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));
91 static struct variable_node *
92 lookup_variable (const struct hmap *map, const struct variable *var, unsigned int hash)
94 struct variable_node *vn = NULL;
95 HMAP_FOR_EACH_WITH_HASH (vn, struct variable_node, node, hash, map)
100 fprintf (stderr, "Warning: Hash table collision\n");
107 struct interact_params
109 /* A map indexed by a interaction_value */
112 const struct interaction *iact;
114 int base_subscript_short;
115 int base_subscript_long;
117 /* The number of distinct values of this interaction */
120 /* An array of integers df_n * df_{n-1} * df_{n-2} ...
121 These are the products of the degrees of freedom for the current
122 variable and all preceeding variables */
127 /* A map of interaction_values indexed by subscript */
128 struct interaction_value **reverse_interaction_value_map;
134 /* Comparison function to sort the reverse_value_map in ascending order */
136 compare_interaction_value_3way (const void *vn1_, const void *vn2_, const void *aux)
138 const struct interaction_value *const *vn1p = vn1_;
139 const struct interaction_value *const *vn2p = vn2_;
141 const struct interact_params *iap = aux;
143 return interaction_case_cmp_3way (iap->iact, (*vn1p)->ccase, (*vn2p)->ccase);
148 /* The weight variable */
149 const struct variable *wv;
151 /* An array of interact_params */
152 struct interact_params *iap;
154 /* Map whose members are the union of the variables which comprise IAP */
157 /* The size of IAP. (ie, the number of interactions involved.) */
160 /* The number of categorical variables which contain entries.
161 In the absence of missing values, this will be equal to N_IAP */
166 /* A map to enable the lookup of variables indexed by subscript.
167 This map considers only the N - 1 of the N variables.
169 int *reverse_variable_map_short;
171 /* Like the above, but uses all N variables */
172 int *reverse_variable_map_long;
178 /* Missing values to be excluded */
179 enum mv_class exclude;
184 const struct payload *payload;
189 categoricals_dump (const struct categoricals *cat)
193 printf ("Reverse Variable Map (short):\n");
194 for (i = 0; i < cat->df_sum; ++i)
196 printf (" %d", cat->reverse_variable_map_short[i]);
200 printf ("Reverse Variable Map (long):\n");
201 for (i = 0; i < cat->n_cats_total; ++i)
203 printf (" %d", cat->reverse_variable_map_long[i]);
208 printf ("Number of interactions %d\n", cat->n_iap);
209 for (i = 0 ; i < cat->n_iap; ++i)
213 const struct interact_params *iap = &cat->iap[i];
214 const struct interaction *iact = iap->iact;
216 ds_init_empty (&str);
217 interaction_to_string (iact, &str);
219 printf ("\nInteraction: %s (n: %d; df: %d ); ", ds_cstr (&str), iap->n_cats, iap->df);
221 printf ("Base subscript: %d\n", iap->base_subscript_short);
224 for (v = 0; v < hmap_count (&iap->ivmap); ++v)
227 const struct interaction_value *iv = iap->reverse_interaction_value_map[v];
229 if (v > 0) printf (" ");
231 for (vv = 0; vv < iact->n_vars; ++vv)
233 const struct variable *var = iact->vars[vv];
234 const union value *val = case_data (iv->ccase, var);
236 printf ("%g", val->f);
237 if (vv < iact->n_vars - 1)
248 categoricals_destroy (struct categoricals *cat)
250 struct variable_node *vn = NULL;
254 for (i = 0; i < cat->n_iap; ++i)
256 struct interaction_value *iv = NULL;
257 /* Interate over each interaction value, and unref any cases that we reffed */
258 HMAP_FOR_EACH (iv, struct interaction_value, node, &cat->iap[i].ivmap)
262 cat->payload->destroy (cat->aux1, iv->user_data);
264 case_unref (iv->ccase);
267 free (cat->iap[i].enc_sum);
268 free (cat->iap[i].df_prod);
269 hmap_destroy (&cat->iap[i].ivmap);
272 /* Interate over each variable and delete its value map */
273 HMAP_FOR_EACH (vn, struct variable_node, node, &cat->varmap)
275 hmap_destroy (&vn->valmap);
278 hmap_destroy (&cat->varmap);
280 pool_destroy (cat->pool);
287 static struct interaction_value *
288 lookup_case (const struct hmap *map, const struct interaction *iact, const struct ccase *c)
290 struct interaction_value *iv = NULL;
291 size_t hash = interaction_case_hash (iact, c, 0);
293 HMAP_FOR_EACH_WITH_HASH (iv, struct interaction_value, node, hash, map)
295 if (interaction_case_equal (iact, c, iv->ccase))
298 fprintf (stderr, "Warning: Hash table collision\n");
305 struct categoricals *
306 categoricals_create (struct interaction *const*inter, size_t n_inter,
307 const struct variable *wv, enum mv_class exclude)
310 struct categoricals *cat = xmalloc (sizeof *cat);
312 cat->n_iap = n_inter;
314 cat->n_cats_total = 0;
316 cat->reverse_variable_map_short = NULL;
317 cat->reverse_variable_map_long = NULL;
318 cat->pool = pool_create ();
319 cat->exclude = exclude;
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;
359 const double weight = cat->wv ? case_data (c, cat->wv)->f : 1.0;
361 assert (NULL == cat->reverse_variable_map_short);
362 assert (NULL == cat->reverse_variable_map_long);
364 /* Interate over each variable, and add the value of that variable
365 to the appropriate map, if it's not already present. */
366 HMAP_FOR_EACH (vn, struct variable_node, node, &cat->varmap)
368 const int width = var_get_width (vn->var);
369 const union value *val = case_data (c, vn->var);
370 unsigned int hash = value_hash (val, width, 0);
372 struct value_node *valn = lookup_value (&vn->valmap, val, hash, width);
375 valn = pool_malloc (cat->pool, sizeof *valn);
376 valn->index = vn->n_vals++;
377 value_init (&valn->val, width);
378 value_copy (&valn->val, val, width);
379 hmap_insert (&vn->valmap, &valn->node, hash);
383 for (i = 0 ; i < cat->n_iap; ++i)
385 const struct interaction *iact = cat->iap[i].iact;
388 struct interaction_value *node;
390 if ( interaction_case_is_missing (iact, c, cat->exclude))
393 hash = interaction_case_hash (iact, c, 0);
394 node = lookup_case (&cat->iap[i].ivmap, iact, c);
398 node = pool_malloc (cat->pool, sizeof *node);
400 node->ccase = case_ref (c);
403 hmap_insert (&cat->iap[i].ivmap, &node->node, hash);
407 node->user_data = cat->payload->create (cat->aux1, cat->aux2);
414 cat->iap[i].cc += weight;
418 double weight = cat->wv ? case_data (c, cat->wv)->f : 1.0;
419 cat->payload->update (cat->aux1, cat->aux2, node->user_data, c, weight);
425 /* Return the number of categories (distinct values) for interction N */
427 categoricals_n_count (const struct categoricals *cat, size_t n)
429 return hmap_count (&cat->iap[n].ivmap);
434 categoricals_df (const struct categoricals *cat, size_t n)
436 const struct interact_params *iap = &cat->iap[n];
437 return iap->df_prod[iap->iact->n_vars - 1];
441 /* Return the total number of categories */
443 categoricals_n_total (const struct categoricals *cat)
445 if (!categoricals_is_complete (cat))
448 return cat->n_cats_total;
452 categoricals_df_total (const struct categoricals *cat)
458 categoricals_is_complete (const struct categoricals *cat)
460 return (NULL != cat->reverse_variable_map_short);
464 /* This function must be called *before* any call to categoricals_get_*_by subscript and
465 *after* all calls to categoricals_update */
467 categoricals_done (const struct categoricals *cat_)
469 /* Implementation Note: Whilst this function is O(n) in cat->n_cats_total, in most
470 uses it will be more efficient that using a tree based structure, since it
471 is called only once, and means that subsequent lookups will be O(1).
473 1 call of O(n) + 10^9 calls of O(1) is better than 10^9 calls of O(log n).
475 struct categoricals *cat = CONST_CAST (struct categoricals *, cat_);
481 cat->n_cats_total = 0;
483 /* Calculate the degrees of freedom, and the number of categories */
484 for (i = 0 ; i < cat->n_iap; ++i)
487 const struct interaction *iact = cat->iap[i].iact;
489 cat->iap[i].df_prod = xcalloc (iact->n_vars, sizeof (int));
491 cat->iap[i].n_cats = 1;
493 for (v = 0 ; v < iact->n_vars; ++v)
495 const struct variable *var = iact->vars[v];
497 struct variable_node *vn = lookup_variable (&cat->varmap, var, hash_pointer (var, 0));
499 if (hmap_count (&vn->valmap) == 0)
502 cat->iap[i].df_prod[v] = df * (hmap_count (&vn->valmap) - 1);
503 df = cat->iap[i].df_prod[v];
505 cat->iap[i].n_cats *= hmap_count (&vn->valmap);
508 cat->df_sum += cat->iap[i].df_prod [v - 1];
509 cat->n_cats_total += cat->iap[i].n_cats;
513 cat->reverse_variable_map_short = pool_calloc (cat->pool,
515 sizeof *cat->reverse_variable_map_short);
517 cat->reverse_variable_map_long = pool_calloc (cat->pool,
519 sizeof *cat->reverse_variable_map_long);
521 for (i = 0 ; i < cat->n_iap; ++i)
523 struct interaction_value *ivn = NULL;
526 struct interact_params *iap = &cat->iap[i];
528 iap->base_subscript_short = idx_short;
529 iap->base_subscript_long = idx_long;
531 iap->reverse_interaction_value_map = pool_calloc (cat->pool, iap->n_cats,
532 sizeof *iap->reverse_interaction_value_map);
534 HMAP_FOR_EACH (ivn, struct interaction_value, node, &iap->ivmap)
536 iap->reverse_interaction_value_map[x++] = ivn;
539 assert (x <= iap->n_cats);
541 /* For some purposes (eg CONTRASTS in ONEWAY) the values need to be sorted */
542 sort (iap->reverse_interaction_value_map, x, sizeof (*iap->reverse_interaction_value_map),
543 compare_interaction_value_3way, iap);
545 /* Fill the remaining values with null */
546 for (ii = x ; ii < iap->n_cats; ++ii)
547 iap->reverse_interaction_value_map[ii] = NULL;
549 /* Populate the reverse variable maps. */
550 for (ii = 0; ii < iap->df_prod [iap->iact->n_vars - 1]; ++ii)
551 cat->reverse_variable_map_short[idx_short++] = i;
553 for (ii = 0; ii < iap->n_cats; ++ii)
554 cat->reverse_variable_map_long[idx_long++] = i;
557 assert (cat->n_vars <= cat->n_iap);
559 // categoricals_dump (cat);
561 /* Tally up the sums for all the encodings */
562 for (i = 0 ; i < cat->n_iap; ++i)
565 struct interact_params *iap = &cat->iap[i];
566 const struct interaction *iact = iap->iact;
568 const int df = iap->df_prod [iact->n_vars - 1];
570 iap->enc_sum = xcalloc (df, sizeof (*(iap->enc_sum)));
572 for (y = 0; y < hmap_count (&iap->ivmap); ++y)
574 struct interaction_value *iv = iap->reverse_interaction_value_map[y];
575 for (x = iap->base_subscript_short; x < iap->base_subscript_short + df ;++x)
577 const double bin = categoricals_get_code_for_case (cat, x, iv->ccase); \
578 iap->enc_sum [x - iap->base_subscript_short] += bin * iv->cc;
588 reverse_variable_lookup_short (const struct categoricals *cat, int subscript)
590 assert (cat->reverse_variable_map_short);
591 assert (subscript >= 0);
592 assert (subscript < cat->df_sum);
594 return cat->reverse_variable_map_short[subscript];
598 reverse_variable_lookup_long (const struct categoricals *cat, int subscript)
600 assert (cat->reverse_variable_map_long);
601 assert (subscript >= 0);
602 assert (subscript < cat->n_cats_total);
604 return cat->reverse_variable_map_long[subscript];
608 /* Return the interaction corresponding to SUBSCRIPT */
609 const struct interaction *
610 categoricals_get_interaction_by_subscript (const struct categoricals *cat, int subscript)
612 int index = reverse_variable_lookup_short (cat, subscript);
614 return cat->iap[index].iact;
618 categoricals_get_weight_by_subscript (const struct categoricals *cat, int subscript)
620 int vindex = reverse_variable_lookup_short (cat, subscript);
621 const struct interact_params *vp = &cat->iap[vindex];
627 categoricals_get_sum_by_subscript (const struct categoricals *cat, int subscript)
629 int vindex = reverse_variable_lookup_short (cat, subscript);
630 const struct interact_params *vp = &cat->iap[vindex];
632 return vp->enc_sum[subscript - vp->base_subscript_short];
635 /* Returns unity if the value in case C at SUBSCRIPT is equal to the category
636 for that subscript */
638 categoricals_get_code_for_case (const struct categoricals *cat, int subscript,
639 const struct ccase *c)
641 const struct interaction *iact = categoricals_get_interaction_by_subscript (cat, subscript);
643 const int i = reverse_variable_lookup_short (cat, subscript);
645 const int base_index = cat->iap[i].base_subscript_short;
650 const struct interact_params *iap = &cat->iap[i];
653 for (v = 0; v < iact->n_vars; ++v)
655 const struct variable *var = iact->vars[v];
657 const union value *val = case_data (c, var);
658 const int width = var_get_width (var);
659 const struct variable_node *vn = lookup_variable (&cat->varmap, var, hash_pointer (var, 0));
661 const unsigned int hash = value_hash (val, width, 0);
662 const struct value_node *valn = lookup_value (&vn->valmap, val, hash, width);
666 const double df = iap->df_prod[v] / dfp;
668 /* Translate the subscript into an index for the individual variable */
669 const int index = ((subscript - base_index) % iap->df_prod[v] ) / dfp;
670 dfp = iap->df_prod [v];
673 if ( valn->index == df )
677 if ( valn->index != index )
688 categoricals_get_n_variables (const struct categoricals *cat)
690 printf ("%s\n", __FUNCTION__);
695 /* Return a case containing the set of values corresponding to
696 the Nth Category of the IACTth interaction */
698 categoricals_get_case_by_category_real (const struct categoricals *cat, int iact, int n)
700 const struct interact_params *vp = &cat->iap[iact];
701 const struct interaction_value *vn = vp->reverse_interaction_value_map [n];
706 /* Return a the user data corresponding to the Nth Category of the IACTth interaction. */
708 categoricals_get_user_data_by_category_real (const struct categoricals *cat, int iact, int n)
710 const struct interact_params *vp = &cat->iap[iact];
711 const struct interaction_value *iv = vp->reverse_interaction_value_map [n];
713 return iv->user_data;
718 /* Return a case containing the set of values corresponding to SUBSCRIPT */
720 categoricals_get_case_by_category (const struct categoricals *cat, int subscript)
722 int vindex = reverse_variable_lookup_long (cat, subscript);
723 const struct interact_params *vp = &cat->iap[vindex];
724 const struct interaction_value *vn = vp->reverse_interaction_value_map [subscript - vp->base_subscript_long];
730 categoricals_get_user_data_by_category (const struct categoricals *cat, int subscript)
732 int vindex = reverse_variable_lookup_long (cat, subscript);
733 const struct interact_params *vp = &cat->iap[vindex];
735 const struct interaction_value *iv = vp->reverse_interaction_value_map [subscript - vp->base_subscript_long];
736 return iv->user_data;
743 categoricals_set_payload (struct categoricals *cat, const struct payload *p,
744 const void *aux1, void *aux2)