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
37 #define EFFECTS_CODING 1
41 struct hmap_node node; /* Node in hash map. */
43 union value val; /* The value */
45 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 this value */
54 double cc; /* Total of the weights of cases matching this interaction */
56 void *user_data; /* A pointer to data which the caller can store stuff */
59 static struct value_node *
60 lookup_value (const struct hmap *map, const union value *val, unsigned int hash, int width)
62 struct value_node *vn = NULL;
63 HMAP_FOR_EACH_WITH_HASH (vn, struct value_node, node, hash, map)
65 if (value_equal (&vn->val, val, width))
74 struct hmap_node node; /* Node in hash map. */
75 const struct variable *var; /* The variable */
77 struct hmap valmap; /* A map of value nodes */
78 int n_vals; /* Number of values for this variable */
81 static struct variable_node *
82 lookup_variable (const struct hmap *map, const struct variable *var, unsigned int hash)
84 struct variable_node *vn = NULL;
85 HMAP_FOR_EACH_WITH_HASH (vn, struct variable_node, node, hash, map)
90 fprintf (stderr, "Warning: Hash table collision\n");
97 struct interact_params
99 /* A map indexed by a interaction_value */
102 const struct interaction *iact;
104 int base_subscript_short;
105 int base_subscript_long;
107 /* The number of distinct values of this interaction */
110 /* An array of integers df_n * df_{n-1} * df_{n-2} ...
111 These are the products of the degrees of freedom for the current
112 variable and all preceeding variables */
117 /* A map of interaction_values indexed by subscript */
118 struct interaction_value **reverse_interaction_value_map;
124 /* Comparison function to sort the reverse_value_map in ascending order */
126 compare_interaction_value_3way (const void *vn1_, const void *vn2_, const void *aux)
128 const struct interaction_value *const *vn1p = vn1_;
129 const struct interaction_value *const *vn2p = vn2_;
131 const struct interact_params *iap = aux;
133 return interaction_case_cmp_3way (iap->iact, (*vn1p)->ccase, (*vn2p)->ccase);
138 /* The weight variable */
139 const struct variable *wv;
141 /* An array of interact_params */
142 struct interact_params *iap;
144 /* Map whose members are the union of the variables which comprise IAP */
147 /* The size of IAP. (ie, the number of interactions involved.) */
150 /* The number of categorical variables which contain entries.
151 In the absence of missing values, this will be equal to N_IAP */
156 /* A map to enable the lookup of variables indexed by subscript.
157 This map considers only the N - 1 of the N variables.
159 int *reverse_variable_map_short;
161 /* Like the above, but uses all N variables */
162 int *reverse_variable_map_long;
168 /* Missing values in the dependent varirable to be excluded */
169 enum mv_class dep_excl;
171 /* Missing values in the factor variables to be excluded */
172 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)
256 cat->payload->destroy (cat->aux1, iv->user_data);
258 case_unref (iv->ccase);
261 free (cat->iap[i].enc_sum);
262 free (cat->iap[i].df_prod);
263 hmap_destroy (&cat->iap[i].ivmap);
266 /* Interate over each variable and delete its value map */
267 HMAP_FOR_EACH (vn, struct variable_node, node, &cat->varmap)
269 hmap_destroy (&vn->valmap);
272 hmap_destroy (&cat->varmap);
274 pool_destroy (cat->pool);
281 static struct interaction_value *
282 lookup_case (const struct hmap *map, const struct interaction *iact, const struct ccase *c)
284 struct interaction_value *iv = NULL;
285 size_t hash = interaction_case_hash (iact, c, 0);
287 HMAP_FOR_EACH_WITH_HASH (iv, struct interaction_value, node, hash, map)
289 if (interaction_case_equal (iact, c, iv->ccase))
292 fprintf (stderr, "Warning: Hash table collision\n");
299 struct categoricals *
300 categoricals_create (struct interaction *const*inter, size_t n_inter,
301 const struct variable *wv, enum mv_class dep_excl, enum mv_class fctr_excl)
304 struct categoricals *cat = xmalloc (sizeof *cat);
306 cat->n_iap = n_inter;
308 cat->n_cats_total = 0;
310 cat->reverse_variable_map_short = NULL;
311 cat->reverse_variable_map_long = NULL;
312 cat->pool = pool_create ();
313 cat->dep_excl = dep_excl;
314 cat->fctr_excl = fctr_excl;
318 cat->iap = pool_calloc (cat->pool, cat->n_iap, sizeof *cat->iap);
320 hmap_init (&cat->varmap);
321 for (i = 0 ; i < cat->n_iap; ++i)
324 hmap_init (&cat->iap[i].ivmap);
325 cat->iap[i].iact = inter[i];
326 cat->iap[i].cc = 0.0;
327 for (v = 0; v < inter[i]->n_vars; ++v)
329 const struct variable *var = inter[i]->vars[v];
330 unsigned int hash = hash_pointer (var, 0);
331 struct variable_node *vn = lookup_variable (&cat->varmap, var, hash);
334 vn = pool_malloc (cat->pool, sizeof *vn);
337 hmap_init (&vn->valmap);
339 hmap_insert (&cat->varmap, &vn->node, hash);
350 categoricals_update (struct categoricals *cat, const struct ccase *c)
353 struct variable_node *vn = NULL;
354 const double weight = cat->wv ? case_data (c, cat->wv)->f : 1.0;
356 assert (NULL == cat->reverse_variable_map_short);
357 assert (NULL == cat->reverse_variable_map_long);
359 /* Interate over each variable, and add the value of that variable
360 to the appropriate map, if it's not already present. */
361 HMAP_FOR_EACH (vn, struct variable_node, node, &cat->varmap)
363 const int width = var_get_width (vn->var);
364 const union value *val = case_data (c, vn->var);
365 unsigned int hash = value_hash (val, width, 0);
367 struct value_node *valn = lookup_value (&vn->valmap, val, hash, width);
370 valn = pool_malloc (cat->pool, sizeof *valn);
371 valn->index = vn->n_vals++;
372 value_init (&valn->val, width);
373 value_copy (&valn->val, val, width);
374 hmap_insert (&vn->valmap, &valn->node, hash);
378 for (i = 0 ; i < cat->n_iap; ++i)
380 const struct interaction *iact = cat->iap[i].iact;
383 struct interaction_value *node;
385 if ( interaction_case_is_missing (iact, c, cat->fctr_excl))
388 hash = interaction_case_hash (iact, c, 0);
389 node = lookup_case (&cat->iap[i].ivmap, iact, c);
393 node = pool_malloc (cat->pool, sizeof *node);
395 node->ccase = case_ref (c);
398 hmap_insert (&cat->iap[i].ivmap, &node->node, hash);
402 node->user_data = cat->payload->create (cat->aux1, cat->aux2);
409 cat->iap[i].cc += weight;
413 double weight = cat->wv ? case_data (c, cat->wv)->f : 1.0;
414 cat->payload->update (cat->aux1, cat->aux2, node->user_data, c, weight);
420 /* Return the number of categories (distinct values) for interction N */
422 categoricals_n_count (const struct categoricals *cat, size_t n)
424 return hmap_count (&cat->iap[n].ivmap);
429 categoricals_df (const struct categoricals *cat, size_t n)
431 const struct interact_params *iap = &cat->iap[n];
432 return iap->df_prod[iap->iact->n_vars - 1];
436 /* Return the total number of categories */
438 categoricals_n_total (const struct categoricals *cat)
440 if (!categoricals_is_complete (cat))
443 return cat->n_cats_total;
447 categoricals_df_total (const struct categoricals *cat)
453 categoricals_is_complete (const struct categoricals *cat)
455 return (NULL != cat->reverse_variable_map_short);
459 /* This function must be called *before* any call to categoricals_get_*_by subscript and
460 *after* all calls to categoricals_update */
462 categoricals_done (const struct categoricals *cat_)
464 /* Implementation Note: Whilst this function is O(n) in cat->n_cats_total, in most
465 uses it will be more efficient that using a tree based structure, since it
466 is called only once, and means that subsequent lookups will be O(1).
468 1 call of O(n) + 10^9 calls of O(1) is better than 10^9 calls of O(log n).
470 struct categoricals *cat = CONST_CAST (struct categoricals *, cat_);
476 cat->n_cats_total = 0;
478 /* Calculate the degrees of freedom, and the number of categories */
479 for (i = 0 ; i < cat->n_iap; ++i)
482 const struct interaction *iact = cat->iap[i].iact;
484 cat->iap[i].df_prod = iact->n_vars ? xcalloc (iact->n_vars, sizeof (int)) : NULL;
486 cat->iap[i].n_cats = 1;
488 for (v = 0 ; v < iact->n_vars; ++v)
490 const struct variable *var = iact->vars[v];
492 struct variable_node *vn = lookup_variable (&cat->varmap, var, hash_pointer (var, 0));
494 if (hmap_count (&vn->valmap) == 0)
497 cat->iap[i].df_prod[v] = df * (hmap_count (&vn->valmap) - 1);
498 df = cat->iap[i].df_prod[v];
500 cat->iap[i].n_cats *= hmap_count (&vn->valmap);
503 assert (v == iact->n_vars);
505 cat->df_sum += cat->iap[i].df_prod [v - 1];
507 cat->n_cats_total += cat->iap[i].n_cats;
511 cat->reverse_variable_map_short = pool_calloc (cat->pool,
513 sizeof *cat->reverse_variable_map_short);
515 cat->reverse_variable_map_long = pool_calloc (cat->pool,
517 sizeof *cat->reverse_variable_map_long);
519 for (i = 0 ; i < cat->n_iap; ++i)
521 struct interaction_value *ivn = NULL;
524 struct interact_params *iap = &cat->iap[i];
526 iap->base_subscript_short = idx_short;
527 iap->base_subscript_long = idx_long;
529 iap->reverse_interaction_value_map = pool_calloc (cat->pool, iap->n_cats,
530 sizeof *iap->reverse_interaction_value_map);
532 HMAP_FOR_EACH (ivn, struct interaction_value, node, &iap->ivmap)
534 iap->reverse_interaction_value_map[x++] = ivn;
538 assert (x <= iap->n_cats);
540 /* For some purposes (eg CONTRASTS in ONEWAY) the values need to be sorted */
541 sort (iap->reverse_interaction_value_map, x, sizeof (*iap->reverse_interaction_value_map),
542 compare_interaction_value_3way, iap);
544 /* Fill the remaining values with null */
545 for (ii = x ; ii < iap->n_cats; ++ii)
546 iap->reverse_interaction_value_map[ii] = NULL;
548 /* Populate the reverse variable maps. */
551 for (ii = 0; ii < iap->df_prod [iap->iact->n_vars - 1]; ++ii)
552 cat->reverse_variable_map_short[idx_short++] = i;
555 for (ii = 0; ii < iap->n_cats; ++ii)
556 cat->reverse_variable_map_long[idx_long++] = i;
559 assert (cat->n_vars <= cat->n_iap);
561 categoricals_dump (cat);
563 /* Tally up the sums for all the encodings */
564 for (i = 0 ; i < cat->n_iap; ++i)
567 struct interact_params *iap = &cat->iap[i];
568 const struct interaction *iact = iap->iact;
570 const int df = iap->df_prod ? iap->df_prod [iact->n_vars - 1] : 0;
572 iap->enc_sum = xcalloc (df, sizeof (*(iap->enc_sum)));
574 for (y = 0; y < hmap_count (&iap->ivmap); ++y)
576 struct interaction_value *iv = iap->reverse_interaction_value_map[y];
577 for (x = iap->base_subscript_short; x < iap->base_subscript_short + df ;++x)
579 const double bin = categoricals_get_code_for_case (cat, x, iv->ccase); \
580 iap->enc_sum [x - iap->base_subscript_short] += bin * iv->cc;
582 if (cat->payload && cat->payload->destroy)
583 cat->payload->destroy (cat->aux1, cat->aux2, iv->user_data);
592 reverse_variable_lookup_short (const struct categoricals *cat, int subscript)
594 assert (cat->reverse_variable_map_short);
595 assert (subscript >= 0);
596 assert (subscript < cat->df_sum);
598 return cat->reverse_variable_map_short[subscript];
602 reverse_variable_lookup_long (const struct categoricals *cat, int subscript)
604 assert (cat->reverse_variable_map_long);
605 assert (subscript >= 0);
606 assert (subscript < cat->n_cats_total);
608 return cat->reverse_variable_map_long[subscript];
612 /* Return the interaction corresponding to SUBSCRIPT */
613 const struct interaction *
614 categoricals_get_interaction_by_subscript (const struct categoricals *cat, int subscript)
616 int index = reverse_variable_lookup_short (cat, subscript);
618 return cat->iap[index].iact;
622 categoricals_get_weight_by_subscript (const struct categoricals *cat, int subscript)
624 int vindex = reverse_variable_lookup_short (cat, subscript);
625 const struct interact_params *vp = &cat->iap[vindex];
631 categoricals_get_sum_by_subscript (const struct categoricals *cat, int subscript)
633 int vindex = reverse_variable_lookup_short (cat, subscript);
634 const struct interact_params *vp = &cat->iap[vindex];
636 return vp->enc_sum[subscript - vp->base_subscript_short];
639 /* Returns unity if the value in case C at SUBSCRIPT is equal to the category
640 for that subscript */
642 categoricals_get_code_for_case (const struct categoricals *cat, int subscript,
643 const struct ccase *c)
645 const struct interaction *iact = categoricals_get_interaction_by_subscript (cat, subscript);
647 const int i = reverse_variable_lookup_short (cat, subscript);
649 const int base_index = cat->iap[i].base_subscript_short;
654 const struct interact_params *iap = &cat->iap[i];
657 for (v = 0; v < iact->n_vars; ++v)
659 const struct variable *var = iact->vars[v];
661 const union value *val = case_data (c, var);
662 const int width = var_get_width (var);
663 const struct variable_node *vn = lookup_variable (&cat->varmap, var, hash_pointer (var, 0));
665 const unsigned int hash = value_hash (val, width, 0);
666 const struct value_node *valn = lookup_value (&vn->valmap, val, hash, width);
670 const double df = iap->df_prod[v] / dfp;
672 /* Translate the subscript into an index for the individual variable */
673 const int index = ((subscript - base_index) % iap->df_prod[v] ) / dfp;
674 dfp = iap->df_prod [v];
677 if ( valn->index == df )
681 if ( valn->index != index )
692 categoricals_get_n_variables (const struct categoricals *cat)
694 printf ("%s\n", __FUNCTION__);
699 /* Return a case containing the set of values corresponding to
700 the Nth Category of the IACTth interaction */
702 categoricals_get_case_by_category_real (const struct categoricals *cat, int iact, int n)
704 const struct interaction_value *vn;
706 const struct interact_params *vp = &cat->iap[iact];
708 if ( n >= hmap_count (&vp->ivmap))
711 vn = vp->reverse_interaction_value_map [n];
716 /* Return a the user data corresponding to the Nth Category of the IACTth interaction. */
718 categoricals_get_user_data_by_category_real (const struct categoricals *cat, int iact, int n)
720 const struct interact_params *vp = &cat->iap[iact];
721 const struct interaction_value *iv ;
723 if ( n >= hmap_count (&vp->ivmap))
726 iv = vp->reverse_interaction_value_map [n];
728 return iv->user_data;
733 /* Return a case containing the set of values corresponding to SUBSCRIPT */
735 categoricals_get_case_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];
739 const struct interaction_value *vn = vp->reverse_interaction_value_map [subscript - vp->base_subscript_long];
745 categoricals_get_user_data_by_category (const struct categoricals *cat, int subscript)
747 int vindex = reverse_variable_lookup_long (cat, subscript);
748 const struct interact_params *vp = &cat->iap[vindex];
750 const struct interaction_value *iv = vp->reverse_interaction_value_map [subscript - vp->base_subscript_long];
751 return iv->user_data;
758 categoricals_set_payload (struct categoricals *cat, const struct payload *p,
759 const void *aux1, void *aux2)