs, int diff)
    {
        degs_op(v, vweight, eweight, degs, g,
                [&](auto kin, auto kout, auto n)
                {
                    int dk = diff * n;
                    auto& h = _hist[r];
                    auto deg = make_pair(kin, kout);
                    auto iter = h.insert({deg, 0}).first;
                    iter->second += dk;
                    if (iter->second == 0)
                        h.erase(iter);
                    _em[r] += dk * deg.first;
                    _ep[r] += dk * deg.second;
                });
550
551
552
553
554
555
556
557
558
    }

    template <class Graph, class VWeight, class EWeight, class Degs>
    void remove_vertex(size_t v, size_t r, bool deg_corr, Graph& g,
                       VWeight& vweight, EWeight& eweight, Degs& degs)
    {
        if (r == null_group || vweight[v] == 0)
            return;
        r = get_r(r);
559
560
561
        change_vertex(v, r, vweight, -1);
        if (deg_corr)
            change_vertex_degs(v, r, g, vweight, eweight, degs, -1);
562
563
564
565
566
567
568
569
    }

    template <class Graph, class VWeight, class EWeight, class Degs>
    void add_vertex(size_t v, size_t nr, bool deg_corr, Graph& g,
                    VWeight& vweight, EWeight& eweight, Degs& degs)
    {
        if (nr == null_group || vweight[v] == 0)
            return;
570
        nr = get_r(nr);
571
572
573
        change_vertex(v, nr, vweight, 1);
        if (deg_corr)
            change_vertex_degs(v, nr, g, vweight, eweight, degs, 1);
574
575
    }

576
577
578
579
580
    void change_E(int dE)
    {
        _E += dE;
    }

581
582
583
584
585
    size_t get_N()
    {
        return _N;
    }

586
587
588
589
590
    size_t get_E()
    {
        return _E;
    }

591
592
593
594
595
596
597
598
599
600
    size_t get_actual_B()
    {
        return _actual_B;
    }

    void add_block()
    {
        _total_B++;
    }

601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
    template <class Graph, class VProp, class VWeight, class EWeight, class Degs>
    bool check_degs(Graph& g, VProp& b, VWeight& vweight, EWeight& eweight, Degs& degs)
    {
        vector<map_t> dhist;
        for (auto v : vertices_range(g))
        {
            degs_op(v, vweight, eweight, degs, g,
                    [&](auto kin, auto kout, auto n)
                    {
                        auto r = get_r(b[v]);
                        if (r >= dhist.size())
                            dhist.resize(r + 1);
                        dhist[r][{kin, kout}] += n;
                    });
        }

        for (size_t r = 0; r < dhist.size(); ++r)
        {
            for (auto& kn : dhist[r])
            {
                auto count = (r >= _hist.size()) ? 0 : _hist[r][kn.first];
                if (kn.second != count)
                {
                    assert(false);
                    return false;
                }
            }
        }

        for (size_t r = 0; r < _hist.size(); ++r)
        {
            for (auto& kn : _hist[r])
            {
                auto count = (r >= dhist.size()) ? 0 : dhist[r][kn.first];
                if (kn.second != count)
                {
                    assert(false);
                    return false;
                }
            }
        }

        return true;
    }
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659

private:
    vector<size_t>& _bmap;
    size_t _N;
    size_t _E;
    size_t _actual_B;
    size_t _total_B;
    vector<map_t> _hist;
    vector<int> _total;
    vector<int> _ep;
    vector<int> _em;
};

} //namespace graph_tool
#endif // GRAPH_BLOCKMODEL_PARTITION_HH