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 */
79 int n_vals; /* Number of values for this variable */
83 /* Comparison function to sort value_nodes in ascending order */
85 compare_value_node_3way (const void *vn1_, const void *vn2_, const void *aux)
87 const struct value_node *const *vn1p = vn1_;
88 const struct value_node *const *vn2p = vn2_;
90 const struct variable_node *vn = aux;
93 return value_compare_3way (&(*vn1p)->val, &(*vn2p)->val, var_get_width (vn->var));
98 static struct variable_node *
99 lookup_variable (const struct hmap *map, const struct variable *var, unsigned int hash)
101 struct variable_node *vn = NULL;
102 HMAP_FOR_EACH_WITH_HASH (vn, struct variable_node, node, hash, map)
112 struct interact_params
114 /* An example of each interaction that appears in the data, like a frequency
115 table for 'iact'. By construction, the number of elements must be less
116 than or equal to 'n_cats'.
118 categoricals_update() updates 'ivmap' case-by-case, then
119 categoricals_done() dumps 'ivmap' into 'reverse_interaction_value_map' and
122 struct interaction_value **reverse_interaction_value_map;
124 const struct interaction *iact;
126 int base_subscript_short;
127 int base_subscript_long;
129 /* Product of hmap_count(&varnodes[*]->valmap), that is, the maximum number
130 of distinct values of this interaction. */
133 /* An array of integers df_n * df_{n-1} * df_{n-2} ...
134 These are the products of the degrees of freedom for the current
135 variable and all preceding variables */
140 /* Sum of reverse_interaction_value_map[*]->cc. */
145 /* Comparison function to sort the reverse_value_map in ascending order */
147 compare_interaction_value_3way (const void *vn1_, const void *vn2_, const void *aux)
149 const struct interaction_value *const *vn1p = vn1_;
150 const struct interaction_value *const *vn2p = vn2_;
152 const struct interact_params *iap = aux;
154 return interaction_case_cmp_3way (iap->iact, (*vn1p)->ccase, (*vn2p)->ccase);
159 /* The weight variable */
160 const struct variable *wv;
162 /* An array of interact_params */
163 struct interact_params *iap;
166 /* Map whose members are the union of the variables which comprise IAP */
169 /* The number of categorical variables which contain entries.
170 In the absence of missing values, this will be equal to N_IAP */
173 /* A map to enable the lookup of variables indexed by subscript.
174 This map considers only the N - 1 of the N variables.
176 int *reverse_variable_map_short; /* 'df_sum' elements. */
179 /* Like the above, but uses all N variables */
180 int *reverse_variable_map_long; /* 'n_cats_total' elements. */
185 /* Missing values in the factor variables to be excluded */
186 enum mv_class fctr_excl;
193 const struct payload *payload;
198 categoricals_isbalanced (const struct categoricals *cat)
202 for (i = 0 ; i < cat->n_iap; ++i)
205 const struct interact_params *iap = &cat->iap[i];
208 for (v = 0; v < hmap_count (&iap->ivmap); ++v)
210 const struct interaction_value *iv = iap->reverse_interaction_value_map[v];
222 categoricals_dump (const struct categoricals *cat)
224 if (CATEGORICALS_DEBUG)
228 printf ("Reverse Variable Map (short):\n");
229 for (i = 0; i < cat->df_sum; ++i)
231 printf (" %d", cat->reverse_variable_map_short[i]);
235 printf ("Reverse Variable Map (long):\n");
236 for (i = 0; i < cat->n_cats_total; ++i)
238 printf (" %d", cat->reverse_variable_map_long[i]);
242 printf ("Number of interactions %zu\n", cat->n_iap);
243 for (i = 0 ; i < cat->n_iap; ++i)
247 const struct interact_params *iap = &cat->iap[i];
248 const struct interaction *iact = iap->iact;
250 ds_init_empty (&str);
251 interaction_to_string (iact, &str);
253 printf ("\nInteraction: \"%s\" (number of categories: %d); ", ds_cstr (&str), iap->n_cats);
255 printf ("Base index (short/long): %d/%d\n", iap->base_subscript_short, iap->base_subscript_long);
258 for (v = 0; v < hmap_count (&iap->ivmap); ++v)
261 const struct interaction_value *iv = iap->reverse_interaction_value_map[v];
263 if (v > 0) printf (" ");
265 for (vv = 0; vv < iact->n_vars; ++vv)
267 const struct variable *var = iact->vars[vv];
268 const union value *val = case_data (iv->ccase, var);
269 unsigned int varhash = hash_pointer (var, 0);
270 struct variable_node *vn = lookup_variable (&cat->varmap, var, varhash);
272 const int width = var_get_width (var);
273 unsigned int valhash = value_hash (val, width, 0);
274 struct value_node *valn = lookup_value (&vn->valmap, val, valhash, width);
276 assert (vn->var == var);
278 printf ("%.*g(%d)", DBL_DIG + 1, val->f, valn->index);
279 if (vv < iact->n_vars - 1)
290 categoricals_destroy (struct categoricals *cat)
292 struct variable_node *vn = NULL;
297 for (i = 0; i < cat->n_iap; ++i)
299 struct interaction_value *iv = NULL;
300 /* Interate over each interaction value, and unref any cases that we reffed */
301 HMAP_FOR_EACH (iv, struct interaction_value, node, &cat->iap[i].ivmap)
303 if (cat->payload && cat->payload->destroy)
304 cat->payload->destroy (cat->aux1, cat->aux2, iv->user_data);
305 case_unref (iv->ccase);
308 free (cat->iap[i].enc_sum);
309 free (cat->iap[i].df_prod);
310 hmap_destroy (&cat->iap[i].ivmap);
313 /* Interate over each variable and delete its value map */
314 HMAP_FOR_EACH (vn, struct variable_node, node, &cat->varmap)
316 hmap_destroy (&vn->valmap);
319 hmap_destroy (&cat->varmap);
321 pool_destroy (cat->pool);
328 static struct interaction_value *
329 lookup_case (const struct hmap *map, const struct interaction *iact, const struct ccase *c)
331 struct interaction_value *iv = NULL;
332 size_t hash = interaction_case_hash (iact, c, 0);
334 HMAP_FOR_EACH_WITH_HASH (iv, struct interaction_value, node, hash, map)
336 if (interaction_case_equal (iact, c, iv->ccase))
343 /* Returns true iff CAT is sane, that is, if it is complete and has at least
346 categoricals_sane (const struct categoricals *cat)
351 /* Creates and returns a new categoricals object whose variables come from the
352 N_INTER interactions objects in the array starting at INTER. (The INTER
353 objects must outlive the categoricals object because it uses them
356 FCTR_EXCL determines which cases are listwise ignored by
357 categoricals_update(). */
358 struct categoricals *
359 categoricals_create (struct interaction *const*inter, size_t n_inter,
360 const struct variable *wv, enum mv_class fctr_excl)
363 struct categoricals *cat = xmalloc (sizeof *cat);
365 cat->n_iap = n_inter;
367 cat->n_cats_total = 0;
369 cat->reverse_variable_map_short = NULL;
370 cat->reverse_variable_map_long = NULL;
371 cat->pool = pool_create ();
372 cat->fctr_excl = fctr_excl;
377 cat->iap = pool_calloc (cat->pool, cat->n_iap, sizeof *cat->iap);
379 hmap_init (&cat->varmap);
380 for (i = 0 ; i < cat->n_iap; ++i)
383 hmap_init (&cat->iap[i].ivmap);
384 cat->iap[i].iact = inter[i];
385 cat->iap[i].cc = 0.0;
386 for (v = 0; v < inter[i]->n_vars; ++v)
388 const struct variable *var = inter[i]->vars[v];
389 unsigned int hash = hash_pointer (var, 0);
390 struct variable_node *vn = lookup_variable (&cat->varmap, var, hash);
393 vn = pool_malloc (cat->pool, sizeof *vn);
396 hmap_init (&vn->valmap);
398 hmap_insert (&cat->varmap, &vn->node, hash);
409 categoricals_update (struct categoricals *cat, const struct ccase *c)
412 struct variable_node *vn = NULL;
418 weight = cat->wv ? case_data (c, cat->wv)->f : 1.0;
419 weight = var_force_valid_weight (cat->wv, weight, NULL);
421 assert (NULL == cat->reverse_variable_map_short);
422 assert (NULL == cat->reverse_variable_map_long);
424 /* Interate over each variable, and add the value of that variable
425 to the appropriate map, if it's not already present. */
426 HMAP_FOR_EACH (vn, struct variable_node, node, &cat->varmap)
428 const int width = var_get_width (vn->var);
429 const union value *val = case_data (c, vn->var);
430 unsigned int hash = value_hash (val, width, 0);
432 struct value_node *valn = lookup_value (&vn->valmap, val, hash, width);
435 valn = pool_malloc (cat->pool, sizeof *valn);
438 value_init (&valn->val, width);
439 value_copy (&valn->val, val, width);
440 hmap_insert (&vn->valmap, &valn->node, hash);
444 for (i = 0 ; i < cat->n_iap; ++i)
446 const struct interaction *iact = cat->iap[i].iact;
449 struct interaction_value *node;
451 if ( interaction_case_is_missing (iact, c, cat->fctr_excl))
454 hash = interaction_case_hash (iact, c, 0);
455 node = lookup_case (&cat->iap[i].ivmap, iact, c);
459 node = pool_malloc (cat->pool, sizeof *node);
460 node->ccase = case_ref (c);
463 hmap_insert (&cat->iap[i].ivmap, &node->node, hash);
467 node->user_data = cat->payload->create (cat->aux1, cat->aux2);
474 cat->iap[i].cc += weight;
478 cat->payload->update (cat->aux1, cat->aux2, node->user_data, c, weight);
483 /* Return the number of categories (distinct values) for interaction IDX in
486 categoricals_n_count (const struct categoricals *cat, size_t n)
488 return hmap_count (&cat->iap[n].ivmap);
492 /* Returns the number of degrees of freedom for interaction IDX within CAT. */
494 categoricals_df (const struct categoricals *cat, size_t n)
496 const struct interact_params *iap = &cat->iap[n];
497 return iap->df_prod[iap->iact->n_vars - 1];
501 /* Return the total number of categories across all interactions in CAT. */
503 categoricals_n_total (const struct categoricals *cat)
505 if (!categoricals_is_complete (cat))
508 return cat->n_cats_total;
511 /* Returns the total degrees of freedom for CAT. */
513 categoricals_df_total (const struct categoricals *cat)
521 /* Returns true iff categoricals_done() has been called for CAT. */
523 categoricals_is_complete (const struct categoricals *cat)
525 return (NULL != cat->reverse_variable_map_short);
529 /* This function must be called (once) before any call to the *_by_subscript or
530 *_by_category functions, but AFTER any calls to categoricals_update. If this
531 function returns false, then no calls to _by_subscript or *_by_category are
534 categoricals_done (const struct categoricals *cat_)
536 /* Implementation Note: Whilst this function is O(n) in cat->n_cats_total, in most
537 uses it will be more efficient that using a tree based structure, since it
538 is called only once, and means that subsequent lookups will be O(1).
540 1 call of O(n) + 10^9 calls of O(1) is better than 10^9 calls of O(log n).
542 struct categoricals *cat = CONST_CAST (struct categoricals *, cat_);
552 cat->n_cats_total = 0;
554 /* Calculate the degrees of freedom, and the number of categories */
555 for (i = 0 ; i < cat->n_iap; ++i)
558 const struct interaction *iact = cat->iap[i].iact;
560 cat->iap[i].df_prod = iact->n_vars ? xcalloc (iact->n_vars, sizeof (int)) : NULL;
562 cat->iap[i].n_cats = 1;
564 for (v = 0 ; v < iact->n_vars; ++v)
567 const struct variable *var = iact->vars[v];
569 struct variable_node *vn = lookup_variable (&cat->varmap, var, hash_pointer (var, 0));
571 struct value_node *valnd = NULL;
572 struct value_node **array ;
574 assert (vn->n_vals == hmap_count (&vn->valmap));
582 /* Sort the VALMAP here */
583 array = xcalloc (sizeof *array, vn->n_vals);
585 HMAP_FOR_EACH (valnd, struct value_node, node, &vn->valmap)
587 /* Note: This loop is probably superfluous, it could be done in the
588 update stage (at the expense of a realloc) */
592 sort (array, vn->n_vals, sizeof (*array),
593 compare_value_node_3way, vn);
595 for (x = 0; x < vn->n_vals; ++x)
597 struct value_node *vvv = array[x];
602 cat->iap[i].df_prod[v] = df * (vn->n_vals - 1);
603 df = cat->iap[i].df_prod[v];
605 cat->iap[i].n_cats *= vn->n_vals;
609 cat->df_sum += cat->iap[i].df_prod [v - 1];
611 cat->n_cats_total += cat->iap[i].n_cats;
615 cat->reverse_variable_map_short = pool_calloc (cat->pool,
617 sizeof *cat->reverse_variable_map_short);
619 cat->reverse_variable_map_long = pool_calloc (cat->pool,
621 sizeof *cat->reverse_variable_map_long);
623 for (i = 0 ; i < cat->n_iap; ++i)
625 struct interaction_value *ivn = NULL;
628 struct interact_params *iap = &cat->iap[i];
630 iap->base_subscript_short = idx_short;
631 iap->base_subscript_long = idx_long;
633 iap->reverse_interaction_value_map = pool_calloc (cat->pool, iap->n_cats,
634 sizeof *iap->reverse_interaction_value_map);
636 HMAP_FOR_EACH (ivn, struct interaction_value, node, &iap->ivmap)
638 iap->reverse_interaction_value_map[x++] = ivn;
641 assert (x <= iap->n_cats);
643 /* For some purposes (eg CONTRASTS in ONEWAY) the values need to be sorted */
644 sort (iap->reverse_interaction_value_map, x, sizeof (*iap->reverse_interaction_value_map),
645 compare_interaction_value_3way, iap);
647 /* Fill the remaining values with null */
648 for (ii = x ; ii < iap->n_cats; ++ii)
649 iap->reverse_interaction_value_map[ii] = NULL;
651 /* Populate the reverse variable maps. */
654 for (ii = 0; ii < iap->df_prod [iap->iact->n_vars - 1]; ++ii)
655 cat->reverse_variable_map_short[idx_short++] = i;
658 for (ii = 0; ii < iap->n_cats; ++ii)
659 cat->reverse_variable_map_long[idx_long++] = i;
662 assert (cat->n_vars <= cat->n_iap);
664 categoricals_dump (cat);
666 /* Tally up the sums for all the encodings */
667 for (i = 0 ; i < cat->n_iap; ++i)
670 struct interact_params *iap = &cat->iap[i];
671 const struct interaction *iact = iap->iact;
673 const int df = iap->df_prod ? iap->df_prod [iact->n_vars - 1] : 0;
675 iap->enc_sum = xcalloc (df, sizeof (*(iap->enc_sum)));
677 for (y = 0; y < hmap_count (&iap->ivmap); ++y)
679 struct interaction_value *iv = iap->reverse_interaction_value_map[y];
680 for (x = iap->base_subscript_short; x < iap->base_subscript_short + df ;++x)
682 const double bin = categoricals_get_effects_code_for_case (cat, x, iv->ccase);
683 iap->enc_sum [x - iap->base_subscript_short] += bin * iv->cc;
685 if (cat->payload && cat->payload->calculate)
686 cat->payload->calculate (cat->aux1, cat->aux2, iv->user_data);
695 reverse_variable_lookup_short (const struct categoricals *cat, int subscript)
697 assert (cat->reverse_variable_map_short);
698 assert (subscript >= 0);
699 assert (subscript < cat->df_sum);
701 return cat->reverse_variable_map_short[subscript];
705 reverse_variable_lookup_long (const struct categoricals *cat, int subscript)
707 assert (cat->reverse_variable_map_long);
708 assert (subscript >= 0);
709 assert (subscript < cat->n_cats_total);
711 return cat->reverse_variable_map_long[subscript];
715 /* Return the interaction corresponding to SUBSCRIPT */
716 const struct interaction *
717 categoricals_get_interaction_by_subscript (const struct categoricals *cat, int subscript)
719 int index = reverse_variable_lookup_short (cat, subscript);
721 return cat->iap[index].iact;
725 categoricals_get_weight_by_subscript (const struct categoricals *cat, int subscript)
727 int vindex = reverse_variable_lookup_short (cat, subscript);
728 const struct interact_params *vp = &cat->iap[vindex];
734 categoricals_get_sum_by_subscript (const struct categoricals *cat, int subscript)
736 int vindex = reverse_variable_lookup_short (cat, subscript);
737 const struct interact_params *vp = &cat->iap[vindex];
739 return vp->enc_sum[subscript - vp->base_subscript_short];
743 /* Returns unity if the value in case C at SUBSCRIPT is equal to the category
744 for that subscript */
746 categoricals_get_code_for_case (const struct categoricals *cat, int subscript,
747 const struct ccase *c,
750 const struct interaction *iact = categoricals_get_interaction_by_subscript (cat, subscript);
752 const int i = reverse_variable_lookup_short (cat, subscript);
754 const int base_index = cat->iap[i].base_subscript_short;
759 const struct interact_params *iap = &cat->iap[i];
762 for (v = 0; v < iact->n_vars; ++v)
764 const struct variable *var = iact->vars[v];
766 const union value *val = case_data (c, var);
767 const int width = var_get_width (var);
768 const struct variable_node *vn = lookup_variable (&cat->varmap, var, hash_pointer (var, 0));
770 const unsigned int hash = value_hash (val, width, 0);
771 const struct value_node *valn = lookup_value (&vn->valmap, val, hash, width);
775 const double df = iap->df_prod[v] / dfp;
777 /* Translate the subscript into an index for the individual variable */
778 const int index = ((subscript - base_index) % iap->df_prod[v] ) / dfp;
779 dfp = iap->df_prod [v];
781 if (effects_coding && valn->index == df )
783 else if ( valn->index != index )
793 /* Returns unity if the value in case C at SUBSCRIPT is equal to the category
794 for that subscript */
796 categoricals_get_dummy_code_for_case (const struct categoricals *cat, int subscript,
797 const struct ccase *c)
799 return categoricals_get_code_for_case (cat, subscript, c, false);
802 /* Returns unity if the value in case C at SUBSCRIPT is equal to the category
804 Else if it is the last category, return -1.
808 categoricals_get_effects_code_for_case (const struct categoricals *cat, int subscript,
809 const struct ccase *c)
811 return categoricals_get_code_for_case (cat, subscript, c, true);
814 /* Return a case containing the set of values corresponding to
815 the Nth Category of the IACTth interaction */
817 categoricals_get_case_by_category_real (const struct categoricals *cat, int iact, int n)
819 const struct interaction_value *vn;
821 const struct interact_params *vp = &cat->iap[iact];
823 if ( n >= hmap_count (&vp->ivmap))
826 vn = vp->reverse_interaction_value_map [n];
831 /* Return a the user data corresponding to the Nth Category of the IACTth interaction. */
833 categoricals_get_user_data_by_category_real (const struct categoricals *cat, int iact, int n)
835 const struct interact_params *vp = &cat->iap[iact];
836 const struct interaction_value *iv ;
838 if ( n >= hmap_count (&vp->ivmap))
841 iv = vp->reverse_interaction_value_map [n];
843 return iv->user_data;
848 /* Return a case containing the set of values corresponding to SUBSCRIPT */
850 categoricals_get_case_by_category (const struct categoricals *cat, int subscript)
852 int vindex = reverse_variable_lookup_long (cat, subscript);
853 const struct interact_params *vp = &cat->iap[vindex];
854 const struct interaction_value *vn = vp->reverse_interaction_value_map [subscript - vp->base_subscript_long];
860 categoricals_get_user_data_by_category (const struct categoricals *cat, int subscript)
862 int vindex = reverse_variable_lookup_long (cat, subscript);
863 const struct interact_params *vp = &cat->iap[vindex];
865 const struct interaction_value *iv = vp->reverse_interaction_value_map [subscript - vp->base_subscript_long];
866 return iv->user_data;
873 categoricals_set_payload (struct categoricals *cat, const struct payload *p,
874 const void *aux1, void *aux2)