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 /* The interaction, and an array with iact->n_vars elements such that
115 varnodes[x] points to the variable_node for iact->vars[x]. */
116 const struct interaction *iact;
117 struct variable_node **varnodes;
119 /* An example of each interaction that appears in the data, like a frequency
120 table for 'iact'. By construction, the number of elements must be less
121 than or equal to 'n_cats'.
123 categoricals_update() updates 'ivmap' case-by-case, then
124 categoricals_done() dumps 'ivmap' into 'ivs' and sorts it. */
126 struct interaction_value **ivs;
131 /* Product of hmap_count(&varnodes[*]->valmap), that is, the maximum number
132 of distinct values of this interaction. */
135 /* An array of integers df_n * df_{n-1} * df_{n-2} ...
136 These are the products of the degrees of freedom for the current
137 variable and all preceding variables */
142 /* Sum of ivs[*]->cc. */
148 compare_interaction_value_3way (const void *vn1_, const void *vn2_, const void *aux)
150 const struct interaction_value *const *vn1p = vn1_;
151 const struct interaction_value *const *vn2p = vn2_;
153 const struct interact_params *iap = aux;
155 return interaction_case_cmp_3way (iap->iact, (*vn1p)->ccase, (*vn2p)->ccase);
160 /* The weight variable */
161 const struct variable *wv;
163 /* An array of interact_params */
164 struct interact_params *iap;
167 /* Map whose members are the union of the variables which comprise IAP */
170 /* The number of categorical variables which contain entries.
171 In the absence of missing values, this will be equal to N_IAP */
174 /* A map to enable the lookup of variables indexed by subscript.
175 This map considers only the N - 1 of the N variables.
177 int *df_to_iact; /* 'df_sum' elements. */
180 /* Like the above, but uses all N variables */
181 int *cat_to_iact; /* 'n_cats_total' elements. */
186 /* Missing values in the factor variables to be excluded */
187 enum mv_class fctr_excl;
194 const struct payload *payload;
199 categoricals_isbalanced (const struct categoricals *cat)
203 for (i = 0 ; i < cat->n_iap; ++i)
206 const struct interact_params *iap = &cat->iap[i];
209 for (v = 0; v < hmap_count (&iap->ivmap); ++v)
211 const struct interaction_value *iv = iap->ivs[v];
223 categoricals_dump (const struct categoricals *cat)
225 if (CATEGORICALS_DEBUG)
229 printf ("df to interaction map:\n");
230 for (i = 0; i < cat->df_sum; ++i)
232 printf (" %d", cat->df_to_iact[i]);
236 printf ("Category to interaction map:\n");
237 for (i = 0; i < cat->n_cats_total; ++i)
239 printf (" %d", cat->cat_to_iact[i]);
243 printf ("Number of interactions %zu\n", cat->n_iap);
244 for (i = 0 ; i < cat->n_iap; ++i)
248 const struct interact_params *iap = &cat->iap[i];
249 const struct interaction *iact = iap->iact;
251 ds_init_empty (&str);
252 interaction_to_string (iact, &str);
254 printf ("\nInteraction: \"%s\" (number of categories: %d); ", ds_cstr (&str), iap->n_cats);
256 printf ("Base index (df/categories): %d/%d\n", iap->base_df, iap->base_cats);
259 for (v = 0; v < hmap_count (&iap->ivmap); ++v)
262 const struct interaction_value *iv = iap->ivs[v];
264 if (v > 0) printf (" ");
266 for (vv = 0; vv < iact->n_vars; ++vv)
268 const struct variable *var = iact->vars[vv];
269 const union value *val = case_data (iv->ccase, var);
270 struct variable_node *vn = iap->varnodes[vv];
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->df_to_iact = NULL;
370 cat->cat_to_iact = 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 cat->iap[i].varnodes = pool_nmalloc (cat->pool, cat->iap[i].iact->n_vars,
387 sizeof *cat->iap[i].varnodes);
388 for (v = 0; v < inter[i]->n_vars; ++v)
390 const struct variable *var = inter[i]->vars[v];
391 unsigned int hash = hash_pointer (var, 0);
392 struct variable_node *vn = lookup_variable (&cat->varmap, var, hash);
395 vn = pool_malloc (cat->pool, sizeof *vn);
398 hmap_init (&vn->valmap);
400 hmap_insert (&cat->varmap, &vn->node, hash);
402 cat->iap[i].varnodes[v] = vn;
412 categoricals_update (struct categoricals *cat, const struct ccase *c)
415 struct variable_node *vn = NULL;
421 weight = cat->wv ? case_data (c, cat->wv)->f : 1.0;
422 weight = var_force_valid_weight (cat->wv, weight, NULL);
424 assert (NULL == cat->df_to_iact);
425 assert (NULL == cat->cat_to_iact);
427 /* Interate over each variable, and add the value of that variable
428 to the appropriate map, if it's not already present. */
429 HMAP_FOR_EACH (vn, struct variable_node, node, &cat->varmap)
431 const int width = var_get_width (vn->var);
432 const union value *val = case_data (c, vn->var);
433 unsigned int hash = value_hash (val, width, 0);
435 struct value_node *valn = lookup_value (&vn->valmap, val, hash, width);
438 valn = pool_malloc (cat->pool, sizeof *valn);
441 value_init (&valn->val, width);
442 value_copy (&valn->val, val, width);
443 hmap_insert (&vn->valmap, &valn->node, hash);
447 for (i = 0 ; i < cat->n_iap; ++i)
449 const struct interaction *iact = cat->iap[i].iact;
452 struct interaction_value *node;
454 if ( interaction_case_is_missing (iact, c, cat->fctr_excl))
457 hash = interaction_case_hash (iact, c, 0);
458 node = lookup_case (&cat->iap[i].ivmap, iact, c);
462 node = pool_malloc (cat->pool, sizeof *node);
463 node->ccase = case_ref (c);
466 hmap_insert (&cat->iap[i].ivmap, &node->node, hash);
470 node->user_data = cat->payload->create (cat->aux1, cat->aux2);
477 cat->iap[i].cc += weight;
481 cat->payload->update (cat->aux1, cat->aux2, node->user_data, c, weight);
486 /* Return the number of categories (distinct values) for interaction IDX in
489 categoricals_n_count (const struct categoricals *cat, size_t n)
491 return hmap_count (&cat->iap[n].ivmap);
495 /* Returns the number of degrees of freedom for interaction IDX within CAT. */
497 categoricals_df (const struct categoricals *cat, size_t n)
499 const struct interact_params *iap = &cat->iap[n];
500 return iap->df_prod[iap->iact->n_vars - 1];
504 /* Return the total number of categories across all interactions in CAT. */
506 categoricals_n_total (const struct categoricals *cat)
508 if (!categoricals_is_complete (cat))
511 return cat->n_cats_total;
514 /* Returns the total degrees of freedom for CAT. */
516 categoricals_df_total (const struct categoricals *cat)
524 /* Returns true iff categoricals_done() has been called for CAT. */
526 categoricals_is_complete (const struct categoricals *cat)
528 return (NULL != cat->df_to_iact);
532 /* This function must be called (once) before any call to the *_by_subscript or
533 *_by_category functions, but AFTER any calls to categoricals_update. If this
534 function returns false, then no calls to _by_subscript or *_by_category are
537 categoricals_done (const struct categoricals *cat_)
539 /* Implementation Note: Whilst this function is O(n) in cat->n_cats_total, in most
540 uses it will be more efficient that using a tree based structure, since it
541 is called only once, and means that subsequent lookups will be O(1).
543 1 call of O(n) + 10^9 calls of O(1) is better than 10^9 calls of O(log n).
545 struct categoricals *cat = CONST_CAST (struct categoricals *, cat_);
555 cat->n_cats_total = 0;
557 /* Assign 'index' to each variables' value_nodes, counting up from 0 in
558 ascending order by value. */
559 struct variable_node *vn;
560 HMAP_FOR_EACH (vn, struct variable_node, node, &cat->varmap)
562 size_t n_vals = hmap_count (&vn->valmap);
569 struct value_node **nodes = xcalloc (sizeof *nodes, n_vals);
571 struct value_node *valnd;
572 HMAP_FOR_EACH (valnd, struct value_node, node, &vn->valmap)
574 sort (nodes, n_vals, sizeof *nodes, compare_value_node_3way, vn);
575 for (x = 0; x < n_vals; ++x)
580 /* Calculate the degrees of freedom, and the number of categories */
581 for (i = 0 ; i < cat->n_iap; ++i)
584 const struct interaction *iact = cat->iap[i].iact;
586 cat->iap[i].df_prod = iact->n_vars ? xcalloc (iact->n_vars, sizeof (int)) : NULL;
588 cat->iap[i].n_cats = 1;
590 for (v = 0 ; v < iact->n_vars; ++v)
592 struct variable_node *vn = cat->iap[i].varnodes[v];
593 cat->iap[i].df_prod[v] = df * (vn->n_vals - 1);
594 df = cat->iap[i].df_prod[v];
596 cat->iap[i].n_cats *= vn->n_vals;
600 cat->df_sum += cat->iap[i].df_prod [v - 1];
602 cat->n_cats_total += cat->iap[i].n_cats;
606 cat->df_to_iact = pool_calloc (cat->pool, cat->df_sum,
607 sizeof *cat->df_to_iact);
609 cat->cat_to_iact = pool_calloc (cat->pool, cat->n_cats_total,
610 sizeof *cat->cat_to_iact);
612 for (i = 0 ; i < cat->n_iap; ++i)
614 struct interaction_value *ivn = NULL;
617 struct interact_params *iap = &cat->iap[i];
619 iap->base_df = idx_df;
620 iap->base_cats = idx_cat;
622 iap->ivs = pool_calloc (cat->pool, iap->n_cats, sizeof *iap->ivs);
624 HMAP_FOR_EACH (ivn, struct interaction_value, node, &iap->ivmap)
629 assert (x <= iap->n_cats);
631 /* For some purposes (eg CONTRASTS in ONEWAY) the values need to be sorted */
632 sort (iap->ivs, x, sizeof *iap->ivs,
633 compare_interaction_value_3way, iap);
635 /* Fill the remaining values with null */
636 for (ii = x ; ii < iap->n_cats; ++ii)
639 /* Populate the variable maps. */
642 for (ii = 0; ii < iap->df_prod [iap->iact->n_vars - 1]; ++ii)
643 cat->df_to_iact[idx_df++] = i;
646 for (ii = 0; ii < iap->n_cats; ++ii)
647 cat->cat_to_iact[idx_cat++] = i;
650 assert (cat->n_vars <= cat->n_iap);
652 categoricals_dump (cat);
654 /* Tally up the sums for all the encodings */
655 for (i = 0 ; i < cat->n_iap; ++i)
658 struct interact_params *iap = &cat->iap[i];
659 const struct interaction *iact = iap->iact;
661 const int df = iap->df_prod ? iap->df_prod [iact->n_vars - 1] : 0;
663 iap->enc_sum = xcalloc (df, sizeof (*(iap->enc_sum)));
665 for (y = 0; y < hmap_count (&iap->ivmap); ++y)
667 struct interaction_value *iv = iap->ivs[y];
668 for (x = iap->base_df; x < iap->base_df + df ;++x)
670 const double bin = categoricals_get_effects_code_for_case (cat, x, iv->ccase);
671 iap->enc_sum [x - iap->base_df] += bin * iv->cc;
673 if (cat->payload && cat->payload->calculate)
674 cat->payload->calculate (cat->aux1, cat->aux2, iv->user_data);
683 df_to_iap (const struct categoricals *cat, int subscript)
685 assert (cat->df_to_iact);
686 assert (subscript >= 0);
687 assert (subscript < cat->df_sum);
689 return cat->df_to_iact[subscript];
693 cat_index_to_iap (const struct categoricals *cat, int subscript)
695 assert (cat->cat_to_iact);
696 assert (subscript >= 0);
697 assert (subscript < cat->n_cats_total);
699 return cat->cat_to_iact[subscript];
703 /* Return the interaction corresponding to SUBSCRIPT */
704 const struct interaction *
705 categoricals_get_interaction_by_subscript (const struct categoricals *cat, int subscript)
707 int index = df_to_iap (cat, subscript);
709 return cat->iap[index].iact;
713 categoricals_get_weight_by_subscript (const struct categoricals *cat, int subscript)
715 int vindex = df_to_iap (cat, subscript);
716 const struct interact_params *vp = &cat->iap[vindex];
722 categoricals_get_sum_by_subscript (const struct categoricals *cat, int subscript)
724 int vindex = df_to_iap (cat, subscript);
725 const struct interact_params *vp = &cat->iap[vindex];
727 return vp->enc_sum[subscript - vp->base_df];
731 /* Returns unity if the value in case C at SUBSCRIPT is equal to the category
732 for that subscript */
734 categoricals_get_code_for_case (const struct categoricals *cat, int subscript,
735 const struct ccase *c,
738 const struct interaction *iact = categoricals_get_interaction_by_subscript (cat, subscript);
740 const int i = df_to_iap (cat, subscript);
742 const int base_index = cat->iap[i].base_df;
747 const struct interact_params *iap = &cat->iap[i];
750 for (v = 0; v < iact->n_vars; ++v)
752 const struct variable *var = iact->vars[v];
754 const union value *val = case_data (c, var);
755 const int width = var_get_width (var);
756 const struct variable_node *vn = iap->varnodes[v];
758 const unsigned int hash = value_hash (val, width, 0);
759 const struct value_node *valn = lookup_value (&vn->valmap, val, hash, width);
763 const double df = iap->df_prod[v] / dfp;
765 /* Translate the subscript into an index for the individual variable */
766 const int index = ((subscript - base_index) % iap->df_prod[v] ) / dfp;
767 dfp = iap->df_prod [v];
769 if (effects_coding && valn->index == df )
771 else if ( valn->index != index )
781 /* Returns unity if the value in case C at SUBSCRIPT is equal to the category
782 for that subscript */
784 categoricals_get_dummy_code_for_case (const struct categoricals *cat, int subscript,
785 const struct ccase *c)
787 return categoricals_get_code_for_case (cat, subscript, c, false);
790 /* Returns unity if the value in case C at SUBSCRIPT is equal to the category
792 Else if it is the last category, return -1.
796 categoricals_get_effects_code_for_case (const struct categoricals *cat, int subscript,
797 const struct ccase *c)
799 return categoricals_get_code_for_case (cat, subscript, c, true);
802 /* Return a case containing the set of values corresponding to
803 the Nth Category of the IACTth interaction */
805 categoricals_get_case_by_category_real (const struct categoricals *cat, int iact, int n)
807 const struct interaction_value *vn;
809 const struct interact_params *vp = &cat->iap[iact];
811 if ( n >= hmap_count (&vp->ivmap))
819 /* Return a the user data corresponding to the Nth Category of the IACTth interaction. */
821 categoricals_get_user_data_by_category_real (const struct categoricals *cat, int iact, int n)
823 const struct interact_params *vp = &cat->iap[iact];
824 const struct interaction_value *iv ;
826 if ( n >= hmap_count (&vp->ivmap))
831 return iv->user_data;
836 /* Return a case containing the set of values corresponding to SUBSCRIPT */
838 categoricals_get_case_by_category (const struct categoricals *cat, int subscript)
840 int vindex = cat_index_to_iap (cat, subscript);
841 const struct interact_params *vp = &cat->iap[vindex];
842 const struct interaction_value *vn = vp->ivs [subscript - vp->base_cats];
848 categoricals_get_user_data_by_category (const struct categoricals *cat, int subscript)
850 int vindex = cat_index_to_iap (cat, subscript);
851 const struct interact_params *vp = &cat->iap[vindex];
853 const struct interaction_value *iv = vp->ivs [subscript - vp->base_cats];
854 return iv->user_data;
861 categoricals_set_payload (struct categoricals *cat, const struct payload *p,
862 const void *aux1, void *aux2)