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
39 struct hmap_node node; /* Node in hash map. */
41 union value val; /* The value */
43 int index; /* A zero based unique index for this value */
46 struct interaction_value
48 struct hmap_node node; /* Node in hash map */
50 struct ccase *ccase; /* A case (probably the first in the dataset) which matches this value */
52 double cc; /* Total of the weights of cases matching this interaction */
54 void *user_data; /* A pointer to data which the caller can store stuff */
57 static struct value_node *
58 lookup_value (const struct hmap *map, const union value *val, unsigned int hash, int width)
60 struct value_node *vn = NULL;
61 HMAP_FOR_EACH_WITH_HASH (vn, struct value_node, node, hash, map)
63 if (value_equal (&vn->val, val, width))
72 struct hmap_node node; /* Node in hash map. */
73 const struct variable *var; /* The variable */
75 struct hmap valmap; /* A map of value nodes */
76 int n_vals; /* Number of values for this variable */
79 static struct variable_node *
80 lookup_variable (const struct hmap *map, const struct variable *var, unsigned int hash)
82 struct variable_node *vn = NULL;
83 HMAP_FOR_EACH_WITH_HASH (vn, struct variable_node, node, hash, map)
88 fprintf (stderr, "Warning: Hash table collision\n");
95 struct interact_params
97 /* A map of cases indexed by a interaction_value */
100 struct interaction *iact;
102 int base_subscript_short;
103 int base_subscript_long;
105 /* The number of distinct values of this interaction */
108 /* An array of integers df_n * df_{n-1} * df_{n-2} ...
109 These are the products of the degrees of freedom for the current
110 variable and all preceeding variables */
115 /* A map of interaction_values indexed by subscript */
116 struct interaction_value **reverse_interaction_value_map;
122 /* Comparison function to sort the reverse_value_map in ascending order */
124 compare_interaction_value_3way (const void *vn1_, const void *vn2_, const void *aux)
126 const struct interaction_value *const *vn1p = vn1_;
127 const struct interaction_value *const *vn2p = vn2_;
129 const struct interact_params *iap = aux;
131 return interaction_case_cmp_3way (iap->iact, (*vn1p)->ccase, (*vn2p)->ccase);
136 /* The weight variable */
137 const struct variable *wv;
139 /* An array of interact_params */
140 struct interact_params *iap;
142 /* Map whose members are the union of the variables which comprise IAP */
145 /* The size of IAP. (ie, the number of interactions involved.) */
148 /* The number of categorical variables which contain entries.
149 In the absence of missing values, this will be equal to N_IAP */
154 /* A map to enable the lookup of variables indexed by subscript.
155 This map considers only the N - 1 of the N variables.
157 int *reverse_variable_map_short;
159 /* Like the above, but uses all N variables */
160 int *reverse_variable_map_long;
166 /* Missing values in the dependent varirable to be excluded */
167 enum mv_class dep_excl;
169 /* Missing values in the factor variables to be excluded */
170 enum mv_class fctr_excl;
177 const struct payload *payload;
181 categoricals_dump (const struct categoricals *cat)
183 if (CATEGORICALS_DEBUG)
187 printf ("Reverse Variable Map (short):\n");
188 for (i = 0; i < cat->df_sum; ++i)
190 printf (" %d", cat->reverse_variable_map_short[i]);
194 printf ("Reverse Variable Map (long):\n");
195 for (i = 0; i < cat->n_cats_total; ++i)
197 printf (" %d", cat->reverse_variable_map_long[i]);
202 printf ("Number of interactions %d\n", cat->n_iap);
203 for (i = 0 ; i < cat->n_iap; ++i)
207 const struct interact_params *iap = &cat->iap[i];
208 const struct interaction *iact = iap->iact;
210 ds_init_empty (&str);
211 interaction_to_string (iact, &str);
213 printf ("\nInteraction: %s (n: %d); ", ds_cstr (&str), iap->n_cats);
215 printf ("Base subscript: %d\n", iap->base_subscript_short);
218 for (v = 0; v < hmap_count (&iap->ivmap); ++v)
221 const struct interaction_value *iv = iap->reverse_interaction_value_map[v];
223 if (v > 0) printf (" ");
225 for (vv = 0; vv < iact->n_vars; ++vv)
227 const struct variable *var = iact->vars[vv];
228 const union value *val = case_data (iv->ccase, var);
230 printf ("%g", val->f);
231 if (vv < iact->n_vars - 1)
242 categoricals_destroy (struct categoricals *cat)
244 struct variable_node *vn = NULL;
248 for (i = 0; i < cat->n_iap; ++i)
250 struct interaction_value *iv = NULL;
251 /* Interate over each interaction value, and unref any cases that we reffed */
252 HMAP_FOR_EACH (iv, struct interaction_value, node, &cat->iap[i].ivmap)
254 if (cat->payload && cat->payload->destroy)
255 cat->payload->destroy (cat->aux1, cat->aux2, iv->user_data);
256 case_unref (iv->ccase);
259 free (cat->iap[i].enc_sum);
260 free (cat->iap[i].df_prod);
261 hmap_destroy (&cat->iap[i].ivmap);
262 interaction_destroy (cat->iap[i].iact);
265 /* Interate over each variable and delete its value map */
266 HMAP_FOR_EACH (vn, struct variable_node, node, &cat->varmap)
268 hmap_destroy (&vn->valmap);
271 hmap_destroy (&cat->varmap);
273 pool_destroy (cat->pool);
280 static struct interaction_value *
281 lookup_case (const struct hmap *map, const struct interaction *iact, const struct ccase *c)
283 struct interaction_value *iv = NULL;
284 size_t hash = interaction_case_hash (iact, c, 0);
286 HMAP_FOR_EACH_WITH_HASH (iv, struct interaction_value, node, hash, map)
288 if (interaction_case_equal (iact, c, iv->ccase))
291 fprintf (stderr, "Warning: Hash table collision\n");
298 categoricals_sane (const struct categoricals *cat)
303 struct categoricals *
304 categoricals_create (struct interaction *const*inter, size_t n_inter,
305 const struct variable *wv, enum mv_class dep_excl, enum mv_class fctr_excl)
308 struct categoricals *cat = xmalloc (sizeof *cat);
310 cat->n_iap = n_inter;
312 cat->n_cats_total = 0;
314 cat->reverse_variable_map_short = NULL;
315 cat->reverse_variable_map_long = NULL;
316 cat->pool = pool_create ();
317 cat->dep_excl = dep_excl;
318 cat->fctr_excl = fctr_excl;
323 cat->iap = pool_calloc (cat->pool, cat->n_iap, sizeof *cat->iap);
325 hmap_init (&cat->varmap);
326 for (i = 0 ; i < cat->n_iap; ++i)
329 hmap_init (&cat->iap[i].ivmap);
330 cat->iap[i].iact = inter[i];
331 cat->iap[i].cc = 0.0;
332 for (v = 0; v < inter[i]->n_vars; ++v)
334 const struct variable *var = inter[i]->vars[v];
335 unsigned int hash = hash_pointer (var, 0);
336 struct variable_node *vn = lookup_variable (&cat->varmap, var, hash);
339 vn = pool_malloc (cat->pool, sizeof *vn);
342 hmap_init (&vn->valmap);
344 hmap_insert (&cat->varmap, &vn->node, hash);
355 categoricals_update (struct categoricals *cat, const struct ccase *c)
358 struct variable_node *vn = NULL;
364 weight = cat->wv ? case_data (c, cat->wv)->f : 1.0;
366 assert (NULL == cat->reverse_variable_map_short);
367 assert (NULL == cat->reverse_variable_map_long);
369 /* Interate over each variable, and add the value of that variable
370 to the appropriate map, if it's not already present. */
371 HMAP_FOR_EACH (vn, struct variable_node, node, &cat->varmap)
373 const int width = var_get_width (vn->var);
374 const union value *val = case_data (c, vn->var);
375 unsigned int hash = value_hash (val, width, 0);
377 struct value_node *valn = lookup_value (&vn->valmap, val, hash, width);
380 valn = pool_malloc (cat->pool, sizeof *valn);
381 valn->index = vn->n_vals++;
382 value_init (&valn->val, width);
383 value_copy (&valn->val, val, width);
384 hmap_insert (&vn->valmap, &valn->node, hash);
388 for (i = 0 ; i < cat->n_iap; ++i)
390 const struct interaction *iact = cat->iap[i].iact;
393 struct interaction_value *node;
395 if ( interaction_case_is_missing (iact, c, cat->fctr_excl))
398 hash = interaction_case_hash (iact, c, 0);
399 node = lookup_case (&cat->iap[i].ivmap, iact, c);
403 node = pool_malloc (cat->pool, sizeof *node);
405 node->ccase = case_ref (c);
408 hmap_insert (&cat->iap[i].ivmap, &node->node, hash);
412 node->user_data = cat->payload->create (cat->aux1, cat->aux2);
419 cat->iap[i].cc += weight;
423 double weight = cat->wv ? case_data (c, cat->wv)->f : 1.0;
424 cat->payload->update (cat->aux1, cat->aux2, node->user_data, c, weight);
430 /* Return the number of categories (distinct values) for interction N */
432 categoricals_n_count (const struct categoricals *cat, size_t n)
434 return hmap_count (&cat->iap[n].ivmap);
439 categoricals_df (const struct categoricals *cat, size_t n)
441 const struct interact_params *iap = &cat->iap[n];
442 return iap->df_prod[iap->iact->n_vars - 1];
446 /* Return the total number of categories */
448 categoricals_n_total (const struct categoricals *cat)
450 if (!categoricals_is_complete (cat))
453 return cat->n_cats_total;
457 categoricals_df_total (const struct categoricals *cat)
466 categoricals_is_complete (const struct categoricals *cat)
468 return (NULL != cat->reverse_variable_map_short);
472 /* This function must be called *before* any call to categoricals_get_*_by subscript and
473 *after* all calls to categoricals_update */
475 categoricals_done (const struct categoricals *cat_)
477 /* Implementation Note: Whilst this function is O(n) in cat->n_cats_total, in most
478 uses it will be more efficient that using a tree based structure, since it
479 is called only once, and means that subsequent lookups will be O(1).
481 1 call of O(n) + 10^9 calls of O(1) is better than 10^9 calls of O(log n).
483 struct categoricals *cat = CONST_CAST (struct categoricals *, cat_);
493 cat->n_cats_total = 0;
495 /* Calculate the degrees of freedom, and the number of categories */
496 for (i = 0 ; i < cat->n_iap; ++i)
499 const struct interaction *iact = cat->iap[i].iact;
501 cat->iap[i].df_prod = iact->n_vars ? xcalloc (iact->n_vars, sizeof (int)) : NULL;
503 cat->iap[i].n_cats = 1;
505 for (v = 0 ; v < iact->n_vars; ++v)
507 const struct variable *var = iact->vars[v];
509 struct variable_node *vn = lookup_variable (&cat->varmap, var, hash_pointer (var, 0));
511 if (hmap_count (&vn->valmap) == 0)
517 cat->iap[i].df_prod[v] = df * (hmap_count (&vn->valmap) - 1);
518 df = cat->iap[i].df_prod[v];
520 cat->iap[i].n_cats *= hmap_count (&vn->valmap);
523 assert (v == iact->n_vars);
525 cat->df_sum += cat->iap[i].df_prod [v - 1];
527 cat->n_cats_total += cat->iap[i].n_cats;
531 cat->reverse_variable_map_short = pool_calloc (cat->pool,
533 sizeof *cat->reverse_variable_map_short);
535 cat->reverse_variable_map_long = pool_calloc (cat->pool,
537 sizeof *cat->reverse_variable_map_long);
539 for (i = 0 ; i < cat->n_iap; ++i)
541 struct interaction_value *ivn = NULL;
544 struct interact_params *iap = &cat->iap[i];
546 iap->base_subscript_short = idx_short;
547 iap->base_subscript_long = idx_long;
549 iap->reverse_interaction_value_map = pool_calloc (cat->pool, iap->n_cats,
550 sizeof *iap->reverse_interaction_value_map);
552 HMAP_FOR_EACH (ivn, struct interaction_value, node, &iap->ivmap)
554 iap->reverse_interaction_value_map[x++] = ivn;
557 assert (x <= iap->n_cats);
559 /* For some purposes (eg CONTRASTS in ONEWAY) the values need to be sorted */
560 sort (iap->reverse_interaction_value_map, x, sizeof (*iap->reverse_interaction_value_map),
561 compare_interaction_value_3way, iap);
563 /* Fill the remaining values with null */
564 for (ii = x ; ii < iap->n_cats; ++ii)
565 iap->reverse_interaction_value_map[ii] = NULL;
567 /* Populate the reverse variable maps. */
570 for (ii = 0; ii < iap->df_prod [iap->iact->n_vars - 1]; ++ii)
571 cat->reverse_variable_map_short[idx_short++] = i;
574 for (ii = 0; ii < iap->n_cats; ++ii)
575 cat->reverse_variable_map_long[idx_long++] = i;
578 assert (cat->n_vars <= cat->n_iap);
580 categoricals_dump (cat);
582 /* Tally up the sums for all the encodings */
583 for (i = 0 ; i < cat->n_iap; ++i)
586 struct interact_params *iap = &cat->iap[i];
587 const struct interaction *iact = iap->iact;
589 const int df = iap->df_prod ? iap->df_prod [iact->n_vars - 1] : 0;
591 iap->enc_sum = xcalloc (df, sizeof (*(iap->enc_sum)));
593 for (y = 0; y < hmap_count (&iap->ivmap); ++y)
595 struct interaction_value *iv = iap->reverse_interaction_value_map[y];
596 for (x = iap->base_subscript_short; x < iap->base_subscript_short + df ;++x)
598 const double bin = categoricals_get_effects_code_for_case (cat, x, iv->ccase);
599 iap->enc_sum [x - iap->base_subscript_short] += bin * iv->cc;
601 if (cat->payload && cat->payload->calculate)
602 cat->payload->calculate (cat->aux1, cat->aux2, iv->user_data);
611 reverse_variable_lookup_short (const struct categoricals *cat, int subscript)
613 assert (cat->reverse_variable_map_short);
614 assert (subscript >= 0);
615 assert (subscript < cat->df_sum);
617 return cat->reverse_variable_map_short[subscript];
621 reverse_variable_lookup_long (const struct categoricals *cat, int subscript)
623 assert (cat->reverse_variable_map_long);
624 assert (subscript >= 0);
625 assert (subscript < cat->n_cats_total);
627 return cat->reverse_variable_map_long[subscript];
631 /* Return the interaction corresponding to SUBSCRIPT */
632 const struct interaction *
633 categoricals_get_interaction_by_subscript (const struct categoricals *cat, int subscript)
635 int index = reverse_variable_lookup_short (cat, subscript);
637 return cat->iap[index].iact;
641 categoricals_get_weight_by_subscript (const struct categoricals *cat, int subscript)
643 int vindex = reverse_variable_lookup_short (cat, subscript);
644 const struct interact_params *vp = &cat->iap[vindex];
650 categoricals_get_sum_by_subscript (const struct categoricals *cat, int subscript)
652 int vindex = reverse_variable_lookup_short (cat, subscript);
653 const struct interact_params *vp = &cat->iap[vindex];
655 return vp->enc_sum[subscript - vp->base_subscript_short];
659 /* Returns unity if the value in case C at SUBSCRIPT is equal to the category
660 for that subscript */
662 categoricals_get_code_for_case (const struct categoricals *cat, int subscript,
663 const struct ccase *c,
666 const struct interaction *iact = categoricals_get_interaction_by_subscript (cat, subscript);
668 const int i = reverse_variable_lookup_short (cat, subscript);
670 const int base_index = cat->iap[i].base_subscript_short;
675 const struct interact_params *iap = &cat->iap[i];
678 for (v = 0; v < iact->n_vars; ++v)
680 const struct variable *var = iact->vars[v];
682 const union value *val = case_data (c, var);
683 const int width = var_get_width (var);
684 const struct variable_node *vn = lookup_variable (&cat->varmap, var, hash_pointer (var, 0));
686 const unsigned int hash = value_hash (val, width, 0);
687 const struct value_node *valn = lookup_value (&vn->valmap, val, hash, width);
691 const double df = iap->df_prod[v] / dfp;
693 /* Translate the subscript into an index for the individual variable */
694 const int index = ((subscript - base_index) % iap->df_prod[v] ) / dfp;
695 dfp = iap->df_prod [v];
697 if (effects_coding && valn->index == df )
699 else if ( valn->index != index )
709 /* Returns unity if the value in case C at SUBSCRIPT is equal to the category
710 for that subscript */
712 categoricals_get_dummy_code_for_case (const struct categoricals *cat, int subscript,
713 const struct ccase *c)
715 return categoricals_get_code_for_case (cat, subscript, c, false);
718 /* Returns unity if the value in case C at SUBSCRIPT is equal to the category
720 Else if it is the last category, return -1.
724 categoricals_get_effects_code_for_case (const struct categoricals *cat, int subscript,
725 const struct ccase *c)
727 return categoricals_get_code_for_case (cat, subscript, c, true);
732 categoricals_get_n_variables (const struct categoricals *cat)
734 printf ("%s\n", __FUNCTION__);
739 /* Return a case containing the set of values corresponding to
740 the Nth Category of the IACTth interaction */
742 categoricals_get_case_by_category_real (const struct categoricals *cat, int iact, int n)
744 const struct interaction_value *vn;
746 const struct interact_params *vp = &cat->iap[iact];
748 if ( n >= hmap_count (&vp->ivmap))
751 vn = vp->reverse_interaction_value_map [n];
756 /* Return a the user data corresponding to the Nth Category of the IACTth interaction. */
758 categoricals_get_user_data_by_category_real (const struct categoricals *cat, int iact, int n)
760 const struct interact_params *vp = &cat->iap[iact];
761 const struct interaction_value *iv ;
763 if ( n >= hmap_count (&vp->ivmap))
766 iv = vp->reverse_interaction_value_map [n];
768 return iv->user_data;
773 /* Return a case containing the set of values corresponding to SUBSCRIPT */
775 categoricals_get_case_by_category (const struct categoricals *cat, int subscript)
777 int vindex = reverse_variable_lookup_long (cat, subscript);
778 const struct interact_params *vp = &cat->iap[vindex];
779 const struct interaction_value *vn = vp->reverse_interaction_value_map [subscript - vp->base_subscript_long];
785 categoricals_get_user_data_by_category (const struct categoricals *cat, int subscript)
787 int vindex = reverse_variable_lookup_long (cat, subscript);
788 const struct interact_params *vp = &cat->iap[vindex];
790 const struct interaction_value *iv = vp->reverse_interaction_value_map [subscript - vp->base_subscript_long];
791 return iv->user_data;
798 categoricals_set_payload (struct categoricals *cat, const struct payload *p,
799 const void *aux1, void *aux2)