/* PSPP - a program for statistical analysis.
- Copyright (C) 2009 Free Software Foundation, Inc.
+ Copyright (C) 2009, 2010, 2011 Free Software Foundation, Inc.
This program is free software: you can redistribute it and/or modify
it under the terms of the GNU General Public License as published by
#include <config.h>
+#include "math/categoricals.h"
+#include "math/interaction.h"
+
#include <stdio.h>
-#include "categoricals.h"
+#include "data/case.h"
+#include "data/value.h"
+#include "data/variable.h"
+#include "libpspp/array.h"
+#include "libpspp/hmap.h"
+#include "libpspp/pool.h"
+#include "libpspp/str.h"
+#include "libpspp/hash-functions.h"
+
+#include "gl/xalloc.h"
+
+struct value_node
+{
+ struct hmap_node node; /* Node in hash map. */
+
+ union value val; /* The value */
+
+};
+
+struct interaction_value
+{
+ struct hmap_node node; /* Node in hash map. */
+
+ struct ccase *ccase; /* A case (probably the first in the dataset) which matches this value */
+
+ /* Total of the weights of cases matching this interaction */
+ double cc;
+
+ void *user_data; /* A pointer to data which the caller can store stuff */
+};
+
+static struct value_node *
+lookup_value (const struct hmap *map, const union value *val, unsigned int hash, int width)
+{
+ struct value_node *vn = NULL;
+ HMAP_FOR_EACH_WITH_HASH (vn, struct value_node, node, hash, map)
+ {
+ if (value_equal (&vn->val, val, width))
+ break;
+ }
+
+ return vn;
+}
+
+struct variable_node
+{
+ struct hmap_node node; /* Node in hash map. */
+ const struct variable *var; /* The variable */
+
+ struct hmap valmap; /* A map of value nodes */
+};
+
+static void
+dump_interaction (const struct interaction *iact)
+{
+ struct string str = DS_EMPTY_INITIALIZER;
+ interaction_to_string (iact, &str);
+ printf ("Interaction: %s\n", ds_cstr (&str));
+ ds_destroy (&str);
+}
+
+
+static struct variable_node *
+lookup_variable (const struct hmap *map, const struct variable *var, unsigned int hash)
+{
+ struct variable_node *vn = NULL;
+ HMAP_FOR_EACH_WITH_HASH (vn, struct variable_node, node, hash, map)
+ {
+ if (vn->var == var)
+ break;
+
+ fprintf (stderr, "Warning: Hash table collision\n");
+ }
+
+ return vn;
+}
+
+
+struct interact_params
+{
+ /* A map indexed by a interaction_value */
+ struct hmap ivmap;
+
+ const struct interaction *iact;
+
+ int base_subscript_short;
+ int base_subscript_long;
+
+ /* The number of distinct values of this interaction */
+ int n_cats;
+
+ /* The degrees of freedom for this interaction */
+ int df;
+
+ /* A map of interaction_values indexed by subscript */
+ struct interaction_value **reverse_interaction_value_map;
+
+ double cc;
+};
+
+
+/* Comparison function to sort the reverse_value_map in ascending order */
+static int
+compare_interaction_value_3way (const void *vn1_, const void *vn2_, const void *aux)
+{
+ const struct interaction_value *const *vn1p = vn1_;
+ const struct interaction_value *const *vn2p = vn2_;
+
+ const struct interact_params *iap = aux;
-#include <gl/xalloc.h>
-#include <data/variable.h>
-#include <data/case.h>
-#include <data/value.h>
-#include <libpspp/hmap.h>
+ return interaction_case_cmp_3way (iap->iact, (*vn1p)->ccase, (*vn2p)->ccase);
+}
struct categoricals
{
- const struct variable **vars;
+ /* The weight variable */
+ const struct variable *wv;
+
+ /* An array of interact_params */
+ struct interact_params *iap;
+
+ /* Map whose members are the union of the variables which comprise IAP */
+ struct hmap varmap;
+
+ /* The size of IAP. (ie, the number of interactions involved.) */
+ size_t n_iap;
+
+ /* The number of categorical variables which contain entries.
+ In the absence of missing values, this will be equal to N_IAP */
size_t n_vars;
- const struct variable *wv;
+ size_t df_sum;
+
+ /* A map to enable the lookup of variables indexed by subscript.
+ This map considers only the N - 1 of the N variables.
+ */
+ int *reverse_variable_map_short;
+
+ /* Like the above, but uses all N variables */
+ int *reverse_variable_map_long;
+
+ size_t n_cats_total;
+
+ struct pool *pool;
+
+ /* Missing values to be excluded */
+ enum mv_class exclude;
- struct hmap *map;
+ /* Function to be called on each update */
+ update_func *update;
- int *next_index;
+ /* Function specified by the caller to create user_data */
+ user_data_create_func *user_data_create;
- size_t n_cats;
+ /* Auxilliary data to be passed to update and user_data_create_func*/
+ void *aux1;
+ void *aux2;
};
-struct value_node
- {
- struct hmap_node node; /* Node in hash map. */
- union value value; /* The value being labeled. */
- double cc; /* The total of the weights of cases with this value */
- int index; /* A zero based integer, unique within the variable.
- Can be used as an index into an array */
- };
+static void
+categoricals_dump (const struct categoricals *cat)
+{
+ int i;
+ printf ("Reverse Variable Map (short):\n");
+ for (i = 0; i < cat->df_sum; ++i)
+ {
+ printf (" %d", cat->reverse_variable_map_short[i]);
+ }
+ printf ("\n");
-static struct value_node *
-lookup_value (const struct hmap *map, const struct variable *var, const union value *val)
+ printf ("Reverse Variable Map (long):\n");
+ for (i = 0; i < cat->n_cats_total; ++i)
+ {
+ printf (" %d", cat->reverse_variable_map_long[i]);
+ }
+ printf ("\n");
+
+
+ printf ("Number of interactions %d\n", cat->n_iap);
+ for (i = 0 ; i < cat->n_iap; ++i)
+ {
+ int v;
+ struct string str;
+ const struct interact_params *iap = &cat->iap[i];
+ const struct interaction *iact = iap->iact;
+
+ ds_init_empty (&str);
+ interaction_to_string (iact, &str);
+
+ printf ("\nInteraction: %s (n: %d; df: %d ); ", ds_cstr (&str), iap->n_cats, iap->df);
+ ds_destroy (&str);
+ printf ("Base subscript: %d\n", iap->base_subscript_short);
+
+ printf ("\t(");
+ for (v = 0; v < hmap_count (&iap->ivmap); ++v)
+ {
+ int vv;
+ const struct interaction_value *iv = iap->reverse_interaction_value_map[v];
+
+ if (v > 0) printf (" ");
+ printf ("{");
+ for (vv = 0; vv < iact->n_vars; ++vv)
+ {
+ const struct variable *var = iact->vars[vv];
+ const union value *val = case_data (iv->ccase, var);
+
+ printf ("%g", val->f);
+ if (vv < iact->n_vars - 1)
+ printf (", ");
+ }
+ printf ("}");
+ }
+ printf (")\n");
+ }
+}
+
+
+void
+categoricals_destroy (struct categoricals *cat)
{
- struct value_node *foo;
- unsigned int width = var_get_width (var);
- size_t hash = value_hash (val, width, 0);
+ struct variable_node *vn = NULL;
+ int i;
+ for (i = 0; i < cat->n_iap; ++i)
+ {
+ struct interaction_value *iv = NULL;
+ /* Interate over each interaction value, and unref any cases that we reffed */
+ HMAP_FOR_EACH (iv, struct interaction_value, node, &cat->iap[i].ivmap)
+ {
+ case_unref (iv->ccase);
+ }
+ hmap_destroy (&cat->iap[i].ivmap);
+ }
- HMAP_FOR_EACH_WITH_HASH (foo, struct value_node, node, hash, map)
+ /* Interate over each variable and delete its value map */
+ HMAP_FOR_EACH (vn, struct variable_node, node, &cat->varmap)
{
- if (value_equal (val, &foo->value, width))
+ hmap_destroy (&vn->valmap);
+ }
+
+ hmap_destroy (&cat->varmap);
+
+ pool_destroy (cat->pool);
+
+ free (cat);
+}
+
+
+
+static struct interaction_value *
+lookup_case (const struct hmap *map, const struct interaction *iact, const struct ccase *c)
+{
+ struct interaction_value *iv = NULL;
+ size_t hash = interaction_case_hash (iact, c);
+
+ HMAP_FOR_EACH_WITH_HASH (iv, struct interaction_value, node, hash, map)
+ {
+ if (interaction_case_equal (iact, c, iv->ccase))
break;
+
+ fprintf (stderr, "Warning: Hash table collision\n");
}
- return foo;
+ return iv;
}
struct categoricals *
-categoricals_create (const struct variable **v, size_t n_vars, const struct variable *wv)
+categoricals_create (struct interaction *const*inter, size_t n_inter,
+ const struct variable *wv, enum mv_class exclude,
+ user_data_create_func *udf,
+ update_func *update, void *aux1, void *aux2
+ )
{
size_t i;
struct categoricals *cat = xmalloc (sizeof *cat);
- cat->vars = v;
- cat->n_vars = n_vars;
+ cat->n_iap = n_inter;
cat->wv = wv;
- cat->n_cats = 0;
-
- cat->map = xmalloc (sizeof *cat->map * n_vars);
- cat->next_index = xcalloc (sizeof *cat->next_index, n_vars);
-
- for (i = 0 ; i < cat->n_vars; ++i)
+ cat->n_cats_total = 0;
+ cat->n_vars = 0;
+ cat->reverse_variable_map_short = NULL;
+ cat->reverse_variable_map_long = NULL;
+ cat->pool = pool_create ();
+ cat->exclude = exclude;
+ cat->update = update;
+ cat->user_data_create = udf;
+
+ cat->aux1 = aux1;
+ cat->aux2 = aux2;
+
+ cat->iap = pool_calloc (cat->pool, cat->n_iap, sizeof *cat->iap);
+
+ hmap_init (&cat->varmap);
+ for (i = 0 ; i < cat->n_iap; ++i)
{
- hmap_init (&cat->map[i]);
+ int v;
+ hmap_init (&cat->iap[i].ivmap);
+ cat->iap[i].iact = inter[i];
+ cat->iap[i].cc = 0.0;
+ for (v = 0; v < inter[i]->n_vars; ++v)
+ {
+ const struct variable *var = inter[i]->vars[v];
+ unsigned int hash = hash_pointer (var, 0);
+ struct variable_node *vn = lookup_variable (&cat->varmap, var, hash);
+ if (vn == NULL)
+ {
+ vn = pool_malloc (cat->pool, sizeof *vn);
+ vn->var = var;
+ hmap_init (&vn->valmap);
+
+ hmap_insert (&cat->varmap, &vn->node, hash);
+ }
+ }
}
return cat;
}
+
void
categoricals_update (struct categoricals *cat, const struct ccase *c)
{
- size_t i;
-
+ int i;
+ struct variable_node *vn = NULL;
const double weight = cat->wv ? case_data (c, cat->wv)->f : 1.0;
- for (i = 0 ; i < cat->n_vars; ++i)
+ assert (NULL == cat->reverse_variable_map_short);
+ assert (NULL == cat->reverse_variable_map_long);
+
+ /* Interate over each variable, and add the value of that variable
+ to the appropriate map, if it's not already present. */
+ HMAP_FOR_EACH (vn, struct variable_node, node, &cat->varmap)
+ {
+ const int width = var_get_width (vn->var);
+ const union value *val = case_data (c, vn->var);
+ unsigned int hash = value_hash (val, width, 0);
+
+ struct value_node *valn = lookup_value (&vn->valmap, val, hash, width);
+ if (valn == NULL)
+ {
+ valn = pool_malloc (cat->pool, sizeof *valn);
+ value_init (&valn->val, width);
+ value_copy (&valn->val, val, width);
+ hmap_insert (&vn->valmap, &valn->node, hash);
+ }
+ }
+
+
+ for (i = 0 ; i < cat->n_iap; ++i)
{
- unsigned int width = var_get_width (cat->vars[i]);
- const union value *val = case_data (c, cat->vars[i]);
- size_t hash = value_hash (val, width, 0);
+ const struct interaction *iact = cat->iap[i].iact;
- struct value_node *node = lookup_value (&cat->map[i], cat->vars[i], val);
+ // if ( interaction_case_is_missing (iact, c, cat->exclude))
+ // continue;
+
+ size_t hash = interaction_case_hash (iact, c);
+ struct interaction_value *node = lookup_case (&cat->iap[i].ivmap, iact, c);
if ( NULL == node)
{
- node = xmalloc (sizeof *node);
+ node = pool_malloc (cat->pool, sizeof *node);
+
+ node->ccase = case_ref (c);
+ node->cc = weight;
- value_init (&node->value, width);
- value_copy (&node->value, val, width);
- node->cc = 0.0;
+ hmap_insert (&cat->iap[i].ivmap, &node->node, hash);
- hmap_insert (&cat->map[i], &node->node, hash);
- cat->n_cats ++;
- node->index = cat->next_index[i]++ ;
+ if (cat->user_data_create)
+ node->user_data = cat->user_data_create (cat->aux1, cat->aux2);
+ }
+ else
+ {
+ node->cc += weight;
}
+ cat->iap[i].cc += weight;
- node->cc += weight;
+ if (cat->update)
+ cat->update (node->user_data, cat->exclude, cat->wv, NULL, c, cat->aux1, cat->aux2);
}
}
-/* Return the number of categories (distinct values) for variable N */
+/* Return the number of categories (distinct values) for interction N */
size_t
categoricals_n_count (const struct categoricals *cat, size_t n)
{
- return hmap_count (&cat->map[n]);
+ return hmap_count (&cat->iap[n].ivmap);
+}
+
+
+size_t
+categoricals_df (const struct categoricals *cat, size_t n)
+{
+ return cat->iap[n].df;
+}
+
+
+/* Return the total number of categories */
+size_t
+categoricals_n_total (const struct categoricals *cat)
+{
+ assert (cat->reverse_variable_map_long);
+
+ return cat->n_cats_total;
+}
+
+size_t
+categoricals_df_total (const struct categoricals *cat)
+{
+ return cat->df_sum;
+}
+
+/* This function must be called *before* any call to categoricals_get_*_by subscript and
+ *after* all calls to categoricals_update */
+void
+categoricals_done (const struct categoricals *cat_)
+{
+ /* Implementation Note: Whilst this function is O(n) in cat->n_cats_total, in most
+ uses it will be more efficient that using a tree based structure, since it
+ is called only once, and means that subsequent lookups will be O(1).
+
+ 1 call of O(n) + 10^9 calls of O(1) is better than 10^9 calls of O(log n).
+ */
+ struct categoricals *cat = CONST_CAST (struct categoricals *, cat_);
+ int v;
+ int i;
+ int idx_short = 0;
+ int idx_long = 0;
+ cat->df_sum = 0;
+ cat->n_cats_total = 0;
+
+ /* Calculate the degrees of freedom, and the number of categories */
+ for (i = 0 ; i < cat->n_iap; ++i)
+ {
+ const struct interaction *iact = cat->iap[i].iact;
+
+ cat->iap[i].df = 1;
+ cat->iap[i].n_cats = 1;
+
+ for (v = 0 ; v < iact->n_vars; ++v)
+ {
+ const struct variable *var = iact->vars[v];
+
+ struct variable_node *vn = lookup_variable (&cat->varmap, var, hash_pointer (var, 0));
+
+ cat->iap[i].df *= hmap_count (&vn->valmap) - 1;
+ cat->iap[i].n_cats *= hmap_count (&vn->valmap);
+ }
+
+ cat->df_sum += cat->iap[i].df;
+ cat->n_cats_total += cat->iap[i].n_cats;
+ }
+
+
+ cat->reverse_variable_map_short = pool_calloc (cat->pool,
+ cat->df_sum,
+ sizeof *cat->reverse_variable_map_short);
+
+ cat->reverse_variable_map_long = pool_calloc (cat->pool,
+ cat->n_cats_total,
+ sizeof *cat->reverse_variable_map_long);
+
+ for (i = 0 ; i < cat->n_iap; ++i)
+ {
+ struct interaction_value *ivn = NULL;
+ int x = 0;
+ int ii;
+ struct interact_params *iap = &cat->iap[i];
+
+ iap->base_subscript_short = idx_short;
+ iap->base_subscript_long = idx_long;
+
+ iap->reverse_interaction_value_map = pool_calloc (cat->pool, iap->n_cats,
+ sizeof *iap->reverse_interaction_value_map);
+
+ HMAP_FOR_EACH (ivn, struct interaction_value, node, &iap->ivmap)
+ {
+ iap->reverse_interaction_value_map[x++] = ivn;
+ }
+
+ assert (x <= iap->n_cats);
+
+ /* For some purposes (eg CONTRASTS in ONEWAY) the values need to be sorted */
+ sort (iap->reverse_interaction_value_map, x, sizeof (*iap->reverse_interaction_value_map),
+ compare_interaction_value_3way, iap);
+
+ /* Fill the remaining values with null */
+ for (ii = x ; ii < iap->n_cats; ++ii)
+ iap->reverse_interaction_value_map[ii] = NULL;
+
+ /* Populate the reverse variable maps. */
+ for (ii = 0; ii < iap->df; ++ii)
+ cat->reverse_variable_map_short[idx_short++] = i;
+
+ for (ii = 0; ii < iap->n_cats; ++ii)
+ cat->reverse_variable_map_long[idx_long++] = i;
+ }
+
+ assert (cat->n_vars <= cat->n_iap);
+
+ // categoricals_dump (cat);
+}
+
+
+static int
+reverse_variable_lookup_short (const struct categoricals *cat, int subscript)
+{
+ assert (cat->reverse_variable_map_short);
+ assert (subscript >= 0);
+ assert (subscript < cat->df_sum);
+
+ return cat->reverse_variable_map_short[subscript];
+}
+
+static int
+reverse_variable_lookup_long (const struct categoricals *cat, int subscript)
+{
+ assert (cat->reverse_variable_map_long);
+ assert (subscript >= 0);
+ assert (subscript < cat->n_cats_total);
+
+ return cat->reverse_variable_map_long[subscript];
+}
+
+
+/* Return the interaction corresponding to SUBSCRIPT */
+const struct interaction *
+categoricals_get_interaction_by_subscript (const struct categoricals *cat, int subscript)
+{
+ int index = reverse_variable_lookup_short (cat, subscript);
+
+ return cat->iap[index].iact;
}
-/* Return the index for value VAL in the Nth variable */
-int
-categoricals_index (const struct categoricals *cat, size_t n, const union value *val)
+/* Return the case corresponding to SUBSCRIPT */
+static const struct ccase *
+categoricals_get_case_by_subscript (const struct categoricals *cat, int subscript)
{
- struct value_node *vn = lookup_value (&cat->map[n], cat->vars[n], val);
+ int vindex = reverse_variable_lookup_short (cat, subscript);
+ const struct interact_params *vp = &cat->iap[vindex];
+ const struct interaction_value *vn = vp->reverse_interaction_value_map [subscript - vp->base_subscript_short];
if ( vn == NULL)
- return -1;
+ return NULL;
- return vn->index;
+ return vn->ccase;
}
-/* Return the total number of categories */
-size_t
-categoricals_total (const struct categoricals *cat)
+double
+categoricals_get_weight_by_subscript (const struct categoricals *cat, int subscript)
{
- return cat->n_cats;
+ int vindex = reverse_variable_lookup_short (cat, subscript);
+ const struct interact_params *vp = &cat->iap[vindex];
+
+ return vp->cc;
}
+double
+categoricals_get_sum_by_subscript (const struct categoricals *cat, int subscript)
+{
+ int vindex = reverse_variable_lookup_short (cat, subscript);
+ const struct interact_params *vp = &cat->iap[vindex];
+ const struct interaction_value *iv = vp->reverse_interaction_value_map [subscript - vp->base_subscript_short];
+ if (iv == NULL)
+ return 0;
+ return iv->cc;
+}
+
+/* Returns unity if the value in case C at SUBSCRIPT is equal to the category
+ for that subscript */
+double
+categoricals_get_binary_by_subscript (const struct categoricals *cat, int subscript,
+ const struct ccase *c)
+{
+ const struct interaction *iact = categoricals_get_interaction_by_subscript (cat, subscript);
+ const struct ccase *c2 = categoricals_get_case_by_subscript (cat, subscript);
+
+ if ( c2 == NULL)
+ return 0;
+
+ return interaction_case_equal (iact, c, c2);
+}
+
+size_t
+categoricals_get_n_variables (const struct categoricals *cat)
+{
+ printf ("%s\n", __FUNCTION__);
+ return cat->n_vars;
+}
+
+/* Return a case containing the set of values corresponding to SUBSCRIPT */
+const struct ccase *
+categoricals_get_case_by_category (const struct categoricals *cat, int subscript)
+{
+ int vindex = reverse_variable_lookup_long (cat, subscript);
+ const struct interact_params *vp = &cat->iap[vindex];
+ const struct interaction_value *vn = vp->reverse_interaction_value_map [subscript - vp->base_subscript_long];
+
+ return vn->ccase;
+}
+
+void *
+categoricals_get_user_data_by_category (const struct categoricals *cat, int subscript)
+{
+ int vindex = reverse_variable_lookup_long (cat, subscript);
+ const struct interact_params *vp = &cat->iap[vindex];
+
+ const struct interaction_value *iv = vp->reverse_interaction_value_map [subscript - vp->base_subscript_long];
+ return iv->user_data;
+}