1 /* PSPP - a program for statistical analysis.
2 Copyright (C) 2009, 2010, 2011 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"
37 struct hmap_node node; /* Node in hash map. */
39 union value val; /* The value */
41 int index; /* A zero based unique index for this value */
44 struct interaction_value
46 struct hmap_node node; /* Node in hash map. */
48 struct ccase *ccase; /* A case (probably the first in the dataset) which matches this value */
50 /* Total of the weights of cases matching this interaction */
53 void *user_data; /* A pointer to data which the caller can store stuff */
56 static struct value_node *
57 lookup_value (const struct hmap *map, const union value *val, unsigned int hash, int width)
59 struct value_node *vn = NULL;
60 HMAP_FOR_EACH_WITH_HASH (vn, struct value_node, node, hash, map)
62 if (value_equal (&vn->val, val, width))
71 struct hmap_node node; /* Node in hash map. */
72 const struct variable *var; /* The variable */
74 struct hmap valmap; /* A map of value nodes */
75 int n_vals; /* Number of values for this variable */
80 dump_interaction (const struct interaction *iact)
82 struct string str = DS_EMPTY_INITIALIZER;
83 interaction_to_string (iact, &str);
84 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 */
124 /* A map of interaction_values indexed by subscript */
125 struct interaction_value **reverse_interaction_value_map;
131 /* Comparison function to sort the reverse_value_map in ascending order */
133 compare_interaction_value_3way (const void *vn1_, const void *vn2_, const void *aux)
135 const struct interaction_value *const *vn1p = vn1_;
136 const struct interaction_value *const *vn2p = vn2_;
138 const struct interact_params *iap = aux;
140 return interaction_case_cmp_3way (iap->iact, (*vn1p)->ccase, (*vn2p)->ccase);
145 /* The weight variable */
146 const struct variable *wv;
148 /* An array of interact_params */
149 struct interact_params *iap;
151 /* Map whose members are the union of the variables which comprise IAP */
154 /* The size of IAP. (ie, the number of interactions involved.) */
157 /* The number of categorical variables which contain entries.
158 In the absence of missing values, this will be equal to N_IAP */
163 /* A map to enable the lookup of variables indexed by subscript.
164 This map considers only the N - 1 of the N variables.
166 int *reverse_variable_map_short;
168 /* Like the above, but uses all N variables */
169 int *reverse_variable_map_long;
175 /* Missing values to be excluded */
176 enum mv_class exclude;
178 /* Function to be called on each update */
181 /* Function specified by the caller to create user_data */
182 user_data_create_func *user_data_create;
184 /* Auxilliary data to be passed to update and user_data_create_func*/
191 categoricals_dump (const struct categoricals *cat)
195 printf ("Reverse Variable Map (short):\n");
196 for (i = 0; i < cat->df_sum; ++i)
198 printf (" %d", cat->reverse_variable_map_short[i]);
202 printf ("Reverse Variable Map (long):\n");
203 for (i = 0; i < cat->n_cats_total; ++i)
205 printf (" %d", cat->reverse_variable_map_long[i]);
210 printf ("Number of interactions %d\n", cat->n_iap);
211 for (i = 0 ; i < cat->n_iap; ++i)
215 const struct interact_params *iap = &cat->iap[i];
216 const struct interaction *iact = iap->iact;
218 ds_init_empty (&str);
219 interaction_to_string (iact, &str);
221 printf ("\nInteraction: %s (n: %d; df: %d ); ", ds_cstr (&str), iap->n_cats, iap->df);
223 printf ("Base subscript: %d\n", iap->base_subscript_short);
226 for (v = 0; v < hmap_count (&iap->ivmap); ++v)
229 const struct interaction_value *iv = iap->reverse_interaction_value_map[v];
231 if (v > 0) printf (" ");
233 for (vv = 0; vv < iact->n_vars; ++vv)
235 const struct variable *var = iact->vars[vv];
236 const union value *val = case_data (iv->ccase, var);
238 printf ("%g", val->f);
239 if (vv < iact->n_vars - 1)
250 categoricals_destroy (struct categoricals *cat)
252 struct variable_node *vn = NULL;
256 for (i = 0; i < cat->n_iap; ++i)
258 struct interaction_value *iv = NULL;
259 /* Interate over each interaction value, and unref any cases that we reffed */
260 HMAP_FOR_EACH (iv, struct interaction_value, node, &cat->iap[i].ivmap)
262 case_unref (iv->ccase);
264 hmap_destroy (&cat->iap[i].ivmap);
267 /* Interate over each variable and delete its value map */
268 HMAP_FOR_EACH (vn, struct variable_node, node, &cat->varmap)
270 hmap_destroy (&vn->valmap);
273 hmap_destroy (&cat->varmap);
275 pool_destroy (cat->pool);
282 static struct interaction_value *
283 lookup_case (const struct hmap *map, const struct interaction *iact, const struct ccase *c)
285 struct interaction_value *iv = NULL;
286 size_t hash = interaction_case_hash (iact, c, 0);
288 HMAP_FOR_EACH_WITH_HASH (iv, struct interaction_value, node, hash, map)
290 if (interaction_case_equal (iact, c, iv->ccase))
293 fprintf (stderr, "Warning: Hash table collision\n");
300 struct categoricals *
301 categoricals_create (struct interaction *const*inter, size_t n_inter,
302 const struct variable *wv, enum mv_class exclude,
303 user_data_create_func *udf,
304 update_func *update, void *aux1, void *aux2
308 struct categoricals *cat = xmalloc (sizeof *cat);
310 cat->n_iap = n_inter;
312 cat->n_cats_total = 0;
314 cat->reverse_variable_map_short = NULL;
315 cat->reverse_variable_map_long = NULL;
316 cat->pool = pool_create ();
317 cat->exclude = exclude;
318 cat->update = update;
319 cat->user_data_create = udf;
324 cat->iap = pool_calloc (cat->pool, cat->n_iap, sizeof *cat->iap);
326 hmap_init (&cat->varmap);
327 for (i = 0 ; i < cat->n_iap; ++i)
330 hmap_init (&cat->iap[i].ivmap);
331 cat->iap[i].iact = inter[i];
332 cat->iap[i].cc = 0.0;
333 for (v = 0; v < inter[i]->n_vars; ++v)
335 const struct variable *var = inter[i]->vars[v];
336 unsigned int hash = hash_pointer (var, 0);
337 struct variable_node *vn = lookup_variable (&cat->varmap, var, hash);
340 vn = pool_malloc (cat->pool, sizeof *vn);
343 hmap_init (&vn->valmap);
345 hmap_insert (&cat->varmap, &vn->node, hash);
356 categoricals_update (struct categoricals *cat, const struct ccase *c)
359 struct variable_node *vn = NULL;
360 const double weight = cat->wv ? case_data (c, cat->wv)->f : 1.0;
362 assert (NULL == cat->reverse_variable_map_short);
363 assert (NULL == cat->reverse_variable_map_long);
365 /* Interate over each variable, and add the value of that variable
366 to the appropriate map, if it's not already present. */
367 HMAP_FOR_EACH (vn, struct variable_node, node, &cat->varmap)
369 const int width = var_get_width (vn->var);
370 const union value *val = case_data (c, vn->var);
371 unsigned int hash = value_hash (val, width, 0);
373 struct value_node *valn = lookup_value (&vn->valmap, val, hash, width);
376 valn = pool_malloc (cat->pool, sizeof *valn);
377 valn->index = vn->n_vals++;
378 value_init (&valn->val, width);
379 value_copy (&valn->val, val, width);
380 hmap_insert (&vn->valmap, &valn->node, hash);
384 for (i = 0 ; i < cat->n_iap; ++i)
386 const struct interaction *iact = cat->iap[i].iact;
388 // if ( interaction_case_is_missing (iact, c, cat->exclude))
391 size_t hash = interaction_case_hash (iact, c, 0);
392 struct interaction_value *node = lookup_case (&cat->iap[i].ivmap, iact, c);
396 node = pool_malloc (cat->pool, sizeof *node);
398 node->ccase = case_ref (c);
401 hmap_insert (&cat->iap[i].ivmap, &node->node, hash);
403 if (cat->user_data_create)
404 node->user_data = cat->user_data_create (cat->aux1, cat->aux2);
410 cat->iap[i].cc += weight;
413 cat->update (node->user_data, cat->exclude, cat->wv, NULL, c, cat->aux1, cat->aux2);
417 /* Return the number of categories (distinct values) for interction N */
419 categoricals_n_count (const struct categoricals *cat, size_t n)
421 return hmap_count (&cat->iap[n].ivmap);
426 categoricals_df (const struct categoricals *cat, size_t n)
428 const struct interact_params *iap = &cat->iap[n];
429 return iap->df_prod[iap->iact->n_vars - 1];
433 /* Return the total number of categories */
435 categoricals_n_total (const struct categoricals *cat)
437 assert (cat->reverse_variable_map_long);
439 return cat->n_cats_total;
443 categoricals_df_total (const struct categoricals *cat)
448 /* This function must be called *before* any call to categoricals_get_*_by subscript and
449 *after* all calls to categoricals_update */
451 categoricals_done (const struct categoricals *cat_)
453 /* Implementation Note: Whilst this function is O(n) in cat->n_cats_total, in most
454 uses it will be more efficient that using a tree based structure, since it
455 is called only once, and means that subsequent lookups will be O(1).
457 1 call of O(n) + 10^9 calls of O(1) is better than 10^9 calls of O(log n).
459 struct categoricals *cat = CONST_CAST (struct categoricals *, cat_);
465 cat->n_cats_total = 0;
467 /* Calculate the degrees of freedom, and the number of categories */
468 for (i = 0 ; i < cat->n_iap; ++i)
471 const struct interaction *iact = cat->iap[i].iact;
473 cat->iap[i].df_prod = xcalloc (iact->n_vars, sizeof (int));
475 cat->iap[i].n_cats = 1;
477 for (v = 0 ; v < iact->n_vars; ++v)
479 const struct variable *var = iact->vars[v];
481 struct variable_node *vn = lookup_variable (&cat->varmap, var, hash_pointer (var, 0));
483 cat->iap[i].df_prod[v] = df * (hmap_count (&vn->valmap) - 1);
484 df = cat->iap[i].df_prod[v];
486 cat->iap[i].n_cats *= hmap_count (&vn->valmap);
489 cat->df_sum += cat->iap[i].df_prod [v - 1];
490 cat->n_cats_total += cat->iap[i].n_cats;
494 cat->reverse_variable_map_short = pool_calloc (cat->pool,
496 sizeof *cat->reverse_variable_map_short);
498 cat->reverse_variable_map_long = pool_calloc (cat->pool,
500 sizeof *cat->reverse_variable_map_long);
502 for (i = 0 ; i < cat->n_iap; ++i)
504 struct interaction_value *ivn = NULL;
507 struct interact_params *iap = &cat->iap[i];
509 iap->base_subscript_short = idx_short;
510 iap->base_subscript_long = idx_long;
512 iap->reverse_interaction_value_map = pool_calloc (cat->pool, iap->n_cats,
513 sizeof *iap->reverse_interaction_value_map);
515 HMAP_FOR_EACH (ivn, struct interaction_value, node, &iap->ivmap)
517 iap->reverse_interaction_value_map[x++] = ivn;
520 assert (x <= iap->n_cats);
522 /* For some purposes (eg CONTRASTS in ONEWAY) the values need to be sorted */
523 sort (iap->reverse_interaction_value_map, x, sizeof (*iap->reverse_interaction_value_map),
524 compare_interaction_value_3way, iap);
526 /* Fill the remaining values with null */
527 for (ii = x ; ii < iap->n_cats; ++ii)
528 iap->reverse_interaction_value_map[ii] = NULL;
530 /* Populate the reverse variable maps. */
531 for (ii = 0; ii < iap->df_prod [iap->iact->n_vars - 1]; ++ii)
532 cat->reverse_variable_map_short[idx_short++] = i;
534 for (ii = 0; ii < iap->n_cats; ++ii)
535 cat->reverse_variable_map_long[idx_long++] = i;
538 assert (cat->n_vars <= cat->n_iap);
540 // categoricals_dump (cat);
545 reverse_variable_lookup_short (const struct categoricals *cat, int subscript)
547 assert (cat->reverse_variable_map_short);
548 assert (subscript >= 0);
549 assert (subscript < cat->df_sum);
551 return cat->reverse_variable_map_short[subscript];
555 reverse_variable_lookup_long (const struct categoricals *cat, int subscript)
557 assert (cat->reverse_variable_map_long);
558 assert (subscript >= 0);
559 assert (subscript < cat->n_cats_total);
561 return cat->reverse_variable_map_long[subscript];
565 /* Return the interaction corresponding to SUBSCRIPT */
566 const struct interaction *
567 categoricals_get_interaction_by_subscript (const struct categoricals *cat, int subscript)
569 int index = reverse_variable_lookup_short (cat, subscript);
571 return cat->iap[index].iact;
574 /* Return the case corresponding to SUBSCRIPT */
575 static const struct ccase *
576 categoricals_get_case_by_subscript (const struct categoricals *cat, int subscript)
578 int vindex = reverse_variable_lookup_short (cat, subscript);
579 const struct interact_params *vp = &cat->iap[vindex];
580 const struct interaction_value *vn = vp->reverse_interaction_value_map [subscript - vp->base_subscript_short];
590 categoricals_get_weight_by_subscript (const struct categoricals *cat, int subscript)
592 int vindex = reverse_variable_lookup_short (cat, subscript);
593 const struct interact_params *vp = &cat->iap[vindex];
599 categoricals_get_sum_by_subscript (const struct categoricals *cat, int subscript)
601 int vindex = reverse_variable_lookup_short (cat, subscript);
602 const struct interact_params *vp = &cat->iap[vindex];
604 const struct interaction_value *iv = vp->reverse_interaction_value_map [subscript - vp->base_subscript_short];
612 /* Returns unity if the value in case C at SUBSCRIPT is equal to the category
613 for that subscript */
615 categoricals_get_binary_by_subscript (const struct categoricals *cat,
617 const struct ccase *c)
619 const struct interaction *iact = categoricals_get_interaction_by_subscript (cat, subscript);
621 const int i = reverse_variable_lookup_short (cat, subscript);
623 const int base_index = cat->iap[i].base_subscript_short;
628 for (v = 0; v < iact->n_vars; ++v)
630 const struct variable *var = iact->vars[v];
632 const union value *val = case_data (c, var);
633 const int width = var_get_width (var);
634 const struct variable_node *vn = lookup_variable (&cat->varmap, var, hash_pointer (var, 0));
636 const unsigned int hash = value_hash (val, width, 0);
637 const struct value_node *valn = lookup_value (&vn->valmap, val, hash, width);
641 /* Translate the subscript into an index for the individual variable */
642 int index = (subscript - base_index) % cat->iap[i].df_prod[v];
644 index /= cat->iap[i].df_prod[v - 1];
647 if ( valn->index == 0)
651 if ( valn->index != index )
662 categoricals_get_n_variables (const struct categoricals *cat)
664 printf ("%s\n", __FUNCTION__);
668 /* Return a case containing the set of values corresponding to SUBSCRIPT */
670 categoricals_get_case_by_category (const struct categoricals *cat, int subscript)
672 int vindex = reverse_variable_lookup_long (cat, subscript);
673 const struct interact_params *vp = &cat->iap[vindex];
674 const struct interaction_value *vn = vp->reverse_interaction_value_map [subscript - vp->base_subscript_long];
680 categoricals_get_user_data_by_category (const struct categoricals *cat, int subscript)
682 int vindex = reverse_variable_lookup_long (cat, subscript);
683 const struct interact_params *vp = &cat->iap[vindex];
685 const struct interaction_value *iv = vp->reverse_interaction_value_map [subscript - vp->base_subscript_long];
686 return iv->user_data;