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"
35 #define EFFECTS_CODING 0
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;
181 /* Function to be called on each update */
184 /* Function specified by the caller to create user_data */
185 user_data_create_func *user_data_create;
187 /* Auxilliary data to be passed to update and user_data_create_func*/
194 categoricals_dump (const struct categoricals *cat)
198 printf ("Reverse Variable Map (short):\n");
199 for (i = 0; i < cat->df_sum; ++i)
201 printf (" %d", cat->reverse_variable_map_short[i]);
205 printf ("Reverse Variable Map (long):\n");
206 for (i = 0; i < cat->n_cats_total; ++i)
208 printf (" %d", cat->reverse_variable_map_long[i]);
213 printf ("Number of interactions %d\n", cat->n_iap);
214 for (i = 0 ; i < cat->n_iap; ++i)
218 const struct interact_params *iap = &cat->iap[i];
219 const struct interaction *iact = iap->iact;
221 ds_init_empty (&str);
222 interaction_to_string (iact, &str);
224 printf ("\nInteraction: %s (n: %d; df: %d ); ", ds_cstr (&str), iap->n_cats, iap->df);
226 printf ("Base subscript: %d\n", iap->base_subscript_short);
229 for (v = 0; v < hmap_count (&iap->ivmap); ++v)
232 const struct interaction_value *iv = iap->reverse_interaction_value_map[v];
234 if (v > 0) printf (" ");
236 for (vv = 0; vv < iact->n_vars; ++vv)
238 const struct variable *var = iact->vars[vv];
239 const union value *val = case_data (iv->ccase, var);
241 printf ("%g", val->f);
242 if (vv < iact->n_vars - 1)
253 categoricals_destroy (struct categoricals *cat)
255 struct variable_node *vn = NULL;
259 for (i = 0; i < cat->n_iap; ++i)
261 struct interaction_value *iv = NULL;
262 /* Interate over each interaction value, and unref any cases that we reffed */
263 HMAP_FOR_EACH (iv, struct interaction_value, node, &cat->iap[i].ivmap)
265 case_unref (iv->ccase);
267 hmap_destroy (&cat->iap[i].ivmap);
270 /* Interate over each variable and delete its value map */
271 HMAP_FOR_EACH (vn, struct variable_node, node, &cat->varmap)
273 hmap_destroy (&vn->valmap);
276 hmap_destroy (&cat->varmap);
278 pool_destroy (cat->pool);
285 static struct interaction_value *
286 lookup_case (const struct hmap *map, const struct interaction *iact, const struct ccase *c)
288 struct interaction_value *iv = NULL;
289 size_t hash = interaction_case_hash (iact, c, 0);
291 HMAP_FOR_EACH_WITH_HASH (iv, struct interaction_value, node, hash, map)
293 if (interaction_case_equal (iact, c, iv->ccase))
296 fprintf (stderr, "Warning: Hash table collision\n");
303 struct categoricals *
304 categoricals_create (struct interaction *const*inter, size_t n_inter,
305 const struct variable *wv, enum mv_class exclude,
306 user_data_create_func *udf,
307 update_func *update, void *aux1, void *aux2
311 struct categoricals *cat = xmalloc (sizeof *cat);
313 cat->n_iap = n_inter;
315 cat->n_cats_total = 0;
317 cat->reverse_variable_map_short = NULL;
318 cat->reverse_variable_map_long = NULL;
319 cat->pool = pool_create ();
320 cat->exclude = exclude;
321 cat->update = update;
322 cat->user_data_create = udf;
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;
391 // if ( interaction_case_is_missing (iact, c, cat->exclude))
394 size_t hash = interaction_case_hash (iact, c, 0);
395 struct interaction_value *node = lookup_case (&cat->iap[i].ivmap, iact, c);
399 node = pool_malloc (cat->pool, sizeof *node);
401 node->ccase = case_ref (c);
404 hmap_insert (&cat->iap[i].ivmap, &node->node, hash);
406 if (cat->user_data_create)
407 node->user_data = cat->user_data_create (cat->aux1, cat->aux2);
413 cat->iap[i].cc += weight;
416 cat->update (node->user_data, cat->exclude, cat->wv, NULL, c, cat->aux1, cat->aux2);
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 assert (cat->reverse_variable_map_long);
442 return cat->n_cats_total;
446 categoricals_df_total (const struct categoricals *cat)
451 /* This function must be called *before* any call to categoricals_get_*_by subscript and
452 *after* all calls to categoricals_update */
454 categoricals_done (const struct categoricals *cat_)
456 /* Implementation Note: Whilst this function is O(n) in cat->n_cats_total, in most
457 uses it will be more efficient that using a tree based structure, since it
458 is called only once, and means that subsequent lookups will be O(1).
460 1 call of O(n) + 10^9 calls of O(1) is better than 10^9 calls of O(log n).
462 struct categoricals *cat = CONST_CAST (struct categoricals *, cat_);
468 cat->n_cats_total = 0;
470 /* Calculate the degrees of freedom, and the number of categories */
471 for (i = 0 ; i < cat->n_iap; ++i)
474 const struct interaction *iact = cat->iap[i].iact;
476 cat->iap[i].df_prod = xcalloc (iact->n_vars, sizeof (int));
478 cat->iap[i].n_cats = 1;
480 for (v = 0 ; v < iact->n_vars; ++v)
482 const struct variable *var = iact->vars[v];
484 struct variable_node *vn = lookup_variable (&cat->varmap, var, hash_pointer (var, 0));
486 cat->iap[i].df_prod[v] = df * (hmap_count (&vn->valmap) - 1);
487 df = cat->iap[i].df_prod[v];
489 cat->iap[i].n_cats *= hmap_count (&vn->valmap);
492 cat->df_sum += cat->iap[i].df_prod [v - 1];
493 cat->n_cats_total += cat->iap[i].n_cats;
497 cat->reverse_variable_map_short = pool_calloc (cat->pool,
499 sizeof *cat->reverse_variable_map_short);
501 cat->reverse_variable_map_long = pool_calloc (cat->pool,
503 sizeof *cat->reverse_variable_map_long);
505 for (i = 0 ; i < cat->n_iap; ++i)
507 struct interaction_value *ivn = NULL;
510 struct interact_params *iap = &cat->iap[i];
512 iap->base_subscript_short = idx_short;
513 iap->base_subscript_long = idx_long;
515 iap->reverse_interaction_value_map = pool_calloc (cat->pool, iap->n_cats,
516 sizeof *iap->reverse_interaction_value_map);
518 HMAP_FOR_EACH (ivn, struct interaction_value, node, &iap->ivmap)
520 iap->reverse_interaction_value_map[x++] = ivn;
523 assert (x <= iap->n_cats);
525 /* For some purposes (eg CONTRASTS in ONEWAY) the values need to be sorted */
526 sort (iap->reverse_interaction_value_map, x, sizeof (*iap->reverse_interaction_value_map),
527 compare_interaction_value_3way, iap);
529 /* Fill the remaining values with null */
530 for (ii = x ; ii < iap->n_cats; ++ii)
531 iap->reverse_interaction_value_map[ii] = NULL;
533 /* Populate the reverse variable maps. */
534 for (ii = 0; ii < iap->df_prod [iap->iact->n_vars - 1]; ++ii)
535 cat->reverse_variable_map_short[idx_short++] = i;
537 for (ii = 0; ii < iap->n_cats; ++ii)
538 cat->reverse_variable_map_long[idx_long++] = i;
541 assert (cat->n_vars <= cat->n_iap);
543 // categoricals_dump (cat);
545 /* Tally up the sums for all the encodings */
546 for (i = 0 ; i < cat->n_iap; ++i)
549 struct interact_params *iap = &cat->iap[i];
550 const struct interaction *iact = iap->iact;
552 const int df = iap->df_prod [iact->n_vars - 1];
554 iap->enc_sum = xcalloc (df, sizeof (*(iap->enc_sum)));
556 for (y = 0; y < hmap_count (&iap->ivmap); ++y)
558 struct interaction_value *iv = iap->reverse_interaction_value_map[y];
559 for (x = iap->base_subscript_short; x < iap->base_subscript_short + df ;++x)
561 const double bin = categoricals_get_binary_by_subscript (cat, x, iv->ccase); \
562 iap->enc_sum [x - iap->base_subscript_short] += bin * iv->cc;
570 reverse_variable_lookup_short (const struct categoricals *cat, int subscript)
572 assert (cat->reverse_variable_map_short);
573 assert (subscript >= 0);
574 assert (subscript < cat->df_sum);
576 return cat->reverse_variable_map_short[subscript];
580 reverse_variable_lookup_long (const struct categoricals *cat, int subscript)
582 assert (cat->reverse_variable_map_long);
583 assert (subscript >= 0);
584 assert (subscript < cat->n_cats_total);
586 return cat->reverse_variable_map_long[subscript];
590 /* Return the interaction corresponding to SUBSCRIPT */
591 const struct interaction *
592 categoricals_get_interaction_by_subscript (const struct categoricals *cat, int subscript)
594 int index = reverse_variable_lookup_short (cat, subscript);
596 return cat->iap[index].iact;
599 /* Return the case corresponding to SUBSCRIPT */
600 static const struct ccase *
601 categoricals_get_case_by_subscript (const struct categoricals *cat, int subscript)
603 int vindex = reverse_variable_lookup_short (cat, subscript);
604 const struct interact_params *vp = &cat->iap[vindex];
605 const struct interaction_value *vn = vp->reverse_interaction_value_map [subscript - vp->base_subscript_short];
615 categoricals_get_weight_by_subscript (const struct categoricals *cat, int subscript)
617 int vindex = reverse_variable_lookup_short (cat, subscript);
618 const struct interact_params *vp = &cat->iap[vindex];
624 categoricals_get_sum_by_subscript (const struct categoricals *cat, int subscript)
626 int vindex = reverse_variable_lookup_short (cat, subscript);
627 const struct interact_params *vp = &cat->iap[vindex];
629 return vp->enc_sum[subscript - vp->base_subscript_short];
632 /* Returns unity if the value in case C at SUBSCRIPT is equal to the category
633 for that subscript */
635 categoricals_get_binary_by_subscript (const struct categoricals *cat,
637 const struct ccase *c)
639 const struct interaction *iact = categoricals_get_interaction_by_subscript (cat, subscript);
641 const int i = reverse_variable_lookup_short (cat, subscript);
643 const int base_index = cat->iap[i].base_subscript_short;
648 const struct interact_params *iap = &cat->iap[i];
650 for (v = 0; v < iact->n_vars; ++v)
652 const struct variable *var = iact->vars[v];
654 const union value *val = case_data (c, var);
655 const int width = var_get_width (var);
656 const struct variable_node *vn = lookup_variable (&cat->varmap, var, hash_pointer (var, 0));
658 const unsigned int hash = value_hash (val, width, 0);
659 const struct value_node *valn = lookup_value (&vn->valmap, val, hash, width);
663 /* Translate the subscript into an index for the individual variable */
664 int index = (subscript - base_index) % iap->df_prod[v];
666 index /= iap->df_prod[v - 1];
669 if ( valn->index == 0)
673 if ( valn->index != index )
684 categoricals_get_n_variables (const struct categoricals *cat)
686 printf ("%s\n", __FUNCTION__);
690 /* Return a case containing the set of values corresponding to SUBSCRIPT */
692 categoricals_get_case_by_category (const struct categoricals *cat, int subscript)
694 int vindex = reverse_variable_lookup_long (cat, subscript);
695 const struct interact_params *vp = &cat->iap[vindex];
696 const struct interaction_value *vn = vp->reverse_interaction_value_map [subscript - vp->base_subscript_long];
702 categoricals_get_user_data_by_category (const struct categoricals *cat, int subscript)
704 int vindex = reverse_variable_lookup_long (cat, subscript);
705 const struct interact_params *vp = &cat->iap[vindex];
707 const struct interaction_value *iv = vp->reverse_interaction_value_map [subscript - vp->base_subscript_long];
708 return iv->user_data;