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)
107 fprintf (stderr, "%s:%d Warning: Hash table collision\n", __FILE__, __LINE__);
114 struct interact_params
116 /* A map of cases indexed by a interaction_value */
119 const struct interaction *iact;
121 int base_subscript_short;
122 int base_subscript_long;
124 /* The number of distinct values of this interaction */
127 /* An array of integers df_n * df_{n-1} * df_{n-2} ...
128 These are the products of the degrees of freedom for the current
129 variable and all preceding variables */
134 /* A map of interaction_values indexed by subscript */
135 struct interaction_value **reverse_interaction_value_map;
141 /* Comparison function to sort the reverse_value_map in ascending order */
143 compare_interaction_value_3way (const void *vn1_, const void *vn2_, const void *aux)
145 const struct interaction_value *const *vn1p = vn1_;
146 const struct interaction_value *const *vn2p = vn2_;
148 const struct interact_params *iap = aux;
150 return interaction_case_cmp_3way (iap->iact, (*vn1p)->ccase, (*vn2p)->ccase);
155 /* The weight variable */
156 const struct variable *wv;
158 /* An array of interact_params */
159 struct interact_params *iap;
161 /* Map whose members are the union of the variables which comprise IAP */
164 /* The size of IAP. (ie, the number of interactions involved.) */
167 /* The number of categorical variables which contain entries.
168 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;
178 /* Like the above, but uses all N variables */
179 int *reverse_variable_map_long;
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))
339 fprintf (stderr, "Warning: Hash table collision\n");
346 categoricals_sane (const struct categoricals *cat)
351 struct categoricals *
352 categoricals_create (struct interaction *const*inter, size_t n_inter,
353 const struct variable *wv, enum mv_class fctr_excl)
356 struct categoricals *cat = xmalloc (sizeof *cat);
358 cat->n_iap = n_inter;
360 cat->n_cats_total = 0;
362 cat->reverse_variable_map_short = NULL;
363 cat->reverse_variable_map_long = NULL;
364 cat->pool = pool_create ();
365 cat->fctr_excl = fctr_excl;
370 cat->iap = pool_calloc (cat->pool, cat->n_iap, sizeof *cat->iap);
372 hmap_init (&cat->varmap);
373 for (i = 0 ; i < cat->n_iap; ++i)
376 hmap_init (&cat->iap[i].ivmap);
377 cat->iap[i].iact = inter[i];
378 cat->iap[i].cc = 0.0;
379 for (v = 0; v < inter[i]->n_vars; ++v)
381 const struct variable *var = inter[i]->vars[v];
382 unsigned int hash = hash_pointer (var, 0);
383 struct variable_node *vn = lookup_variable (&cat->varmap, var, hash);
386 vn = pool_malloc (cat->pool, sizeof *vn);
389 hmap_init (&vn->valmap);
391 hmap_insert (&cat->varmap, &vn->node, hash);
402 categoricals_update (struct categoricals *cat, const struct ccase *c)
405 struct variable_node *vn = NULL;
411 weight = cat->wv ? case_data (c, cat->wv)->f : 1.0;
412 weight = var_force_valid_weight (cat->wv, weight, NULL);
414 assert (NULL == cat->reverse_variable_map_short);
415 assert (NULL == cat->reverse_variable_map_long);
417 /* Interate over each variable, and add the value of that variable
418 to the appropriate map, if it's not already present. */
419 HMAP_FOR_EACH (vn, struct variable_node, node, &cat->varmap)
421 const int width = var_get_width (vn->var);
422 const union value *val = case_data (c, vn->var);
423 unsigned int hash = value_hash (val, width, 0);
425 struct value_node *valn = lookup_value (&vn->valmap, val, hash, width);
428 valn = pool_malloc (cat->pool, sizeof *valn);
431 value_init (&valn->val, width);
432 value_copy (&valn->val, val, width);
433 hmap_insert (&vn->valmap, &valn->node, hash);
437 for (i = 0 ; i < cat->n_iap; ++i)
439 const struct interaction *iact = cat->iap[i].iact;
442 struct interaction_value *node;
444 if ( interaction_case_is_missing (iact, c, cat->fctr_excl))
447 hash = interaction_case_hash (iact, c, 0);
448 node = lookup_case (&cat->iap[i].ivmap, iact, c);
452 node = pool_malloc (cat->pool, sizeof *node);
453 node->ccase = case_ref (c);
456 hmap_insert (&cat->iap[i].ivmap, &node->node, hash);
460 node->user_data = cat->payload->create (cat->aux1, cat->aux2);
467 cat->iap[i].cc += weight;
471 cat->payload->update (cat->aux1, cat->aux2, node->user_data, c, weight);
476 /* Return the number of categories (distinct values) for interction N */
478 categoricals_n_count (const struct categoricals *cat, size_t n)
480 return hmap_count (&cat->iap[n].ivmap);
485 categoricals_df (const struct categoricals *cat, size_t n)
487 const struct interact_params *iap = &cat->iap[n];
488 return iap->df_prod[iap->iact->n_vars - 1];
492 /* Return the total number of categories */
494 categoricals_n_total (const struct categoricals *cat)
496 if (!categoricals_is_complete (cat))
499 return cat->n_cats_total;
503 categoricals_df_total (const struct categoricals *cat)
512 categoricals_is_complete (const struct categoricals *cat)
514 return (NULL != cat->reverse_variable_map_short);
518 /* This function must be called *before* any call to categoricals_get_*_by subscript and
519 *after* all calls to categoricals_update */
521 categoricals_done (const struct categoricals *cat_)
523 /* Implementation Note: Whilst this function is O(n) in cat->n_cats_total, in most
524 uses it will be more efficient that using a tree based structure, since it
525 is called only once, and means that subsequent lookups will be O(1).
527 1 call of O(n) + 10^9 calls of O(1) is better than 10^9 calls of O(log n).
529 struct categoricals *cat = CONST_CAST (struct categoricals *, cat_);
539 cat->n_cats_total = 0;
541 /* Calculate the degrees of freedom, and the number of categories */
542 for (i = 0 ; i < cat->n_iap; ++i)
545 const struct interaction *iact = cat->iap[i].iact;
547 cat->iap[i].df_prod = iact->n_vars ? xcalloc (iact->n_vars, sizeof (int)) : NULL;
549 cat->iap[i].n_cats = 1;
551 for (v = 0 ; v < iact->n_vars; ++v)
554 const struct variable *var = iact->vars[v];
556 struct variable_node *vn = lookup_variable (&cat->varmap, var, hash_pointer (var, 0));
558 struct value_node *valnd = NULL;
559 struct value_node **array ;
561 assert (vn->n_vals == hmap_count (&vn->valmap));
569 /* Sort the VALMAP here */
570 array = xcalloc (sizeof *array, vn->n_vals);
572 HMAP_FOR_EACH (valnd, struct value_node, node, &vn->valmap)
574 /* Note: This loop is probably superfluous, it could be done in the
575 update stage (at the expense of a realloc) */
579 sort (array, vn->n_vals, sizeof (*array),
580 compare_value_node_3way, vn);
582 for (x = 0; x < vn->n_vals; ++x)
584 struct value_node *vvv = array[x];
589 cat->iap[i].df_prod[v] = df * (vn->n_vals - 1);
590 df = cat->iap[i].df_prod[v];
592 cat->iap[i].n_cats *= vn->n_vals;
596 cat->df_sum += cat->iap[i].df_prod [v - 1];
598 cat->n_cats_total += cat->iap[i].n_cats;
602 cat->reverse_variable_map_short = pool_calloc (cat->pool,
604 sizeof *cat->reverse_variable_map_short);
606 cat->reverse_variable_map_long = pool_calloc (cat->pool,
608 sizeof *cat->reverse_variable_map_long);
610 for (i = 0 ; i < cat->n_iap; ++i)
612 struct interaction_value *ivn = NULL;
615 struct interact_params *iap = &cat->iap[i];
617 iap->base_subscript_short = idx_short;
618 iap->base_subscript_long = idx_long;
620 iap->reverse_interaction_value_map = pool_calloc (cat->pool, iap->n_cats,
621 sizeof *iap->reverse_interaction_value_map);
623 HMAP_FOR_EACH (ivn, struct interaction_value, node, &iap->ivmap)
625 iap->reverse_interaction_value_map[x++] = ivn;
628 assert (x <= iap->n_cats);
630 /* For some purposes (eg CONTRASTS in ONEWAY) the values need to be sorted */
631 sort (iap->reverse_interaction_value_map, x, sizeof (*iap->reverse_interaction_value_map),
632 compare_interaction_value_3way, iap);
634 /* Fill the remaining values with null */
635 for (ii = x ; ii < iap->n_cats; ++ii)
636 iap->reverse_interaction_value_map[ii] = NULL;
638 /* Populate the reverse variable maps. */
641 for (ii = 0; ii < iap->df_prod [iap->iact->n_vars - 1]; ++ii)
642 cat->reverse_variable_map_short[idx_short++] = i;
645 for (ii = 0; ii < iap->n_cats; ++ii)
646 cat->reverse_variable_map_long[idx_long++] = i;
649 assert (cat->n_vars <= cat->n_iap);
651 categoricals_dump (cat);
653 /* Tally up the sums for all the encodings */
654 for (i = 0 ; i < cat->n_iap; ++i)
657 struct interact_params *iap = &cat->iap[i];
658 const struct interaction *iact = iap->iact;
660 const int df = iap->df_prod ? iap->df_prod [iact->n_vars - 1] : 0;
662 iap->enc_sum = xcalloc (df, sizeof (*(iap->enc_sum)));
664 for (y = 0; y < hmap_count (&iap->ivmap); ++y)
666 struct interaction_value *iv = iap->reverse_interaction_value_map[y];
667 for (x = iap->base_subscript_short; x < iap->base_subscript_short + df ;++x)
669 const double bin = categoricals_get_effects_code_for_case (cat, x, iv->ccase);
670 iap->enc_sum [x - iap->base_subscript_short] += bin * iv->cc;
672 if (cat->payload && cat->payload->calculate)
673 cat->payload->calculate (cat->aux1, cat->aux2, iv->user_data);
682 reverse_variable_lookup_short (const struct categoricals *cat, int subscript)
684 assert (cat->reverse_variable_map_short);
685 assert (subscript >= 0);
686 assert (subscript < cat->df_sum);
688 return cat->reverse_variable_map_short[subscript];
692 reverse_variable_lookup_long (const struct categoricals *cat, int subscript)
694 assert (cat->reverse_variable_map_long);
695 assert (subscript >= 0);
696 assert (subscript < cat->n_cats_total);
698 return cat->reverse_variable_map_long[subscript];
702 /* Return the interaction corresponding to SUBSCRIPT */
703 const struct interaction *
704 categoricals_get_interaction_by_subscript (const struct categoricals *cat, int subscript)
706 int index = reverse_variable_lookup_short (cat, subscript);
708 return cat->iap[index].iact;
712 categoricals_get_weight_by_subscript (const struct categoricals *cat, int subscript)
714 int vindex = reverse_variable_lookup_short (cat, subscript);
715 const struct interact_params *vp = &cat->iap[vindex];
721 categoricals_get_sum_by_subscript (const struct categoricals *cat, int subscript)
723 int vindex = reverse_variable_lookup_short (cat, subscript);
724 const struct interact_params *vp = &cat->iap[vindex];
726 return vp->enc_sum[subscript - vp->base_subscript_short];
730 /* Returns unity if the value in case C at SUBSCRIPT is equal to the category
731 for that subscript */
733 categoricals_get_code_for_case (const struct categoricals *cat, int subscript,
734 const struct ccase *c,
737 const struct interaction *iact = categoricals_get_interaction_by_subscript (cat, subscript);
739 const int i = reverse_variable_lookup_short (cat, subscript);
741 const int base_index = cat->iap[i].base_subscript_short;
746 const struct interact_params *iap = &cat->iap[i];
749 for (v = 0; v < iact->n_vars; ++v)
751 const struct variable *var = iact->vars[v];
753 const union value *val = case_data (c, var);
754 const int width = var_get_width (var);
755 const struct variable_node *vn = lookup_variable (&cat->varmap, var, hash_pointer (var, 0));
757 const unsigned int hash = value_hash (val, width, 0);
758 const struct value_node *valn = lookup_value (&vn->valmap, val, hash, width);
762 const double df = iap->df_prod[v] / dfp;
764 /* Translate the subscript into an index for the individual variable */
765 const int index = ((subscript - base_index) % iap->df_prod[v] ) / dfp;
766 dfp = iap->df_prod [v];
768 if (effects_coding && valn->index == df )
770 else if ( valn->index != index )
780 /* Returns unity if the value in case C at SUBSCRIPT is equal to the category
781 for that subscript */
783 categoricals_get_dummy_code_for_case (const struct categoricals *cat, int subscript,
784 const struct ccase *c)
786 return categoricals_get_code_for_case (cat, subscript, c, false);
789 /* Returns unity if the value in case C at SUBSCRIPT is equal to the category
791 Else if it is the last category, return -1.
795 categoricals_get_effects_code_for_case (const struct categoricals *cat, int subscript,
796 const struct ccase *c)
798 return categoricals_get_code_for_case (cat, subscript, c, true);
803 categoricals_get_n_variables (const struct categoricals *cat)
805 printf ("%s\n", __FUNCTION__);
810 /* Return a case containing the set of values corresponding to
811 the Nth Category of the IACTth interaction */
813 categoricals_get_case_by_category_real (const struct categoricals *cat, int iact, int n)
815 const struct interaction_value *vn;
817 const struct interact_params *vp = &cat->iap[iact];
819 if ( n >= hmap_count (&vp->ivmap))
822 vn = vp->reverse_interaction_value_map [n];
827 /* Return a the user data corresponding to the Nth Category of the IACTth interaction. */
829 categoricals_get_user_data_by_category_real (const struct categoricals *cat, int iact, int n)
831 const struct interact_params *vp = &cat->iap[iact];
832 const struct interaction_value *iv ;
834 if ( n >= hmap_count (&vp->ivmap))
837 iv = vp->reverse_interaction_value_map [n];
839 return iv->user_data;
844 /* Return a case containing the set of values corresponding to SUBSCRIPT */
846 categoricals_get_case_by_category (const struct categoricals *cat, int subscript)
848 int vindex = reverse_variable_lookup_long (cat, subscript);
849 const struct interact_params *vp = &cat->iap[vindex];
850 const struct interaction_value *vn = vp->reverse_interaction_value_map [subscript - vp->base_subscript_long];
856 categoricals_get_user_data_by_category (const struct categoricals *cat, int subscript)
858 int vindex = reverse_variable_lookup_long (cat, subscript);
859 const struct interact_params *vp = &cat->iap[vindex];
861 const struct interaction_value *iv = vp->reverse_interaction_value_map [subscript - vp->base_subscript_long];
862 return iv->user_data;
869 categoricals_set_payload (struct categoricals *cat, const struct payload *p,
870 const void *aux1, void *aux2)