1 /* PSPP - a program for statistical analysis.
2 Copyright (C) 2009, 2010, 2011, 2012 Free Software Foundation, Inc.
4 This program is free software: you can redistribute it and/or modify
5 it under the terms of the GNU General Public License as published by
6 the Free Software Foundation, either version 3 of the License, or
7 (at your option) any later version.
9 This program is distributed in the hope that it will be useful,
10 but WITHOUT ANY WARRANTY; without even the implied warranty of
11 MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
12 GNU General Public License for more details.
14 You should have received a copy of the GNU General Public License
15 along with this program. If not, see <http://www.gnu.org/licenses/>. */
19 #include "math/categoricals.h"
20 #include "math/interaction.h"
24 #include "data/case.h"
25 #include "data/value.h"
26 #include "data/variable.h"
27 #include "libpspp/array.h"
28 #include "libpspp/hmap.h"
29 #include "libpspp/pool.h"
30 #include "libpspp/str.h"
31 #include "libpspp/hash-functions.h"
33 #include "gl/xalloc.h"
35 #define CATEGORICALS_DEBUG 0
37 #define EFFECTS_CODING 1
41 struct hmap_node node; /* Node in hash map. */
43 union value val; /* The value */
45 int index; /* A zero based unique index for this value */
48 struct interaction_value
50 struct hmap_node node; /* Node in hash map */
52 struct ccase *ccase; /* A case (probably the first in the dataset) which matches this value */
54 double cc; /* Total of the weights of cases matching this interaction */
56 void *user_data; /* A pointer to data which the caller can store stuff */
59 static struct value_node *
60 lookup_value (const struct hmap *map, const union value *val, unsigned int hash, int width)
62 struct value_node *vn = NULL;
63 HMAP_FOR_EACH_WITH_HASH (vn, struct value_node, node, hash, map)
65 if (value_equal (&vn->val, val, width))
74 struct hmap_node node; /* Node in hash map. */
75 const struct variable *var; /* The variable */
77 struct hmap valmap; /* A map of value nodes */
78 int n_vals; /* Number of values for this variable */
82 dump_interaction (const struct interaction *iact)
84 if (CATEGORICALS_DEBUG)
86 struct string str = DS_EMPTY_INITIALIZER;
87 interaction_to_string (iact, &str);
88 printf ("Interaction: %s\n", ds_cstr (&str));
93 static struct variable_node *
94 lookup_variable (const struct hmap *map, const struct variable *var, unsigned int hash)
96 struct variable_node *vn = NULL;
97 HMAP_FOR_EACH_WITH_HASH (vn, struct variable_node, node, hash, map)
102 fprintf (stderr, "Warning: Hash table collision\n");
109 struct interact_params
111 /* A map indexed by a interaction_value */
114 const struct interaction *iact;
116 int base_subscript_short;
117 int base_subscript_long;
119 /* The number of distinct values of this interaction */
122 /* An array of integers df_n * df_{n-1} * df_{n-2} ...
123 These are the products of the degrees of freedom for the current
124 variable and all preceeding variables */
129 /* A map of interaction_values indexed by subscript */
130 struct interaction_value **reverse_interaction_value_map;
136 /* Comparison function to sort the reverse_value_map in ascending order */
138 compare_interaction_value_3way (const void *vn1_, const void *vn2_, const void *aux)
140 const struct interaction_value *const *vn1p = vn1_;
141 const struct interaction_value *const *vn2p = vn2_;
143 const struct interact_params *iap = aux;
145 return interaction_case_cmp_3way (iap->iact, (*vn1p)->ccase, (*vn2p)->ccase);
150 /* The weight variable */
151 const struct variable *wv;
153 /* An array of interact_params */
154 struct interact_params *iap;
156 /* Map whose members are the union of the variables which comprise IAP */
159 /* The size of IAP. (ie, the number of interactions involved.) */
162 /* The number of categorical variables which contain entries.
163 In the absence of missing values, this will be equal to N_IAP */
168 /* A map to enable the lookup of variables indexed by subscript.
169 This map considers only the N - 1 of the N variables.
171 int *reverse_variable_map_short;
173 /* Like the above, but uses all N variables */
174 int *reverse_variable_map_long;
180 /* Missing values to be excluded */
181 enum mv_class exclude;
186 const struct payload *payload;
190 categoricals_dump (const struct categoricals *cat)
192 if (CATEGORICALS_DEBUG)
196 printf ("Reverse Variable Map (short):\n");
197 for (i = 0; i < cat->df_sum; ++i)
199 printf (" %d", cat->reverse_variable_map_short[i]);
203 printf ("Reverse Variable Map (long):\n");
204 for (i = 0; i < cat->n_cats_total; ++i)
206 printf (" %d", cat->reverse_variable_map_long[i]);
211 printf ("Number of interactions %d\n", cat->n_iap);
212 for (i = 0 ; i < cat->n_iap; ++i)
216 const struct interact_params *iap = &cat->iap[i];
217 const struct interaction *iact = iap->iact;
219 ds_init_empty (&str);
220 interaction_to_string (iact, &str);
222 printf ("\nInteraction: %s (n: %d); ", ds_cstr (&str), iap->n_cats);
224 printf ("Base subscript: %d\n", iap->base_subscript_short);
227 for (v = 0; v < hmap_count (&iap->ivmap); ++v)
230 const struct interaction_value *iv = iap->reverse_interaction_value_map[v];
232 if (v > 0) printf (" ");
234 for (vv = 0; vv < iact->n_vars; ++vv)
236 const struct variable *var = iact->vars[vv];
237 const union value *val = case_data (iv->ccase, var);
239 printf ("%g", val->f);
240 if (vv < iact->n_vars - 1)
251 categoricals_destroy (struct categoricals *cat)
253 struct variable_node *vn = NULL;
257 for (i = 0; i < cat->n_iap; ++i)
259 struct interaction_value *iv = NULL;
260 /* Interate over each interaction value, and unref any cases that we reffed */
261 HMAP_FOR_EACH (iv, struct interaction_value, node, &cat->iap[i].ivmap)
265 cat->payload->destroy (cat->aux1, iv->user_data);
267 case_unref (iv->ccase);
270 free (cat->iap[i].enc_sum);
271 free (cat->iap[i].df_prod);
272 hmap_destroy (&cat->iap[i].ivmap);
275 /* Interate over each variable and delete its value map */
276 HMAP_FOR_EACH (vn, struct variable_node, node, &cat->varmap)
278 hmap_destroy (&vn->valmap);
281 hmap_destroy (&cat->varmap);
283 pool_destroy (cat->pool);
290 static struct interaction_value *
291 lookup_case (const struct hmap *map, const struct interaction *iact, const struct ccase *c)
293 struct interaction_value *iv = NULL;
294 size_t hash = interaction_case_hash (iact, c, 0);
296 HMAP_FOR_EACH_WITH_HASH (iv, struct interaction_value, node, hash, map)
298 if (interaction_case_equal (iact, c, iv->ccase))
301 fprintf (stderr, "Warning: Hash table collision\n");
308 struct categoricals *
309 categoricals_create (struct interaction *const*inter, size_t n_inter,
310 const struct variable *wv, enum mv_class exclude)
313 struct categoricals *cat = xmalloc (sizeof *cat);
315 cat->n_iap = n_inter;
317 cat->n_cats_total = 0;
319 cat->reverse_variable_map_short = NULL;
320 cat->reverse_variable_map_long = NULL;
321 cat->pool = pool_create ();
322 cat->exclude = exclude;
326 cat->iap = pool_calloc (cat->pool, cat->n_iap, sizeof *cat->iap);
328 hmap_init (&cat->varmap);
329 for (i = 0 ; i < cat->n_iap; ++i)
332 hmap_init (&cat->iap[i].ivmap);
333 cat->iap[i].iact = inter[i];
334 cat->iap[i].cc = 0.0;
335 for (v = 0; v < inter[i]->n_vars; ++v)
337 const struct variable *var = inter[i]->vars[v];
338 unsigned int hash = hash_pointer (var, 0);
339 struct variable_node *vn = lookup_variable (&cat->varmap, var, hash);
342 vn = pool_malloc (cat->pool, sizeof *vn);
345 hmap_init (&vn->valmap);
347 hmap_insert (&cat->varmap, &vn->node, hash);
358 categoricals_update (struct categoricals *cat, const struct ccase *c)
361 struct variable_node *vn = NULL;
362 const double weight = cat->wv ? case_data (c, cat->wv)->f : 1.0;
364 assert (NULL == cat->reverse_variable_map_short);
365 assert (NULL == cat->reverse_variable_map_long);
367 /* Interate over each variable, and add the value of that variable
368 to the appropriate map, if it's not already present. */
369 HMAP_FOR_EACH (vn, struct variable_node, node, &cat->varmap)
371 const int width = var_get_width (vn->var);
372 const union value *val = case_data (c, vn->var);
373 unsigned int hash = value_hash (val, width, 0);
375 struct value_node *valn = lookup_value (&vn->valmap, val, hash, width);
378 valn = pool_malloc (cat->pool, sizeof *valn);
379 valn->index = vn->n_vals++;
380 value_init (&valn->val, width);
381 value_copy (&valn->val, val, width);
382 hmap_insert (&vn->valmap, &valn->node, hash);
386 for (i = 0 ; i < cat->n_iap; ++i)
388 const struct interaction *iact = cat->iap[i].iact;
391 struct interaction_value *node;
393 if ( interaction_case_is_missing (iact, c, cat->exclude))
396 hash = interaction_case_hash (iact, c, 0);
397 node = lookup_case (&cat->iap[i].ivmap, iact, c);
401 node = pool_malloc (cat->pool, sizeof *node);
403 node->ccase = case_ref (c);
406 hmap_insert (&cat->iap[i].ivmap, &node->node, hash);
410 node->user_data = cat->payload->create (cat->aux1, cat->aux2);
417 cat->iap[i].cc += weight;
421 double weight = cat->wv ? case_data (c, cat->wv)->f : 1.0;
422 cat->payload->update (cat->aux1, cat->aux2, node->user_data, c, weight);
428 /* Return the number of categories (distinct values) for interction N */
430 categoricals_n_count (const struct categoricals *cat, size_t n)
432 return hmap_count (&cat->iap[n].ivmap);
437 categoricals_df (const struct categoricals *cat, size_t n)
439 const struct interact_params *iap = &cat->iap[n];
440 return iap->df_prod[iap->iact->n_vars - 1];
444 /* Return the total number of categories */
446 categoricals_n_total (const struct categoricals *cat)
448 if (!categoricals_is_complete (cat))
451 return cat->n_cats_total;
455 categoricals_df_total (const struct categoricals *cat)
461 categoricals_is_complete (const struct categoricals *cat)
463 return (NULL != cat->reverse_variable_map_short);
467 /* This function must be called *before* any call to categoricals_get_*_by subscript and
468 *after* all calls to categoricals_update */
470 categoricals_done (const struct categoricals *cat_)
472 /* Implementation Note: Whilst this function is O(n) in cat->n_cats_total, in most
473 uses it will be more efficient that using a tree based structure, since it
474 is called only once, and means that subsequent lookups will be O(1).
476 1 call of O(n) + 10^9 calls of O(1) is better than 10^9 calls of O(log n).
478 struct categoricals *cat = CONST_CAST (struct categoricals *, cat_);
484 cat->n_cats_total = 0;
486 /* Calculate the degrees of freedom, and the number of categories */
487 for (i = 0 ; i < cat->n_iap; ++i)
490 const struct interaction *iact = cat->iap[i].iact;
492 cat->iap[i].df_prod = iact->n_vars ? xcalloc (iact->n_vars, sizeof (int)) : NULL;
494 cat->iap[i].n_cats = 1;
496 for (v = 0 ; v < iact->n_vars; ++v)
498 const struct variable *var = iact->vars[v];
500 struct variable_node *vn = lookup_variable (&cat->varmap, var, hash_pointer (var, 0));
502 if (hmap_count (&vn->valmap) == 0)
505 cat->iap[i].df_prod[v] = df * (hmap_count (&vn->valmap) - 1);
506 df = cat->iap[i].df_prod[v];
508 cat->iap[i].n_cats *= hmap_count (&vn->valmap);
511 assert (v == iact->n_vars);
513 cat->df_sum += cat->iap[i].df_prod [v - 1];
515 cat->n_cats_total += cat->iap[i].n_cats;
519 cat->reverse_variable_map_short = pool_calloc (cat->pool,
521 sizeof *cat->reverse_variable_map_short);
523 cat->reverse_variable_map_long = pool_calloc (cat->pool,
525 sizeof *cat->reverse_variable_map_long);
527 for (i = 0 ; i < cat->n_iap; ++i)
529 struct interaction_value *ivn = NULL;
532 struct interact_params *iap = &cat->iap[i];
534 iap->base_subscript_short = idx_short;
535 iap->base_subscript_long = idx_long;
537 iap->reverse_interaction_value_map = pool_calloc (cat->pool, iap->n_cats,
538 sizeof *iap->reverse_interaction_value_map);
540 HMAP_FOR_EACH (ivn, struct interaction_value, node, &iap->ivmap)
542 iap->reverse_interaction_value_map[x++] = ivn;
545 assert (x <= iap->n_cats);
547 /* For some purposes (eg CONTRASTS in ONEWAY) the values need to be sorted */
548 sort (iap->reverse_interaction_value_map, x, sizeof (*iap->reverse_interaction_value_map),
549 compare_interaction_value_3way, iap);
551 /* Fill the remaining values with null */
552 for (ii = x ; ii < iap->n_cats; ++ii)
553 iap->reverse_interaction_value_map[ii] = NULL;
555 /* Populate the reverse variable maps. */
558 for (ii = 0; ii < iap->df_prod [iap->iact->n_vars - 1]; ++ii)
559 cat->reverse_variable_map_short[idx_short++] = i;
562 for (ii = 0; ii < iap->n_cats; ++ii)
563 cat->reverse_variable_map_long[idx_long++] = i;
566 assert (cat->n_vars <= cat->n_iap);
568 // categoricals_dump (cat);
570 /* Tally up the sums for all the encodings */
571 for (i = 0 ; i < cat->n_iap; ++i)
574 struct interact_params *iap = &cat->iap[i];
575 const struct interaction *iact = iap->iact;
577 const int df = iap->df_prod ? iap->df_prod [iact->n_vars - 1] : 0;
579 iap->enc_sum = xcalloc (df, sizeof (*(iap->enc_sum)));
581 for (y = 0; y < hmap_count (&iap->ivmap); ++y)
583 struct interaction_value *iv = iap->reverse_interaction_value_map[y];
584 for (x = iap->base_subscript_short; x < iap->base_subscript_short + df ;++x)
586 const double bin = categoricals_get_code_for_case (cat, x, iv->ccase); \
587 iap->enc_sum [x - iap->base_subscript_short] += bin * iv->cc;
597 reverse_variable_lookup_short (const struct categoricals *cat, int subscript)
599 assert (cat->reverse_variable_map_short);
600 assert (subscript >= 0);
601 assert (subscript < cat->df_sum);
603 return cat->reverse_variable_map_short[subscript];
607 reverse_variable_lookup_long (const struct categoricals *cat, int subscript)
609 assert (cat->reverse_variable_map_long);
610 assert (subscript >= 0);
611 assert (subscript < cat->n_cats_total);
613 return cat->reverse_variable_map_long[subscript];
617 /* Return the interaction corresponding to SUBSCRIPT */
618 const struct interaction *
619 categoricals_get_interaction_by_subscript (const struct categoricals *cat, int subscript)
621 int index = reverse_variable_lookup_short (cat, subscript);
623 return cat->iap[index].iact;
627 categoricals_get_weight_by_subscript (const struct categoricals *cat, int subscript)
629 int vindex = reverse_variable_lookup_short (cat, subscript);
630 const struct interact_params *vp = &cat->iap[vindex];
636 categoricals_get_sum_by_subscript (const struct categoricals *cat, int subscript)
638 int vindex = reverse_variable_lookup_short (cat, subscript);
639 const struct interact_params *vp = &cat->iap[vindex];
641 return vp->enc_sum[subscript - vp->base_subscript_short];
644 /* Returns unity if the value in case C at SUBSCRIPT is equal to the category
645 for that subscript */
647 categoricals_get_code_for_case (const struct categoricals *cat, int subscript,
648 const struct ccase *c)
650 const struct interaction *iact = categoricals_get_interaction_by_subscript (cat, subscript);
652 const int i = reverse_variable_lookup_short (cat, subscript);
654 const int base_index = cat->iap[i].base_subscript_short;
659 const struct interact_params *iap = &cat->iap[i];
662 for (v = 0; v < iact->n_vars; ++v)
664 const struct variable *var = iact->vars[v];
666 const union value *val = case_data (c, var);
667 const int width = var_get_width (var);
668 const struct variable_node *vn = lookup_variable (&cat->varmap, var, hash_pointer (var, 0));
670 const unsigned int hash = value_hash (val, width, 0);
671 const struct value_node *valn = lookup_value (&vn->valmap, val, hash, width);
675 const double df = iap->df_prod[v] / dfp;
677 /* Translate the subscript into an index for the individual variable */
678 const int index = ((subscript - base_index) % iap->df_prod[v] ) / dfp;
679 dfp = iap->df_prod [v];
682 if ( valn->index == df )
686 if ( valn->index != index )
697 categoricals_get_n_variables (const struct categoricals *cat)
699 printf ("%s\n", __FUNCTION__);
704 /* Return a case containing the set of values corresponding to
705 the Nth Category of the IACTth interaction */
707 categoricals_get_case_by_category_real (const struct categoricals *cat, int iact, int n)
709 const struct interact_params *vp = &cat->iap[iact];
710 const struct interaction_value *vn = vp->reverse_interaction_value_map [n];
715 /* Return a the user data corresponding to the Nth Category of the IACTth interaction. */
717 categoricals_get_user_data_by_category_real (const struct categoricals *cat, int iact, int n)
719 const struct interact_params *vp = &cat->iap[iact];
720 const struct interaction_value *iv = vp->reverse_interaction_value_map [n];
722 return iv->user_data;
727 /* Return a case containing the set of values corresponding to SUBSCRIPT */
729 categoricals_get_case_by_category (const struct categoricals *cat, int subscript)
731 int vindex = reverse_variable_lookup_long (cat, subscript);
732 const struct interact_params *vp = &cat->iap[vindex];
733 const struct interaction_value *vn = vp->reverse_interaction_value_map [subscript - vp->base_subscript_long];
739 categoricals_get_user_data_by_category (const struct categoricals *cat, int subscript)
741 int vindex = reverse_variable_lookup_long (cat, subscript);
742 const struct interact_params *vp = &cat->iap[vindex];
744 const struct interaction_value *iv = vp->reverse_interaction_value_map [subscript - vp->base_subscript_long];
745 return iv->user_data;
752 categoricals_set_payload (struct categoricals *cat, const struct payload *p,
753 const void *aux1, void *aux2)