1 /* PSPP - a program for statistical analysis.
2 Copyright (C) 2009, 2010, 2011, 2012, 2014 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"
25 #include "data/case.h"
26 #include "data/value.h"
27 #include "data/variable.h"
28 #include "libpspp/array.h"
29 #include "libpspp/hmap.h"
30 #include "libpspp/pool.h"
31 #include "libpspp/str.h"
32 #include "libpspp/hash-functions.h"
34 #include "gl/xalloc.h"
36 #define CATEGORICALS_DEBUG 0
40 struct hmap_node node; /* Node in hash map. */
42 union value val; /* The value */
44 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
55 double cc; /* Total of the weights of cases matching this interaction */
57 void *user_data; /* A pointer to data which the caller can store stuff */
60 static struct value_node *
61 lookup_value (const struct hmap *map, const union value *val, unsigned int hash, int width)
63 struct value_node *vn = NULL;
64 HMAP_FOR_EACH_WITH_HASH (vn, struct value_node, node, hash, map)
66 if (value_equal (&vn->val, val, width))
75 struct hmap_node node; /* Node in hash map. */
76 const struct variable *var; /* The variable */
78 struct hmap valmap; /* A map of value nodes */
82 /* Comparison function to sort value_nodes in ascending order */
84 compare_value_node_3way (const void *vn1_, const void *vn2_, const void *aux)
86 const struct value_node *const *vn1p = vn1_;
87 const struct value_node *const *vn2p = vn2_;
89 const struct variable_node *vn = aux;
92 return value_compare_3way (&(*vn1p)->val, &(*vn2p)->val, var_get_width (vn->var));
97 static struct variable_node *
98 lookup_variable (const struct hmap *map, const struct variable *var, unsigned int hash)
100 struct variable_node *vn = NULL;
101 HMAP_FOR_EACH_WITH_HASH (vn, struct variable_node, node, hash, map)
111 struct interact_params
113 /* The interaction, and an array with iact->n_vars elements such that
114 varnodes[x] points to the variable_node for iact->vars[x]. */
115 const struct interaction *iact;
116 struct variable_node **varnodes;
118 /* An example of each interaction that appears in the data, like a frequency
119 table for 'iact'. By construction, the number of elements must be less
120 than or equal to 'n_cats'.
122 categoricals_update() updates 'ivmap' case-by-case, then
123 categoricals_done() dumps 'ivmap' into 'ivs' and sorts it. */
125 struct interaction_value **ivs;
130 /* Product of hmap_count(&varnodes[*]->valmap), that is, the maximum number
131 of distinct values of this interaction. */
134 /* Product of degrees of freedom of all the variables. */
139 /* Sum of ivs[*]->cc. */
145 compare_interaction_value_3way (const void *vn1_, const void *vn2_, const void *aux)
147 const struct interaction_value *const *vn1p = vn1_;
148 const struct interaction_value *const *vn2p = vn2_;
150 const struct interact_params *iap = aux;
152 return interaction_case_cmp_3way (iap->iact, (*vn1p)->ccase, (*vn2p)->ccase);
157 /* The weight variable */
158 const struct variable *wv;
160 /* An array of interact_params */
161 struct interact_params *iap;
164 /* Map whose members are the union of the variables which comprise IAP */
167 /* The number of categorical variables which contain entries.
168 In the absence of missing values, this will be equal to N_IAP */
171 /* A map to enable the lookup of variables indexed by subscript.
172 This map considers only the N - 1 of the N variables.
174 int *df_to_iact; /* 'df_sum' elements. */
177 /* Like the above, but uses all N variables */
178 int *cat_to_iact; /* 'n_cats_total' elements. */
183 /* Missing values in the factor variables to be excluded */
184 enum mv_class fctr_excl;
191 const struct payload *payload;
196 categoricals_isbalanced (const struct categoricals *cat)
200 for (i = 0 ; i < cat->n_iap; ++i)
203 const struct interact_params *iap = &cat->iap[i];
206 for (v = 0; v < hmap_count (&iap->ivmap); ++v)
208 const struct interaction_value *iv = iap->ivs[v];
220 categoricals_dump (const struct categoricals *cat)
222 if (CATEGORICALS_DEBUG)
226 printf ("df to interaction map:\n");
227 for (i = 0; i < cat->df_sum; ++i)
229 printf (" %d", cat->df_to_iact[i]);
233 printf ("Category to interaction map:\n");
234 for (i = 0; i < cat->n_cats_total; ++i)
236 printf (" %d", cat->cat_to_iact[i]);
240 printf ("Number of interactions %zu\n", cat->n_iap);
241 for (i = 0 ; i < cat->n_iap; ++i)
245 const struct interact_params *iap = &cat->iap[i];
246 const struct interaction *iact = iap->iact;
248 ds_init_empty (&str);
249 interaction_to_string (iact, &str);
251 printf ("\nInteraction: \"%s\" (number of categories: %d); ", ds_cstr (&str), iap->n_cats);
253 printf ("Base index (df/categories): %d/%d\n", iap->base_df, iap->base_cats);
256 for (v = 0; v < hmap_count (&iap->ivmap); ++v)
259 const struct interaction_value *iv = iap->ivs[v];
261 if (v > 0) printf (" ");
263 for (vv = 0; vv < iact->n_vars; ++vv)
265 const struct variable *var = iact->vars[vv];
266 const union value *val = case_data (iv->ccase, var);
267 struct variable_node *vn = iap->varnodes[vv];
269 const int width = var_get_width (var);
270 unsigned int valhash = value_hash (val, width, 0);
271 struct value_node *valn = lookup_value (&vn->valmap, val, valhash, width);
273 assert (vn->var == var);
275 printf ("%.*g(%d)", DBL_DIG + 1, val->f, valn->index);
276 if (vv < iact->n_vars - 1)
287 categoricals_destroy (struct categoricals *cat)
289 struct variable_node *vn = NULL;
294 for (i = 0; i < cat->n_iap; ++i)
296 struct interaction_value *iv = NULL;
297 /* Interate over each interaction value, and unref any cases that we reffed */
298 HMAP_FOR_EACH (iv, struct interaction_value, node, &cat->iap[i].ivmap)
300 if (cat->payload && cat->payload->destroy)
301 cat->payload->destroy (cat->aux1, cat->aux2, iv->user_data);
302 case_unref (iv->ccase);
305 free (cat->iap[i].enc_sum);
306 hmap_destroy (&cat->iap[i].ivmap);
309 /* Interate over each variable and delete its value map */
310 HMAP_FOR_EACH (vn, struct variable_node, node, &cat->varmap)
312 hmap_destroy (&vn->valmap);
315 hmap_destroy (&cat->varmap);
317 pool_destroy (cat->pool);
324 static struct interaction_value *
325 lookup_case (const struct hmap *map, const struct interaction *iact, const struct ccase *c)
327 struct interaction_value *iv = NULL;
328 size_t hash = interaction_case_hash (iact, c, 0);
330 HMAP_FOR_EACH_WITH_HASH (iv, struct interaction_value, node, hash, map)
332 if (interaction_case_equal (iact, c, iv->ccase))
339 /* Returns true iff CAT is sane, that is, if it is complete and has at least
342 categoricals_sane (const struct categoricals *cat)
347 /* Creates and returns a new categoricals object whose variables come from the
348 N_INTER interactions objects in the array starting at INTER. (The INTER
349 objects must outlive the categoricals object because it uses them
352 FCTR_EXCL determines which cases are listwise ignored by
353 categoricals_update(). */
354 struct categoricals *
355 categoricals_create (struct interaction *const*inter, size_t n_inter,
356 const struct variable *wv, enum mv_class fctr_excl)
359 struct categoricals *cat = xmalloc (sizeof *cat);
361 cat->n_iap = n_inter;
363 cat->n_cats_total = 0;
365 cat->df_to_iact = NULL;
366 cat->cat_to_iact = NULL;
367 cat->pool = pool_create ();
368 cat->fctr_excl = fctr_excl;
373 cat->iap = pool_calloc (cat->pool, cat->n_iap, sizeof *cat->iap);
375 hmap_init (&cat->varmap);
376 for (i = 0 ; i < cat->n_iap; ++i)
379 hmap_init (&cat->iap[i].ivmap);
380 cat->iap[i].iact = inter[i];
381 cat->iap[i].cc = 0.0;
382 cat->iap[i].varnodes = pool_nmalloc (cat->pool, cat->iap[i].iact->n_vars,
383 sizeof *cat->iap[i].varnodes);
384 for (v = 0; v < inter[i]->n_vars; ++v)
386 const struct variable *var = inter[i]->vars[v];
387 unsigned int hash = hash_pointer (var, 0);
388 struct variable_node *vn = lookup_variable (&cat->varmap, var, hash);
391 vn = pool_malloc (cat->pool, sizeof *vn);
393 hmap_init (&vn->valmap);
395 hmap_insert (&cat->varmap, &vn->node, hash);
397 cat->iap[i].varnodes[v] = vn;
407 categoricals_update (struct categoricals *cat, const struct ccase *c)
410 struct variable_node *vn = NULL;
416 weight = cat->wv ? case_data (c, cat->wv)->f : 1.0;
417 weight = var_force_valid_weight (cat->wv, weight, NULL);
419 assert (NULL == cat->df_to_iact);
420 assert (NULL == cat->cat_to_iact);
422 /* Interate over each variable, and add the value of that variable
423 to the appropriate map, if it's not already present. */
424 HMAP_FOR_EACH (vn, struct variable_node, node, &cat->varmap)
426 const int width = var_get_width (vn->var);
427 const union value *val = case_data (c, vn->var);
428 unsigned int hash = value_hash (val, width, 0);
430 struct value_node *valn = lookup_value (&vn->valmap, val, hash, width);
433 valn = pool_malloc (cat->pool, sizeof *valn);
435 value_init (&valn->val, width);
436 value_copy (&valn->val, val, width);
437 hmap_insert (&vn->valmap, &valn->node, hash);
441 for (i = 0 ; i < cat->n_iap; ++i)
443 const struct interaction *iact = cat->iap[i].iact;
446 struct interaction_value *node;
448 if ( interaction_case_is_missing (iact, c, cat->fctr_excl))
451 hash = interaction_case_hash (iact, c, 0);
452 node = lookup_case (&cat->iap[i].ivmap, iact, c);
456 node = pool_malloc (cat->pool, sizeof *node);
457 node->ccase = case_ref (c);
460 hmap_insert (&cat->iap[i].ivmap, &node->node, hash);
464 node->user_data = cat->payload->create (cat->aux1, cat->aux2);
471 cat->iap[i].cc += weight;
475 cat->payload->update (cat->aux1, cat->aux2, node->user_data, c, weight);
480 /* Return the number of categories (distinct values) for interaction IDX in
483 categoricals_n_count (const struct categoricals *cat, size_t n)
485 return hmap_count (&cat->iap[n].ivmap);
489 /* Returns the number of degrees of freedom for interaction IDX within CAT. */
491 categoricals_df (const struct categoricals *cat, size_t n)
493 const struct interact_params *iap = &cat->iap[n];
498 /* Return the total number of categories across all interactions in CAT. */
500 categoricals_n_total (const struct categoricals *cat)
502 if (!categoricals_is_complete (cat))
505 return cat->n_cats_total;
508 /* Returns the total degrees of freedom for CAT. */
510 categoricals_df_total (const struct categoricals *cat)
518 /* Returns true iff categoricals_done() has been called for CAT. */
520 categoricals_is_complete (const struct categoricals *cat)
522 return (NULL != cat->df_to_iact);
526 /* This function must be called (once) before any call to the *_by_subscript or
527 *_by_category functions, but AFTER any calls to categoricals_update. If this
528 function returns false, then no calls to _by_subscript or *_by_category are
531 categoricals_done (const struct categoricals *cat_)
533 /* Implementation Note: Whilst this function is O(n) in cat->n_cats_total, in most
534 uses it will be more efficient that using a tree based structure, since it
535 is called only once, and means that subsequent lookups will be O(1).
537 1 call of O(n) + 10^9 calls of O(1) is better than 10^9 calls of O(log n).
539 struct categoricals *cat = CONST_CAST (struct categoricals *, cat_);
549 cat->n_cats_total = 0;
551 /* Assign 'index' to each variables' value_nodes, counting up from 0 in
552 ascending order by value. */
553 struct variable_node *vn;
554 HMAP_FOR_EACH (vn, struct variable_node, node, &cat->varmap)
556 size_t n_vals = hmap_count (&vn->valmap);
563 struct value_node **nodes = xcalloc (sizeof *nodes, n_vals);
565 struct value_node *valnd;
566 HMAP_FOR_EACH (valnd, struct value_node, node, &vn->valmap)
568 sort (nodes, n_vals, sizeof *nodes, compare_value_node_3way, vn);
569 for (x = 0; x < n_vals; ++x)
574 /* Calculate the degrees of freedom, and the number of categories */
575 for (i = 0 ; i < cat->n_iap; ++i)
577 const struct interaction *iact = cat->iap[i].iact;
579 cat->iap[i].n_cats = 1;
580 cat->iap[i].df_prod = 1;
581 for (v = 0 ; v < iact->n_vars; ++v)
583 struct variable_node *vn = cat->iap[i].varnodes[v];
584 size_t n_vals = hmap_count (&vn->valmap);
586 cat->iap[i].df_prod *= n_vals - 1;
587 cat->iap[i].n_cats *= n_vals;
591 cat->df_sum += cat->iap[i].df_prod;
592 cat->n_cats_total += cat->iap[i].n_cats;
596 cat->df_to_iact = pool_calloc (cat->pool, cat->df_sum,
597 sizeof *cat->df_to_iact);
599 cat->cat_to_iact = pool_calloc (cat->pool, cat->n_cats_total,
600 sizeof *cat->cat_to_iact);
602 for (i = 0 ; i < cat->n_iap; ++i)
604 struct interaction_value *ivn = NULL;
607 struct interact_params *iap = &cat->iap[i];
609 iap->base_df = idx_df;
610 iap->base_cats = idx_cat;
612 iap->ivs = pool_calloc (cat->pool, iap->n_cats, sizeof *iap->ivs);
614 HMAP_FOR_EACH (ivn, struct interaction_value, node, &iap->ivmap)
619 assert (x <= iap->n_cats);
621 /* For some purposes (eg CONTRASTS in ONEWAY) the values need to be sorted */
622 sort (iap->ivs, x, sizeof *iap->ivs,
623 compare_interaction_value_3way, iap);
625 /* Fill the remaining values with null */
626 for (ii = x ; ii < iap->n_cats; ++ii)
629 /* Populate the variable maps. */
630 if (iap->iact->n_vars)
632 for (ii = 0; ii < iap->df_prod; ++ii)
633 cat->df_to_iact[idx_df++] = i;
636 for (ii = 0; ii < iap->n_cats; ++ii)
637 cat->cat_to_iact[idx_cat++] = i;
640 assert (cat->n_vars <= cat->n_iap);
642 categoricals_dump (cat);
644 /* Tally up the sums for all the encodings */
645 for (i = 0 ; i < cat->n_iap; ++i)
648 struct interact_params *iap = &cat->iap[i];
649 const struct interaction *iact = iap->iact;
651 const int df = iact->n_vars ? iap->df_prod : 0;
653 iap->enc_sum = xcalloc (df, sizeof (*(iap->enc_sum)));
655 for (y = 0; y < hmap_count (&iap->ivmap); ++y)
657 struct interaction_value *iv = iap->ivs[y];
658 for (x = iap->base_df; x < iap->base_df + df ;++x)
660 const double bin = categoricals_get_effects_code_for_case (cat, x, iv->ccase);
661 iap->enc_sum [x - iap->base_df] += bin * iv->cc;
663 if (cat->payload && cat->payload->calculate)
664 cat->payload->calculate (cat->aux1, cat->aux2, iv->user_data);
673 df_to_iap (const struct categoricals *cat, int subscript)
675 assert (cat->df_to_iact);
676 assert (subscript >= 0);
677 assert (subscript < cat->df_sum);
679 return cat->df_to_iact[subscript];
683 cat_index_to_iap (const struct categoricals *cat, int subscript)
685 assert (cat->cat_to_iact);
686 assert (subscript >= 0);
687 assert (subscript < cat->n_cats_total);
689 return cat->cat_to_iact[subscript];
693 /* Return the interaction corresponding to SUBSCRIPT */
694 const struct interaction *
695 categoricals_get_interaction_by_subscript (const struct categoricals *cat, int subscript)
697 int index = df_to_iap (cat, subscript);
699 return cat->iap[index].iact;
703 categoricals_get_weight_by_subscript (const struct categoricals *cat, int subscript)
705 int vindex = df_to_iap (cat, subscript);
706 const struct interact_params *vp = &cat->iap[vindex];
712 categoricals_get_sum_by_subscript (const struct categoricals *cat, int subscript)
714 int vindex = df_to_iap (cat, subscript);
715 const struct interact_params *vp = &cat->iap[vindex];
717 return vp->enc_sum[subscript - vp->base_df];
721 /* Returns unity if the value in case C at SUBSCRIPT is equal to the category
722 for that subscript */
724 categoricals_get_code_for_case (const struct categoricals *cat, int subscript,
725 const struct ccase *c,
728 const struct interaction *iact = categoricals_get_interaction_by_subscript (cat, subscript);
730 const int i = df_to_iap (cat, subscript);
735 const struct interact_params *iap = &cat->iap[i];
738 for (v = 0; v < iact->n_vars; ++v)
740 const struct variable *var = iact->vars[v];
742 const union value *val = case_data (c, var);
743 const int width = var_get_width (var);
744 const struct variable_node *vn = iap->varnodes[v];
746 const unsigned int hash = value_hash (val, width, 0);
747 const struct value_node *valn = lookup_value (&vn->valmap, val, hash, width);
751 const int df = hmap_count (&iap->varnodes[v]->valmap) - 1;
752 const int dfpn = dfp * df;
753 /* Translate subscript into an index for the individual variable. */
754 const int index = ((subscript - iap->base_df) % dfpn) / dfp;
756 if (effects_coding && valn->index == df )
758 else if ( valn->index != index )
770 /* Returns unity if the value in case C at SUBSCRIPT is equal to the category
771 for that subscript */
773 categoricals_get_dummy_code_for_case (const struct categoricals *cat, int subscript,
774 const struct ccase *c)
776 return categoricals_get_code_for_case (cat, subscript, c, false);
779 /* Returns unity if the value in case C at SUBSCRIPT is equal to the category
781 Else if it is the last category, return -1.
785 categoricals_get_effects_code_for_case (const struct categoricals *cat, int subscript,
786 const struct ccase *c)
788 return categoricals_get_code_for_case (cat, subscript, c, true);
791 /* Return a case containing the set of values corresponding to
792 the Nth Category of the IACTth interaction */
794 categoricals_get_case_by_category_real (const struct categoricals *cat, int iact, int n)
796 const struct interaction_value *vn;
798 const struct interact_params *vp = &cat->iap[iact];
800 if ( n >= hmap_count (&vp->ivmap))
808 /* Return a the user data corresponding to the Nth Category of the IACTth interaction. */
810 categoricals_get_user_data_by_category_real (const struct categoricals *cat, int iact, int n)
812 const struct interact_params *vp = &cat->iap[iact];
813 const struct interaction_value *iv ;
815 if ( n >= hmap_count (&vp->ivmap))
820 return iv->user_data;
825 /* Return a case containing the set of values corresponding to SUBSCRIPT */
827 categoricals_get_case_by_category (const struct categoricals *cat, int subscript)
829 int vindex = cat_index_to_iap (cat, subscript);
830 const struct interact_params *vp = &cat->iap[vindex];
831 const struct interaction_value *vn = vp->ivs [subscript - vp->base_cats];
837 categoricals_get_user_data_by_category (const struct categoricals *cat, int subscript)
839 int vindex = cat_index_to_iap (cat, subscript);
840 const struct interact_params *vp = &cat->iap[vindex];
842 const struct interaction_value *iv = vp->ivs [subscript - vp->base_cats];
843 return iv->user_data;
850 categoricals_set_payload (struct categoricals *cat, const struct payload *p,
851 const void *aux1, void *aux2)