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 EFFECTS_CODING 1
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 */
81 dump_interaction (const struct interaction *iact)
83 struct string str = DS_EMPTY_INITIALIZER;
84 interaction_to_string (iact, &str);
85 printf ("Interaction: %s\n", ds_cstr (&str));
90 static struct variable_node *
91 lookup_variable (const struct hmap *map, const struct variable *var, unsigned int hash)
93 struct variable_node *vn = NULL;
94 HMAP_FOR_EACH_WITH_HASH (vn, struct variable_node, node, hash, map)
99 fprintf (stderr, "Warning: Hash table collision\n");
106 struct interact_params
108 /* A map indexed by a interaction_value */
111 const struct interaction *iact;
113 int base_subscript_short;
114 int base_subscript_long;
116 /* The number of distinct values of this interaction */
119 /* An array of integers df_n * df_{n-1} * df_{n-2} ...
120 These are the products of the degrees of freedom for the current
121 variable and all preceeding variables */
126 /* A map of interaction_values indexed by subscript */
127 struct interaction_value **reverse_interaction_value_map;
133 /* Comparison function to sort the reverse_value_map in ascending order */
135 compare_interaction_value_3way (const void *vn1_, const void *vn2_, const void *aux)
137 const struct interaction_value *const *vn1p = vn1_;
138 const struct interaction_value *const *vn2p = vn2_;
140 const struct interact_params *iap = aux;
142 return interaction_case_cmp_3way (iap->iact, (*vn1p)->ccase, (*vn2p)->ccase);
147 /* The weight variable */
148 const struct variable *wv;
150 /* An array of interact_params */
151 struct interact_params *iap;
153 /* Map whose members are the union of the variables which comprise IAP */
156 /* The size of IAP. (ie, the number of interactions involved.) */
159 /* The number of categorical variables which contain entries.
160 In the absence of missing values, this will be equal to N_IAP */
165 /* A map to enable the lookup of variables indexed by subscript.
166 This map considers only the N - 1 of the N variables.
168 int *reverse_variable_map_short;
170 /* Like the above, but uses all N variables */
171 int *reverse_variable_map_long;
177 /* Missing values to be excluded */
178 enum mv_class exclude;
183 const struct payload *payload;
188 categoricals_dump (const struct categoricals *cat)
192 printf ("Reverse Variable Map (short):\n");
193 for (i = 0; i < cat->df_sum; ++i)
195 printf (" %d", cat->reverse_variable_map_short[i]);
199 printf ("Reverse Variable Map (long):\n");
200 for (i = 0; i < cat->n_cats_total; ++i)
202 printf (" %d", cat->reverse_variable_map_long[i]);
207 printf ("Number of interactions %d\n", cat->n_iap);
208 for (i = 0 ; i < cat->n_iap; ++i)
212 const struct interact_params *iap = &cat->iap[i];
213 const struct interaction *iact = iap->iact;
215 ds_init_empty (&str);
216 interaction_to_string (iact, &str);
218 printf ("\nInteraction: %s (n: %d); ", ds_cstr (&str), iap->n_cats);
220 printf ("Base subscript: %d\n", iap->base_subscript_short);
223 for (v = 0; v < hmap_count (&iap->ivmap); ++v)
226 const struct interaction_value *iv = iap->reverse_interaction_value_map[v];
228 if (v > 0) printf (" ");
230 for (vv = 0; vv < iact->n_vars; ++vv)
232 const struct variable *var = iact->vars[vv];
233 const union value *val = case_data (iv->ccase, var);
235 printf ("%g", val->f);
236 if (vv < iact->n_vars - 1)
247 categoricals_destroy (struct categoricals *cat)
249 struct variable_node *vn = NULL;
253 for (i = 0; i < cat->n_iap; ++i)
255 struct interaction_value *iv = NULL;
256 /* Interate over each interaction value, and unref any cases that we reffed */
257 HMAP_FOR_EACH (iv, struct interaction_value, node, &cat->iap[i].ivmap)
261 cat->payload->destroy (cat->aux1, iv->user_data);
263 case_unref (iv->ccase);
266 free (cat->iap[i].enc_sum);
267 free (cat->iap[i].df_prod);
268 hmap_destroy (&cat->iap[i].ivmap);
271 /* Interate over each variable and delete its value map */
272 HMAP_FOR_EACH (vn, struct variable_node, node, &cat->varmap)
274 hmap_destroy (&vn->valmap);
277 hmap_destroy (&cat->varmap);
279 pool_destroy (cat->pool);
286 static struct interaction_value *
287 lookup_case (const struct hmap *map, const struct interaction *iact, const struct ccase *c)
289 struct interaction_value *iv = NULL;
290 size_t hash = interaction_case_hash (iact, c, 0);
292 HMAP_FOR_EACH_WITH_HASH (iv, struct interaction_value, node, hash, map)
294 if (interaction_case_equal (iact, c, iv->ccase))
297 fprintf (stderr, "Warning: Hash table collision\n");
304 struct categoricals *
305 categoricals_create (struct interaction *const*inter, size_t n_inter,
306 const struct variable *wv, enum mv_class exclude)
309 struct categoricals *cat = xmalloc (sizeof *cat);
311 cat->n_iap = n_inter;
313 cat->n_cats_total = 0;
315 cat->reverse_variable_map_short = NULL;
316 cat->reverse_variable_map_long = NULL;
317 cat->pool = pool_create ();
318 cat->exclude = exclude;
322 cat->iap = pool_calloc (cat->pool, cat->n_iap, sizeof *cat->iap);
324 hmap_init (&cat->varmap);
325 for (i = 0 ; i < cat->n_iap; ++i)
328 hmap_init (&cat->iap[i].ivmap);
329 cat->iap[i].iact = inter[i];
330 cat->iap[i].cc = 0.0;
331 for (v = 0; v < inter[i]->n_vars; ++v)
333 const struct variable *var = inter[i]->vars[v];
334 unsigned int hash = hash_pointer (var, 0);
335 struct variable_node *vn = lookup_variable (&cat->varmap, var, hash);
338 vn = pool_malloc (cat->pool, sizeof *vn);
341 hmap_init (&vn->valmap);
343 hmap_insert (&cat->varmap, &vn->node, hash);
354 categoricals_update (struct categoricals *cat, const struct ccase *c)
357 struct variable_node *vn = NULL;
358 const double weight = cat->wv ? case_data (c, cat->wv)->f : 1.0;
360 assert (NULL == cat->reverse_variable_map_short);
361 assert (NULL == cat->reverse_variable_map_long);
363 /* Interate over each variable, and add the value of that variable
364 to the appropriate map, if it's not already present. */
365 HMAP_FOR_EACH (vn, struct variable_node, node, &cat->varmap)
367 const int width = var_get_width (vn->var);
368 const union value *val = case_data (c, vn->var);
369 unsigned int hash = value_hash (val, width, 0);
371 struct value_node *valn = lookup_value (&vn->valmap, val, hash, width);
374 valn = pool_malloc (cat->pool, sizeof *valn);
375 valn->index = vn->n_vals++;
376 value_init (&valn->val, width);
377 value_copy (&valn->val, val, width);
378 hmap_insert (&vn->valmap, &valn->node, hash);
382 for (i = 0 ; i < cat->n_iap; ++i)
384 const struct interaction *iact = cat->iap[i].iact;
387 struct interaction_value *node;
389 if ( interaction_case_is_missing (iact, c, cat->exclude))
392 hash = interaction_case_hash (iact, c, 0);
393 node = lookup_case (&cat->iap[i].ivmap, iact, c);
397 node = pool_malloc (cat->pool, sizeof *node);
399 node->ccase = case_ref (c);
402 hmap_insert (&cat->iap[i].ivmap, &node->node, hash);
406 node->user_data = cat->payload->create (cat->aux1, cat->aux2);
413 cat->iap[i].cc += weight;
417 double weight = cat->wv ? case_data (c, cat->wv)->f : 1.0;
418 cat->payload->update (cat->aux1, cat->aux2, node->user_data, c, weight);
424 /* Return the number of categories (distinct values) for interction N */
426 categoricals_n_count (const struct categoricals *cat, size_t n)
428 return hmap_count (&cat->iap[n].ivmap);
433 categoricals_df (const struct categoricals *cat, size_t n)
435 const struct interact_params *iap = &cat->iap[n];
436 return iap->df_prod[iap->iact->n_vars - 1];
440 /* Return the total number of categories */
442 categoricals_n_total (const struct categoricals *cat)
444 if (!categoricals_is_complete (cat))
447 return cat->n_cats_total;
451 categoricals_df_total (const struct categoricals *cat)
457 categoricals_is_complete (const struct categoricals *cat)
459 return (NULL != cat->reverse_variable_map_short);
463 /* This function must be called *before* any call to categoricals_get_*_by subscript and
464 *after* all calls to categoricals_update */
466 categoricals_done (const struct categoricals *cat_)
468 /* Implementation Note: Whilst this function is O(n) in cat->n_cats_total, in most
469 uses it will be more efficient that using a tree based structure, since it
470 is called only once, and means that subsequent lookups will be O(1).
472 1 call of O(n) + 10^9 calls of O(1) is better than 10^9 calls of O(log n).
474 struct categoricals *cat = CONST_CAST (struct categoricals *, cat_);
480 cat->n_cats_total = 0;
482 /* Calculate the degrees of freedom, and the number of categories */
483 for (i = 0 ; i < cat->n_iap; ++i)
486 const struct interaction *iact = cat->iap[i].iact;
488 cat->iap[i].df_prod = iact->n_vars ? xcalloc (iact->n_vars, sizeof (int)) : NULL;
490 cat->iap[i].n_cats = 1;
492 for (v = 0 ; v < iact->n_vars; ++v)
494 const struct variable *var = iact->vars[v];
496 struct variable_node *vn = lookup_variable (&cat->varmap, var, hash_pointer (var, 0));
498 if (hmap_count (&vn->valmap) == 0)
501 cat->iap[i].df_prod[v] = df * (hmap_count (&vn->valmap) - 1);
502 df = cat->iap[i].df_prod[v];
504 cat->iap[i].n_cats *= hmap_count (&vn->valmap);
507 assert (v == iact->n_vars);
509 cat->df_sum += cat->iap[i].df_prod [v - 1];
511 cat->n_cats_total += cat->iap[i].n_cats;
515 cat->reverse_variable_map_short = pool_calloc (cat->pool,
517 sizeof *cat->reverse_variable_map_short);
519 cat->reverse_variable_map_long = pool_calloc (cat->pool,
521 sizeof *cat->reverse_variable_map_long);
523 for (i = 0 ; i < cat->n_iap; ++i)
525 struct interaction_value *ivn = NULL;
528 struct interact_params *iap = &cat->iap[i];
530 iap->base_subscript_short = idx_short;
531 iap->base_subscript_long = idx_long;
533 iap->reverse_interaction_value_map = pool_calloc (cat->pool, iap->n_cats,
534 sizeof *iap->reverse_interaction_value_map);
536 HMAP_FOR_EACH (ivn, struct interaction_value, node, &iap->ivmap)
538 iap->reverse_interaction_value_map[x++] = ivn;
541 assert (x <= iap->n_cats);
543 /* For some purposes (eg CONTRASTS in ONEWAY) the values need to be sorted */
544 sort (iap->reverse_interaction_value_map, x, sizeof (*iap->reverse_interaction_value_map),
545 compare_interaction_value_3way, iap);
547 /* Fill the remaining values with null */
548 for (ii = x ; ii < iap->n_cats; ++ii)
549 iap->reverse_interaction_value_map[ii] = NULL;
551 /* Populate the reverse variable maps. */
554 for (ii = 0; ii < iap->df_prod [iap->iact->n_vars - 1]; ++ii)
555 cat->reverse_variable_map_short[idx_short++] = i;
558 for (ii = 0; ii < iap->n_cats; ++ii)
559 cat->reverse_variable_map_long[idx_long++] = i;
562 assert (cat->n_vars <= cat->n_iap);
564 // categoricals_dump (cat);
566 /* Tally up the sums for all the encodings */
567 for (i = 0 ; i < cat->n_iap; ++i)
570 struct interact_params *iap = &cat->iap[i];
571 const struct interaction *iact = iap->iact;
573 const int df = iap->df_prod ? iap->df_prod [iact->n_vars - 1] : 0;
575 iap->enc_sum = xcalloc (df, sizeof (*(iap->enc_sum)));
577 for (y = 0; y < hmap_count (&iap->ivmap); ++y)
579 struct interaction_value *iv = iap->reverse_interaction_value_map[y];
580 for (x = iap->base_subscript_short; x < iap->base_subscript_short + df ;++x)
582 const double bin = categoricals_get_code_for_case (cat, x, iv->ccase); \
583 iap->enc_sum [x - iap->base_subscript_short] += bin * iv->cc;
593 reverse_variable_lookup_short (const struct categoricals *cat, int subscript)
595 assert (cat->reverse_variable_map_short);
596 assert (subscript >= 0);
597 assert (subscript < cat->df_sum);
599 return cat->reverse_variable_map_short[subscript];
603 reverse_variable_lookup_long (const struct categoricals *cat, int subscript)
605 assert (cat->reverse_variable_map_long);
606 assert (subscript >= 0);
607 assert (subscript < cat->n_cats_total);
609 return cat->reverse_variable_map_long[subscript];
613 /* Return the interaction corresponding to SUBSCRIPT */
614 const struct interaction *
615 categoricals_get_interaction_by_subscript (const struct categoricals *cat, int subscript)
617 int index = reverse_variable_lookup_short (cat, subscript);
619 return cat->iap[index].iact;
623 categoricals_get_weight_by_subscript (const struct categoricals *cat, int subscript)
625 int vindex = reverse_variable_lookup_short (cat, subscript);
626 const struct interact_params *vp = &cat->iap[vindex];
632 categoricals_get_sum_by_subscript (const struct categoricals *cat, int subscript)
634 int vindex = reverse_variable_lookup_short (cat, subscript);
635 const struct interact_params *vp = &cat->iap[vindex];
637 return vp->enc_sum[subscript - vp->base_subscript_short];
640 /* Returns unity if the value in case C at SUBSCRIPT is equal to the category
641 for that subscript */
643 categoricals_get_code_for_case (const struct categoricals *cat, int subscript,
644 const struct ccase *c)
646 const struct interaction *iact = categoricals_get_interaction_by_subscript (cat, subscript);
648 const int i = reverse_variable_lookup_short (cat, subscript);
650 const int base_index = cat->iap[i].base_subscript_short;
655 const struct interact_params *iap = &cat->iap[i];
658 for (v = 0; v < iact->n_vars; ++v)
660 const struct variable *var = iact->vars[v];
662 const union value *val = case_data (c, var);
663 const int width = var_get_width (var);
664 const struct variable_node *vn = lookup_variable (&cat->varmap, var, hash_pointer (var, 0));
666 const unsigned int hash = value_hash (val, width, 0);
667 const struct value_node *valn = lookup_value (&vn->valmap, val, hash, width);
671 const double df = iap->df_prod[v] / dfp;
673 /* Translate the subscript into an index for the individual variable */
674 const int index = ((subscript - base_index) % iap->df_prod[v] ) / dfp;
675 dfp = iap->df_prod [v];
678 if ( valn->index == df )
682 if ( valn->index != index )
693 categoricals_get_n_variables (const struct categoricals *cat)
695 printf ("%s\n", __FUNCTION__);
700 /* Return a case containing the set of values corresponding to
701 the Nth Category of the IACTth interaction */
703 categoricals_get_case_by_category_real (const struct categoricals *cat, int iact, int n)
705 const struct interact_params *vp = &cat->iap[iact];
706 const struct interaction_value *vn = vp->reverse_interaction_value_map [n];
711 /* Return a the user data corresponding to the Nth Category of the IACTth interaction. */
713 categoricals_get_user_data_by_category_real (const struct categoricals *cat, int iact, int n)
715 const struct interact_params *vp = &cat->iap[iact];
716 const struct interaction_value *iv = vp->reverse_interaction_value_map [n];
718 return iv->user_data;
723 /* Return a case containing the set of values corresponding to SUBSCRIPT */
725 categoricals_get_case_by_category (const struct categoricals *cat, int subscript)
727 int vindex = reverse_variable_lookup_long (cat, subscript);
728 const struct interact_params *vp = &cat->iap[vindex];
729 const struct interaction_value *vn = vp->reverse_interaction_value_map [subscript - vp->base_subscript_long];
735 categoricals_get_user_data_by_category (const struct categoricals *cat, int subscript)
737 int vindex = reverse_variable_lookup_long (cat, subscript);
738 const struct interact_params *vp = &cat->iap[vindex];
740 const struct interaction_value *iv = vp->reverse_interaction_value_map [subscript - vp->base_subscript_long];
741 return iv->user_data;
748 categoricals_set_payload (struct categoricals *cat, const struct payload *p,
749 const void *aux1, void *aux2)