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 /* An example of each interaction that appears in the data, like a frequency
117 table for 'iact'. By construction, the number of elements must be less
118 than or equal to 'n_cats'.
120 categoricals_update() updates 'ivmap' case-by-case, then
121 categoricals_done() dumps 'ivmap' into 'reverse_interaction_value_map' and
124 struct interaction_value **reverse_interaction_value_map;
126 const struct interaction *iact;
128 int base_subscript_short;
129 int base_subscript_long;
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 reverse_interaction_value_map[*]->cc. */
147 /* Comparison function to sort the reverse_value_map in ascending order */
149 compare_interaction_value_3way (const void *vn1_, const void *vn2_, const void *aux)
151 const struct interaction_value *const *vn1p = vn1_;
152 const struct interaction_value *const *vn2p = vn2_;
154 const struct interact_params *iap = aux;
156 return interaction_case_cmp_3way (iap->iact, (*vn1p)->ccase, (*vn2p)->ccase);
161 /* The weight variable */
162 const struct variable *wv;
164 /* An array of interact_params */
165 struct interact_params *iap;
168 /* Map whose members are the union of the variables which comprise IAP */
171 /* The number of categorical variables which contain entries.
172 In the absence of missing values, this will be equal to N_IAP */
175 /* A map to enable the lookup of variables indexed by subscript.
176 This map considers only the N - 1 of the N variables.
178 int *reverse_variable_map_short; /* 'df_sum' elements. */
181 /* Like the above, but uses all N variables */
182 int *reverse_variable_map_long; /* 'n_cats_total' elements. */
187 /* Missing values in the factor variables to be excluded */
188 enum mv_class fctr_excl;
195 const struct payload *payload;
200 categoricals_isbalanced (const struct categoricals *cat)
204 for (i = 0 ; i < cat->n_iap; ++i)
207 const struct interact_params *iap = &cat->iap[i];
210 for (v = 0; v < hmap_count (&iap->ivmap); ++v)
212 const struct interaction_value *iv = iap->reverse_interaction_value_map[v];
224 categoricals_dump (const struct categoricals *cat)
226 if (CATEGORICALS_DEBUG)
230 printf ("Reverse Variable Map (short):\n");
231 for (i = 0; i < cat->df_sum; ++i)
233 printf (" %d", cat->reverse_variable_map_short[i]);
237 printf ("Reverse Variable Map (long):\n");
238 for (i = 0; i < cat->n_cats_total; ++i)
240 printf (" %d", cat->reverse_variable_map_long[i]);
244 printf ("Number of interactions %zu\n", cat->n_iap);
245 for (i = 0 ; i < cat->n_iap; ++i)
249 const struct interact_params *iap = &cat->iap[i];
250 const struct interaction *iact = iap->iact;
252 ds_init_empty (&str);
253 interaction_to_string (iact, &str);
255 printf ("\nInteraction: \"%s\" (number of categories: %d); ", ds_cstr (&str), iap->n_cats);
257 printf ("Base index (short/long): %d/%d\n", iap->base_subscript_short, iap->base_subscript_long);
260 for (v = 0; v < hmap_count (&iap->ivmap); ++v)
263 const struct interaction_value *iv = iap->reverse_interaction_value_map[v];
265 if (v > 0) printf (" ");
267 for (vv = 0; vv < iact->n_vars; ++vv)
269 const struct variable *var = iact->vars[vv];
270 const union value *val = case_data (iv->ccase, var);
271 unsigned int varhash = hash_pointer (var, 0);
272 struct variable_node *vn = lookup_variable (&cat->varmap, var, varhash);
274 const int width = var_get_width (var);
275 unsigned int valhash = value_hash (val, width, 0);
276 struct value_node *valn = lookup_value (&vn->valmap, val, valhash, width);
278 assert (vn->var == var);
280 printf ("%.*g(%d)", DBL_DIG + 1, val->f, valn->index);
281 if (vv < iact->n_vars - 1)
292 categoricals_destroy (struct categoricals *cat)
294 struct variable_node *vn = NULL;
299 for (i = 0; i < cat->n_iap; ++i)
301 struct interaction_value *iv = NULL;
302 /* Interate over each interaction value, and unref any cases that we reffed */
303 HMAP_FOR_EACH (iv, struct interaction_value, node, &cat->iap[i].ivmap)
305 if (cat->payload && cat->payload->destroy)
306 cat->payload->destroy (cat->aux1, cat->aux2, iv->user_data);
307 case_unref (iv->ccase);
310 free (cat->iap[i].enc_sum);
311 free (cat->iap[i].df_prod);
312 hmap_destroy (&cat->iap[i].ivmap);
315 /* Interate over each variable and delete its value map */
316 HMAP_FOR_EACH (vn, struct variable_node, node, &cat->varmap)
318 hmap_destroy (&vn->valmap);
321 hmap_destroy (&cat->varmap);
323 pool_destroy (cat->pool);
330 static struct interaction_value *
331 lookup_case (const struct hmap *map, const struct interaction *iact, const struct ccase *c)
333 struct interaction_value *iv = NULL;
334 size_t hash = interaction_case_hash (iact, c, 0);
336 HMAP_FOR_EACH_WITH_HASH (iv, struct interaction_value, node, hash, map)
338 if (interaction_case_equal (iact, c, iv->ccase))
341 fprintf (stderr, "Warning: Hash table collision\n");
347 /* Returns true iff CAT is sane, that is, if it is complete and has at least
350 categoricals_sane (const struct categoricals *cat)
355 /* Creates and returns a new categoricals object whose variables come from the
356 N_INTER interactions objects in the array starting at INTER. (The INTER
357 objects must outlive the categoricals object because it uses them
360 FCTR_EXCL determines which cases are listwise ignored by
361 categoricals_update(). */
362 struct categoricals *
363 categoricals_create (struct interaction *const*inter, size_t n_inter,
364 const struct variable *wv, enum mv_class fctr_excl)
367 struct categoricals *cat = xmalloc (sizeof *cat);
369 cat->n_iap = n_inter;
371 cat->n_cats_total = 0;
373 cat->reverse_variable_map_short = NULL;
374 cat->reverse_variable_map_long = NULL;
375 cat->pool = pool_create ();
376 cat->fctr_excl = fctr_excl;
381 cat->iap = pool_calloc (cat->pool, cat->n_iap, sizeof *cat->iap);
383 hmap_init (&cat->varmap);
384 for (i = 0 ; i < cat->n_iap; ++i)
387 hmap_init (&cat->iap[i].ivmap);
388 cat->iap[i].iact = inter[i];
389 cat->iap[i].cc = 0.0;
390 for (v = 0; v < inter[i]->n_vars; ++v)
392 const struct variable *var = inter[i]->vars[v];
393 unsigned int hash = hash_pointer (var, 0);
394 struct variable_node *vn = lookup_variable (&cat->varmap, var, hash);
397 vn = pool_malloc (cat->pool, sizeof *vn);
400 hmap_init (&vn->valmap);
402 hmap_insert (&cat->varmap, &vn->node, hash);
413 categoricals_update (struct categoricals *cat, const struct ccase *c)
416 struct variable_node *vn = NULL;
422 weight = cat->wv ? case_data (c, cat->wv)->f : 1.0;
423 weight = var_force_valid_weight (cat->wv, weight, NULL);
425 assert (NULL == cat->reverse_variable_map_short);
426 assert (NULL == cat->reverse_variable_map_long);
428 /* Interate over each variable, and add the value of that variable
429 to the appropriate map, if it's not already present. */
430 HMAP_FOR_EACH (vn, struct variable_node, node, &cat->varmap)
432 const int width = var_get_width (vn->var);
433 const union value *val = case_data (c, vn->var);
434 unsigned int hash = value_hash (val, width, 0);
436 struct value_node *valn = lookup_value (&vn->valmap, val, hash, width);
439 valn = pool_malloc (cat->pool, sizeof *valn);
442 value_init (&valn->val, width);
443 value_copy (&valn->val, val, width);
444 hmap_insert (&vn->valmap, &valn->node, hash);
448 for (i = 0 ; i < cat->n_iap; ++i)
450 const struct interaction *iact = cat->iap[i].iact;
453 struct interaction_value *node;
455 if ( interaction_case_is_missing (iact, c, cat->fctr_excl))
458 hash = interaction_case_hash (iact, c, 0);
459 node = lookup_case (&cat->iap[i].ivmap, iact, c);
463 node = pool_malloc (cat->pool, sizeof *node);
464 node->ccase = case_ref (c);
467 hmap_insert (&cat->iap[i].ivmap, &node->node, hash);
471 node->user_data = cat->payload->create (cat->aux1, cat->aux2);
478 cat->iap[i].cc += weight;
482 cat->payload->update (cat->aux1, cat->aux2, node->user_data, c, weight);
487 /* Return the number of categories (distinct values) for interaction IDX in
490 categoricals_n_count (const struct categoricals *cat, size_t n)
492 return hmap_count (&cat->iap[n].ivmap);
496 /* Returns the number of degrees of freedom for interaction IDX within CAT. */
498 categoricals_df (const struct categoricals *cat, size_t n)
500 const struct interact_params *iap = &cat->iap[n];
501 return iap->df_prod[iap->iact->n_vars - 1];
505 /* Return the total number of categories across all interactions in CAT. */
507 categoricals_n_total (const struct categoricals *cat)
509 if (!categoricals_is_complete (cat))
512 return cat->n_cats_total;
515 /* Returns the total degrees of freedom for CAT. */
517 categoricals_df_total (const struct categoricals *cat)
525 /* Returns true iff categoricals_done() has been called for CAT. */
527 categoricals_is_complete (const struct categoricals *cat)
529 return (NULL != cat->reverse_variable_map_short);
533 /* This function must be called (once) before any call to the *_by_subscript or
534 *_by_category functions, but AFTER any calls to categoricals_update. If this
535 function returns false, then no calls to _by_subscript or *_by_category are
538 categoricals_done (const struct categoricals *cat_)
540 /* Implementation Note: Whilst this function is O(n) in cat->n_cats_total, in most
541 uses it will be more efficient that using a tree based structure, since it
542 is called only once, and means that subsequent lookups will be O(1).
544 1 call of O(n) + 10^9 calls of O(1) is better than 10^9 calls of O(log n).
546 struct categoricals *cat = CONST_CAST (struct categoricals *, cat_);
556 cat->n_cats_total = 0;
558 /* Calculate the degrees of freedom, and the number of categories */
559 for (i = 0 ; i < cat->n_iap; ++i)
562 const struct interaction *iact = cat->iap[i].iact;
564 cat->iap[i].df_prod = iact->n_vars ? xcalloc (iact->n_vars, sizeof (int)) : NULL;
566 cat->iap[i].n_cats = 1;
568 for (v = 0 ; v < iact->n_vars; ++v)
571 const struct variable *var = iact->vars[v];
573 struct variable_node *vn = lookup_variable (&cat->varmap, var, hash_pointer (var, 0));
575 struct value_node *valnd = NULL;
576 struct value_node **array ;
578 assert (vn->n_vals == hmap_count (&vn->valmap));
586 /* Sort the VALMAP here */
587 array = xcalloc (sizeof *array, vn->n_vals);
589 HMAP_FOR_EACH (valnd, struct value_node, node, &vn->valmap)
591 /* Note: This loop is probably superfluous, it could be done in the
592 update stage (at the expense of a realloc) */
596 sort (array, vn->n_vals, sizeof (*array),
597 compare_value_node_3way, vn);
599 for (x = 0; x < vn->n_vals; ++x)
601 struct value_node *vvv = array[x];
606 cat->iap[i].df_prod[v] = df * (vn->n_vals - 1);
607 df = cat->iap[i].df_prod[v];
609 cat->iap[i].n_cats *= vn->n_vals;
613 cat->df_sum += cat->iap[i].df_prod [v - 1];
615 cat->n_cats_total += cat->iap[i].n_cats;
619 cat->reverse_variable_map_short = pool_calloc (cat->pool,
621 sizeof *cat->reverse_variable_map_short);
623 cat->reverse_variable_map_long = pool_calloc (cat->pool,
625 sizeof *cat->reverse_variable_map_long);
627 for (i = 0 ; i < cat->n_iap; ++i)
629 struct interaction_value *ivn = NULL;
632 struct interact_params *iap = &cat->iap[i];
634 iap->base_subscript_short = idx_short;
635 iap->base_subscript_long = idx_long;
637 iap->reverse_interaction_value_map = pool_calloc (cat->pool, iap->n_cats,
638 sizeof *iap->reverse_interaction_value_map);
640 HMAP_FOR_EACH (ivn, struct interaction_value, node, &iap->ivmap)
642 iap->reverse_interaction_value_map[x++] = ivn;
645 assert (x <= iap->n_cats);
647 /* For some purposes (eg CONTRASTS in ONEWAY) the values need to be sorted */
648 sort (iap->reverse_interaction_value_map, x, sizeof (*iap->reverse_interaction_value_map),
649 compare_interaction_value_3way, iap);
651 /* Fill the remaining values with null */
652 for (ii = x ; ii < iap->n_cats; ++ii)
653 iap->reverse_interaction_value_map[ii] = NULL;
655 /* Populate the reverse variable maps. */
658 for (ii = 0; ii < iap->df_prod [iap->iact->n_vars - 1]; ++ii)
659 cat->reverse_variable_map_short[idx_short++] = i;
662 for (ii = 0; ii < iap->n_cats; ++ii)
663 cat->reverse_variable_map_long[idx_long++] = i;
666 assert (cat->n_vars <= cat->n_iap);
668 categoricals_dump (cat);
670 /* Tally up the sums for all the encodings */
671 for (i = 0 ; i < cat->n_iap; ++i)
674 struct interact_params *iap = &cat->iap[i];
675 const struct interaction *iact = iap->iact;
677 const int df = iap->df_prod ? iap->df_prod [iact->n_vars - 1] : 0;
679 iap->enc_sum = xcalloc (df, sizeof (*(iap->enc_sum)));
681 for (y = 0; y < hmap_count (&iap->ivmap); ++y)
683 struct interaction_value *iv = iap->reverse_interaction_value_map[y];
684 for (x = iap->base_subscript_short; x < iap->base_subscript_short + df ;++x)
686 const double bin = categoricals_get_effects_code_for_case (cat, x, iv->ccase);
687 iap->enc_sum [x - iap->base_subscript_short] += bin * iv->cc;
689 if (cat->payload && cat->payload->calculate)
690 cat->payload->calculate (cat->aux1, cat->aux2, iv->user_data);
699 reverse_variable_lookup_short (const struct categoricals *cat, int subscript)
701 assert (cat->reverse_variable_map_short);
702 assert (subscript >= 0);
703 assert (subscript < cat->df_sum);
705 return cat->reverse_variable_map_short[subscript];
709 reverse_variable_lookup_long (const struct categoricals *cat, int subscript)
711 assert (cat->reverse_variable_map_long);
712 assert (subscript >= 0);
713 assert (subscript < cat->n_cats_total);
715 return cat->reverse_variable_map_long[subscript];
719 /* Return the interaction corresponding to SUBSCRIPT */
720 const struct interaction *
721 categoricals_get_interaction_by_subscript (const struct categoricals *cat, int subscript)
723 int index = reverse_variable_lookup_short (cat, subscript);
725 return cat->iap[index].iact;
729 categoricals_get_weight_by_subscript (const struct categoricals *cat, int subscript)
731 int vindex = reverse_variable_lookup_short (cat, subscript);
732 const struct interact_params *vp = &cat->iap[vindex];
738 categoricals_get_sum_by_subscript (const struct categoricals *cat, int subscript)
740 int vindex = reverse_variable_lookup_short (cat, subscript);
741 const struct interact_params *vp = &cat->iap[vindex];
743 return vp->enc_sum[subscript - vp->base_subscript_short];
747 /* Returns unity if the value in case C at SUBSCRIPT is equal to the category
748 for that subscript */
750 categoricals_get_code_for_case (const struct categoricals *cat, int subscript,
751 const struct ccase *c,
754 const struct interaction *iact = categoricals_get_interaction_by_subscript (cat, subscript);
756 const int i = reverse_variable_lookup_short (cat, subscript);
758 const int base_index = cat->iap[i].base_subscript_short;
763 const struct interact_params *iap = &cat->iap[i];
766 for (v = 0; v < iact->n_vars; ++v)
768 const struct variable *var = iact->vars[v];
770 const union value *val = case_data (c, var);
771 const int width = var_get_width (var);
772 const struct variable_node *vn = lookup_variable (&cat->varmap, var, hash_pointer (var, 0));
774 const unsigned int hash = value_hash (val, width, 0);
775 const struct value_node *valn = lookup_value (&vn->valmap, val, hash, width);
779 const double df = iap->df_prod[v] / dfp;
781 /* Translate the subscript into an index for the individual variable */
782 const int index = ((subscript - base_index) % iap->df_prod[v] ) / dfp;
783 dfp = iap->df_prod [v];
785 if (effects_coding && valn->index == df )
787 else if ( valn->index != index )
797 /* Returns unity if the value in case C at SUBSCRIPT is equal to the category
798 for that subscript */
800 categoricals_get_dummy_code_for_case (const struct categoricals *cat, int subscript,
801 const struct ccase *c)
803 return categoricals_get_code_for_case (cat, subscript, c, false);
806 /* Returns unity if the value in case C at SUBSCRIPT is equal to the category
808 Else if it is the last category, return -1.
812 categoricals_get_effects_code_for_case (const struct categoricals *cat, int subscript,
813 const struct ccase *c)
815 return categoricals_get_code_for_case (cat, subscript, c, true);
820 categoricals_get_n_variables (const struct categoricals *cat)
822 printf ("%s\n", __FUNCTION__);
827 /* Return a case containing the set of values corresponding to
828 the Nth Category of the IACTth interaction */
830 categoricals_get_case_by_category_real (const struct categoricals *cat, int iact, int n)
832 const struct interaction_value *vn;
834 const struct interact_params *vp = &cat->iap[iact];
836 if ( n >= hmap_count (&vp->ivmap))
839 vn = vp->reverse_interaction_value_map [n];
844 /* Return a the user data corresponding to the Nth Category of the IACTth interaction. */
846 categoricals_get_user_data_by_category_real (const struct categoricals *cat, int iact, int n)
848 const struct interact_params *vp = &cat->iap[iact];
849 const struct interaction_value *iv ;
851 if ( n >= hmap_count (&vp->ivmap))
854 iv = vp->reverse_interaction_value_map [n];
856 return iv->user_data;
861 /* Return a case containing the set of values corresponding to SUBSCRIPT */
863 categoricals_get_case_by_category (const struct categoricals *cat, int subscript)
865 int vindex = reverse_variable_lookup_long (cat, subscript);
866 const struct interact_params *vp = &cat->iap[vindex];
867 const struct interaction_value *vn = vp->reverse_interaction_value_map [subscript - vp->base_subscript_long];
873 categoricals_get_user_data_by_category (const struct categoricals *cat, int subscript)
875 int vindex = reverse_variable_lookup_long (cat, subscript);
876 const struct interact_params *vp = &cat->iap[vindex];
878 const struct interaction_value *iv = vp->reverse_interaction_value_map [subscript - vp->base_subscript_long];
879 return iv->user_data;
886 categoricals_set_payload (struct categoricals *cat, const struct payload *p,
887 const void *aux1, void *aux2)