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;
179 const struct payload *payload;
183 categoricals_dump (const struct categoricals *cat)
185 if (CATEGORICALS_DEBUG)
189 printf ("Reverse Variable Map (short):\n");
190 for (i = 0; i < cat->df_sum; ++i)
192 printf (" %d", cat->reverse_variable_map_short[i]);
196 printf ("Reverse Variable Map (long):\n");
197 for (i = 0; i < cat->n_cats_total; ++i)
199 printf (" %d", cat->reverse_variable_map_long[i]);
204 printf ("Number of interactions %d\n", cat->n_iap);
205 for (i = 0 ; i < cat->n_iap; ++i)
209 const struct interact_params *iap = &cat->iap[i];
210 const struct interaction *iact = iap->iact;
212 ds_init_empty (&str);
213 interaction_to_string (iact, &str);
215 printf ("\nInteraction: %s (n: %d); ", ds_cstr (&str), iap->n_cats);
217 printf ("Base subscript: %d\n", iap->base_subscript_short);
220 for (v = 0; v < hmap_count (&iap->ivmap); ++v)
223 const struct interaction_value *iv = iap->reverse_interaction_value_map[v];
225 if (v > 0) printf (" ");
227 for (vv = 0; vv < iact->n_vars; ++vv)
229 const struct variable *var = iact->vars[vv];
230 const union value *val = case_data (iv->ccase, var);
232 printf ("%g", val->f);
233 if (vv < iact->n_vars - 1)
244 categoricals_destroy (struct categoricals *cat)
246 struct variable_node *vn = NULL;
250 for (i = 0; i < cat->n_iap; ++i)
252 struct interaction_value *iv = NULL;
253 /* Interate over each interaction value, and unref any cases that we reffed */
254 HMAP_FOR_EACH (iv, struct interaction_value, node, &cat->iap[i].ivmap)
258 cat->payload->destroy (cat->aux1, iv->user_data);
260 case_unref (iv->ccase);
263 free (cat->iap[i].enc_sum);
264 free (cat->iap[i].df_prod);
265 hmap_destroy (&cat->iap[i].ivmap);
266 interaction_destroy (cat->iap[i].iact);
269 /* Interate over each variable and delete its value map */
270 HMAP_FOR_EACH (vn, struct variable_node, node, &cat->varmap)
272 hmap_destroy (&vn->valmap);
275 hmap_destroy (&cat->varmap);
277 pool_destroy (cat->pool);
284 static struct interaction_value *
285 lookup_case (const struct hmap *map, const struct interaction *iact, const struct ccase *c)
287 struct interaction_value *iv = NULL;
288 size_t hash = interaction_case_hash (iact, c, 0);
290 HMAP_FOR_EACH_WITH_HASH (iv, struct interaction_value, node, hash, map)
292 if (interaction_case_equal (iact, c, iv->ccase))
295 fprintf (stderr, "Warning: Hash table collision\n");
302 categoricals_sane (const struct categoricals *cat)
307 struct categoricals *
308 categoricals_create (struct interaction *const*inter, size_t n_inter,
309 const struct variable *wv, enum mv_class dep_excl, enum mv_class fctr_excl)
312 struct categoricals *cat = xmalloc (sizeof *cat);
314 cat->n_iap = n_inter;
316 cat->n_cats_total = 0;
318 cat->reverse_variable_map_short = NULL;
319 cat->reverse_variable_map_long = NULL;
320 cat->pool = pool_create ();
321 cat->dep_excl = dep_excl;
322 cat->fctr_excl = fctr_excl;
327 cat->iap = pool_calloc (cat->pool, cat->n_iap, sizeof *cat->iap);
329 hmap_init (&cat->varmap);
330 for (i = 0 ; i < cat->n_iap; ++i)
333 hmap_init (&cat->iap[i].ivmap);
334 cat->iap[i].iact = inter[i];
335 cat->iap[i].cc = 0.0;
336 for (v = 0; v < inter[i]->n_vars; ++v)
338 const struct variable *var = inter[i]->vars[v];
339 unsigned int hash = hash_pointer (var, 0);
340 struct variable_node *vn = lookup_variable (&cat->varmap, var, hash);
343 vn = pool_malloc (cat->pool, sizeof *vn);
346 hmap_init (&vn->valmap);
348 hmap_insert (&cat->varmap, &vn->node, hash);
359 categoricals_update (struct categoricals *cat, const struct ccase *c)
362 struct variable_node *vn = NULL;
363 const double weight = cat->wv ? case_data (c, cat->wv)->f : 1.0;
365 assert (NULL == cat->reverse_variable_map_short);
366 assert (NULL == cat->reverse_variable_map_long);
368 /* Interate over each variable, and add the value of that variable
369 to the appropriate map, if it's not already present. */
370 HMAP_FOR_EACH (vn, struct variable_node, node, &cat->varmap)
372 const int width = var_get_width (vn->var);
373 const union value *val = case_data (c, vn->var);
374 unsigned int hash = value_hash (val, width, 0);
376 struct value_node *valn = lookup_value (&vn->valmap, val, hash, width);
379 valn = pool_malloc (cat->pool, sizeof *valn);
380 valn->index = vn->n_vals++;
381 value_init (&valn->val, width);
382 value_copy (&valn->val, val, width);
383 hmap_insert (&vn->valmap, &valn->node, hash);
387 for (i = 0 ; i < cat->n_iap; ++i)
389 const struct interaction *iact = cat->iap[i].iact;
392 struct interaction_value *node;
394 if ( interaction_case_is_missing (iact, c, cat->fctr_excl))
397 hash = interaction_case_hash (iact, c, 0);
398 node = lookup_case (&cat->iap[i].ivmap, iact, c);
402 node = pool_malloc (cat->pool, sizeof *node);
404 node->ccase = case_ref (c);
407 hmap_insert (&cat->iap[i].ivmap, &node->node, hash);
411 node->user_data = cat->payload->create (cat->aux1, cat->aux2);
418 cat->iap[i].cc += weight;
422 double weight = cat->wv ? case_data (c, cat->wv)->f : 1.0;
423 cat->payload->update (cat->aux1, cat->aux2, node->user_data, c, weight);
429 /* Return the number of categories (distinct values) for interction N */
431 categoricals_n_count (const struct categoricals *cat, size_t n)
433 return hmap_count (&cat->iap[n].ivmap);
438 categoricals_df (const struct categoricals *cat, size_t n)
440 const struct interact_params *iap = &cat->iap[n];
441 return iap->df_prod[iap->iact->n_vars - 1];
445 /* Return the total number of categories */
447 categoricals_n_total (const struct categoricals *cat)
449 if (!categoricals_is_complete (cat))
452 return cat->n_cats_total;
456 categoricals_df_total (const struct categoricals *cat)
462 categoricals_is_complete (const struct categoricals *cat)
464 return (NULL != cat->reverse_variable_map_short);
468 /* This function must be called *before* any call to categoricals_get_*_by subscript and
469 *after* all calls to categoricals_update */
471 categoricals_done (const struct categoricals *cat_)
473 /* Implementation Note: Whilst this function is O(n) in cat->n_cats_total, in most
474 uses it will be more efficient that using a tree based structure, since it
475 is called only once, and means that subsequent lookups will be O(1).
477 1 call of O(n) + 10^9 calls of O(1) is better than 10^9 calls of O(log n).
479 struct categoricals *cat = CONST_CAST (struct categoricals *, cat_);
485 cat->n_cats_total = 0;
487 /* Calculate the degrees of freedom, and the number of categories */
488 for (i = 0 ; i < cat->n_iap; ++i)
491 const struct interaction *iact = cat->iap[i].iact;
493 cat->iap[i].df_prod = iact->n_vars ? xcalloc (iact->n_vars, sizeof (int)) : NULL;
495 cat->iap[i].n_cats = 1;
497 for (v = 0 ; v < iact->n_vars; ++v)
499 const struct variable *var = iact->vars[v];
501 struct variable_node *vn = lookup_variable (&cat->varmap, var, hash_pointer (var, 0));
503 if (hmap_count (&vn->valmap) == 0)
509 cat->iap[i].df_prod[v] = df * (hmap_count (&vn->valmap) - 1);
510 df = cat->iap[i].df_prod[v];
512 cat->iap[i].n_cats *= hmap_count (&vn->valmap);
515 assert (v == iact->n_vars);
517 cat->df_sum += cat->iap[i].df_prod [v - 1];
519 cat->n_cats_total += cat->iap[i].n_cats;
523 cat->reverse_variable_map_short = pool_calloc (cat->pool,
525 sizeof *cat->reverse_variable_map_short);
527 cat->reverse_variable_map_long = pool_calloc (cat->pool,
529 sizeof *cat->reverse_variable_map_long);
531 for (i = 0 ; i < cat->n_iap; ++i)
533 struct interaction_value *ivn = NULL;
536 struct interact_params *iap = &cat->iap[i];
538 iap->base_subscript_short = idx_short;
539 iap->base_subscript_long = idx_long;
541 iap->reverse_interaction_value_map = pool_calloc (cat->pool, iap->n_cats,
542 sizeof *iap->reverse_interaction_value_map);
544 HMAP_FOR_EACH (ivn, struct interaction_value, node, &iap->ivmap)
546 iap->reverse_interaction_value_map[x++] = ivn;
550 assert (x <= iap->n_cats);
552 /* For some purposes (eg CONTRASTS in ONEWAY) the values need to be sorted */
553 sort (iap->reverse_interaction_value_map, x, sizeof (*iap->reverse_interaction_value_map),
554 compare_interaction_value_3way, iap);
556 /* Fill the remaining values with null */
557 for (ii = x ; ii < iap->n_cats; ++ii)
558 iap->reverse_interaction_value_map[ii] = NULL;
560 /* Populate the reverse variable maps. */
563 for (ii = 0; ii < iap->df_prod [iap->iact->n_vars - 1]; ++ii)
564 cat->reverse_variable_map_short[idx_short++] = i;
567 for (ii = 0; ii < iap->n_cats; ++ii)
568 cat->reverse_variable_map_long[idx_long++] = i;
571 assert (cat->n_vars <= cat->n_iap);
573 categoricals_dump (cat);
575 /* Tally up the sums for all the encodings */
576 for (i = 0 ; i < cat->n_iap; ++i)
579 struct interact_params *iap = &cat->iap[i];
580 const struct interaction *iact = iap->iact;
582 const int df = iap->df_prod ? iap->df_prod [iact->n_vars - 1] : 0;
584 iap->enc_sum = xcalloc (df, sizeof (*(iap->enc_sum)));
586 for (y = 0; y < hmap_count (&iap->ivmap); ++y)
588 struct interaction_value *iv = iap->reverse_interaction_value_map[y];
589 for (x = iap->base_subscript_short; x < iap->base_subscript_short + df ;++x)
591 const double bin = categoricals_get_code_for_case (cat, x, iv->ccase); \
592 iap->enc_sum [x - iap->base_subscript_short] += bin * iv->cc;
594 if (cat->payload && cat->payload->destroy)
595 cat->payload->destroy (cat->aux1, cat->aux2, iv->user_data);
604 reverse_variable_lookup_short (const struct categoricals *cat, int subscript)
606 assert (cat->reverse_variable_map_short);
607 assert (subscript >= 0);
608 assert (subscript < cat->df_sum);
610 return cat->reverse_variable_map_short[subscript];
614 reverse_variable_lookup_long (const struct categoricals *cat, int subscript)
616 assert (cat->reverse_variable_map_long);
617 assert (subscript >= 0);
618 assert (subscript < cat->n_cats_total);
620 return cat->reverse_variable_map_long[subscript];
624 /* Return the interaction corresponding to SUBSCRIPT */
625 const struct interaction *
626 categoricals_get_interaction_by_subscript (const struct categoricals *cat, int subscript)
628 int index = reverse_variable_lookup_short (cat, subscript);
630 return cat->iap[index].iact;
634 categoricals_get_weight_by_subscript (const struct categoricals *cat, int subscript)
636 int vindex = reverse_variable_lookup_short (cat, subscript);
637 const struct interact_params *vp = &cat->iap[vindex];
643 categoricals_get_sum_by_subscript (const struct categoricals *cat, int subscript)
645 int vindex = reverse_variable_lookup_short (cat, subscript);
646 const struct interact_params *vp = &cat->iap[vindex];
648 return vp->enc_sum[subscript - vp->base_subscript_short];
651 /* Returns unity if the value in case C at SUBSCRIPT is equal to the category
652 for that subscript */
654 categoricals_get_code_for_case (const struct categoricals *cat, int subscript,
655 const struct ccase *c)
657 const struct interaction *iact = categoricals_get_interaction_by_subscript (cat, subscript);
659 const int i = reverse_variable_lookup_short (cat, subscript);
661 const int base_index = cat->iap[i].base_subscript_short;
666 const struct interact_params *iap = &cat->iap[i];
669 for (v = 0; v < iact->n_vars; ++v)
671 const struct variable *var = iact->vars[v];
673 const union value *val = case_data (c, var);
674 const int width = var_get_width (var);
675 const struct variable_node *vn = lookup_variable (&cat->varmap, var, hash_pointer (var, 0));
677 const unsigned int hash = value_hash (val, width, 0);
678 const struct value_node *valn = lookup_value (&vn->valmap, val, hash, width);
682 const double df = iap->df_prod[v] / dfp;
684 /* Translate the subscript into an index for the individual variable */
685 const int index = ((subscript - base_index) % iap->df_prod[v] ) / dfp;
686 dfp = iap->df_prod [v];
689 if ( valn->index == df )
693 if ( valn->index != index )
704 categoricals_get_n_variables (const struct categoricals *cat)
706 printf ("%s\n", __FUNCTION__);
711 /* Return a case containing the set of values corresponding to
712 the Nth Category of the IACTth interaction */
714 categoricals_get_case_by_category_real (const struct categoricals *cat, int iact, int n)
716 const struct interaction_value *vn;
718 const struct interact_params *vp = &cat->iap[iact];
720 if ( n >= hmap_count (&vp->ivmap))
723 vn = vp->reverse_interaction_value_map [n];
728 /* Return a the user data corresponding to the Nth Category of the IACTth interaction. */
730 categoricals_get_user_data_by_category_real (const struct categoricals *cat, int iact, int n)
732 const struct interact_params *vp = &cat->iap[iact];
733 const struct interaction_value *iv ;
735 if ( n >= hmap_count (&vp->ivmap))
738 iv = vp->reverse_interaction_value_map [n];
740 return iv->user_data;
745 /* Return a case containing the set of values corresponding to SUBSCRIPT */
747 categoricals_get_case_by_category (const struct categoricals *cat, int subscript)
749 int vindex = reverse_variable_lookup_long (cat, subscript);
750 const struct interact_params *vp = &cat->iap[vindex];
751 const struct interaction_value *vn = vp->reverse_interaction_value_map [subscript - vp->base_subscript_long];
757 categoricals_get_user_data_by_category (const struct categoricals *cat, int subscript)
759 int vindex = reverse_variable_lookup_long (cat, subscript);
760 const struct interact_params *vp = &cat->iap[vindex];
762 const struct interaction_value *iv = vp->reverse_interaction_value_map [subscript - vp->base_subscript_long];
763 return iv->user_data;
770 categoricals_set_payload (struct categoricals *cat, const struct payload *p,
771 const void *aux1, void *aux2)