/* PSPP - a program for statistical analysis.
- Copyright (C) 2009, 2010, 2011, 2012 Free Software Foundation, Inc.
+ Copyright (C) 2009, 2010, 2011, 2012, 2014 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 "math/categoricals.h"
#include "math/interaction.h"
+#include <float.h>
#include <stdio.h>
#include "data/case.h"
if (value_equal (&vn->val, val, width))
break;
}
-
+
return vn;
}
const struct variable *var; /* The variable */
struct hmap valmap; /* A map of value nodes */
- int n_vals; /* Number of values for this variable */
-
- int *indirection; /* An array (of size n_vals) of integers, which serve to
- permute the index members of the values in valmap.
-
- Doing this, means that categories are considered in the order
- of their values. Mathematically the order is irrelevant.
- However certain procedures (eg logistic regression) want to report
- statisitics for particular categories */
-
};
{
if (vn->var == var)
break;
-
- fprintf (stderr, "Warning: Hash table collision\n");
}
-
+
return vn;
}
struct interact_params
{
- /* A map of cases indexed by a interaction_value */
- struct hmap ivmap;
+ /* The interaction, and an array with iact->n_vars elements such that
+ varnodes[x] points to the variable_node for iact->vars[x]. */
+ const struct interaction *iact;
+ struct variable_node **varnodes;
- struct interaction *iact;
+ /* An example of each interaction that appears in the data, like a frequency
+ table for 'iact'. By construction, the number of elements must be less
+ than or equal to 'n_cats'.
- int base_subscript_short;
- int base_subscript_long;
+ categoricals_update() updates 'ivmap' case-by-case, then
+ categoricals_done() dumps 'ivmap' into 'ivs' and sorts it. */
+ struct hmap ivmap;
+ struct interaction_value **ivs;
+
+ int base_df;
+ int base_cats;
- /* The number of distinct values of this interaction */
+ /* Product of hmap_count(&varnodes[*]->valmap), that is, the maximum number
+ of distinct values of this interaction. */
int n_cats;
/* An array of integers df_n * df_{n-1} * df_{n-2} ...
- These are the products of the degrees of freedom for the current
- variable and all preceeding variables */
- int *df_prod;
+ These are the products of the degrees of freedom for the current
+ variable and all preceding variables */
+ int *df_prod;
double *enc_sum;
- /* A map of interaction_values indexed by subscript */
- struct interaction_value **reverse_interaction_value_map;
-
+ /* Sum of ivs[*]->cc. */
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)
{
/* An array of interact_params */
struct interact_params *iap;
+ size_t n_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;
- 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;
+ int *df_to_iact; /* 'df_sum' elements. */
+ size_t df_sum;
/* Like the above, but uses all N variables */
- int *reverse_variable_map_long;
-
+ int *cat_to_iact; /* 'n_cats_total' elements. */
size_t n_cats_total;
struct pool *pool;
- /* Missing values in the dependent varirable to be excluded */
- enum mv_class dep_excl;
-
/* Missing values in the factor variables to be excluded */
enum mv_class fctr_excl;
const struct payload *payload;
};
+
+bool
+categoricals_isbalanced (const struct categoricals *cat)
+{
+ int i;
+
+ for (i = 0 ; i < cat->n_iap; ++i)
+ {
+ int v;
+ const struct interact_params *iap = &cat->iap[i];
+
+ double oval = -1.0;
+ for (v = 0; v < hmap_count (&iap->ivmap); ++v)
+ {
+ const struct interaction_value *iv = iap->ivs[v];
+ if (oval == -1.0)
+ oval = iv->cc;
+ if (oval != iv->cc)
+ return false;
+ }
+ }
+ return true;
+}
+
+
static void
categoricals_dump (const struct categoricals *cat)
{
{
int i;
- printf ("Reverse Variable Map (short):\n");
+ printf ("df to interaction map:\n");
for (i = 0; i < cat->df_sum; ++i)
{
- printf (" %d", cat->reverse_variable_map_short[i]);
+ printf (" %d", cat->df_to_iact[i]);
}
printf ("\n");
- printf ("Reverse Variable Map (long):\n");
+ printf ("Category to interaction map:\n");
for (i = 0; i < cat->n_cats_total; ++i)
{
- printf (" %d", cat->reverse_variable_map_long[i]);
+ printf (" %d", cat->cat_to_iact[i]);
}
printf ("\n");
- printf ("Number of interactions %d\n", cat->n_iap);
+ printf ("Number of interactions %zu\n", cat->n_iap);
for (i = 0 ; i < cat->n_iap; ++i)
{
int v;
printf ("\nInteraction: \"%s\" (number of categories: %d); ", ds_cstr (&str), iap->n_cats);
ds_destroy (&str);
- printf ("Base index (short/long): %d/%d\n", iap->base_subscript_short, iap->base_subscript_long);
+ printf ("Base index (df/categories): %d/%d\n", iap->base_df, iap->base_cats);
printf ("\t(");
for (v = 0; v < hmap_count (&iap->ivmap); ++v)
{
int vv;
- const struct interaction_value *iv = iap->reverse_interaction_value_map[v];
+ const struct interaction_value *iv = iap->ivs[v];
if (v > 0) printf (" ");
printf ("{");
{
const struct variable *var = iact->vars[vv];
const union value *val = case_data (iv->ccase, var);
- unsigned int varhash = hash_pointer (var, 0);
- struct variable_node *vn = lookup_variable (&cat->varmap, var, varhash);
+ struct variable_node *vn = iap->varnodes[vv];
const int width = var_get_width (var);
unsigned int valhash = value_hash (val, width, 0);
assert (vn->var == var);
- printf ("%g(%d)", val->f, valn->index);
+ printf ("%.*g(%d)", DBL_DIG + 1, val->f, valn->index);
if (vv < iact->n_vars - 1)
printf (", ");
}
int i;
if (NULL == cat)
return;
+
for (i = 0; i < cat->n_iap; ++i)
{
struct interaction_value *iv = NULL;
free (cat->iap[i].enc_sum);
free (cat->iap[i].df_prod);
hmap_destroy (&cat->iap[i].ivmap);
- interaction_destroy (cat->iap[i].iact);
}
/* Interate over each variable and delete its value map */
{
if (interaction_case_equal (iact, c, iv->ccase))
break;
-
- fprintf (stderr, "Warning: Hash table collision\n");
}
return iv;
}
-bool
+/* Returns true iff CAT is sane, that is, if it is complete and has at least
+ one value. */
+bool
categoricals_sane (const struct categoricals *cat)
{
return cat->sane;
}
+/* Creates and returns a new categoricals object whose variables come from the
+ N_INTER interactions objects in the array starting at INTER. (The INTER
+ objects must outlive the categoricals object because it uses them
+ internally.)
+
+ FCTR_EXCL determines which cases are listwise ignored by
+ categoricals_update(). */
struct categoricals *
categoricals_create (struct interaction *const*inter, size_t n_inter,
- const struct variable *wv, enum mv_class dep_excl, enum mv_class fctr_excl)
+ const struct variable *wv, enum mv_class fctr_excl)
{
size_t i;
struct categoricals *cat = xmalloc (sizeof *cat);
-
+
cat->n_iap = n_inter;
cat->wv = wv;
cat->n_cats_total = 0;
cat->n_vars = 0;
- cat->reverse_variable_map_short = NULL;
- cat->reverse_variable_map_long = NULL;
+ cat->df_to_iact = NULL;
+ cat->cat_to_iact = NULL;
cat->pool = pool_create ();
- cat->dep_excl = dep_excl;
cat->fctr_excl = fctr_excl;
cat->payload = NULL;
cat->aux2 = NULL;
hmap_init (&cat->iap[i].ivmap);
cat->iap[i].iact = inter[i];
cat->iap[i].cc = 0.0;
+ cat->iap[i].varnodes = pool_nmalloc (cat->pool, cat->iap[i].iact->n_vars,
+ sizeof *cat->iap[i].varnodes);
for (v = 0; v < inter[i]->n_vars; ++v)
{
const struct variable *var = inter[i]->vars[v];
{
vn = pool_malloc (cat->pool, sizeof *vn);
vn->var = var;
- vn->n_vals = 0;
hmap_init (&vn->valmap);
hmap_insert (&cat->varmap, &vn->node, hash);
}
+ cat->iap[i].varnodes[v] = vn;
}
}
return;
weight = cat->wv ? case_data (c, cat->wv)->f : 1.0;
+ weight = var_force_valid_weight (cat->wv, weight, NULL);
- assert (NULL == cat->reverse_variable_map_short);
- assert (NULL == cat->reverse_variable_map_long);
+ assert (NULL == cat->df_to_iact);
+ assert (NULL == cat->cat_to_iact);
/* Interate over each variable, and add the value of that variable
to the appropriate map, if it's not already present. */
if (valn == NULL)
{
valn = pool_malloc (cat->pool, sizeof *valn);
- valn->index = vn->n_vals++;
+ valn->index = -1;
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)
{
const struct interaction *iact = cat->iap[i].iact;
hmap_insert (&cat->iap[i].ivmap, &node->node, hash);
- if (cat->payload)
+ if (cat->payload)
{
node->user_data = cat->payload->create (cat->aux1, cat->aux2);
}
if (cat->payload)
{
- double weight = cat->wv ? case_data (c, cat->wv)->f : 1.0;
cat->payload->update (cat->aux1, cat->aux2, node->user_data, c, weight);
}
-
}
}
-/* Return the number of categories (distinct values) for interction N */
+/* Return the number of categories (distinct values) for interaction IDX in
+ CAT. */
size_t
categoricals_n_count (const struct categoricals *cat, size_t n)
{
}
+/* Returns the number of degrees of freedom for interaction IDX within CAT. */
size_t
categoricals_df (const struct categoricals *cat, size_t n)
{
}
-/* Return the total number of categories */
+/* Return the total number of categories across all interactions in CAT. */
size_t
categoricals_n_total (const struct categoricals *cat)
{
return cat->n_cats_total;
}
+/* Returns the total degrees of freedom for CAT. */
size_t
categoricals_df_total (const struct categoricals *cat)
{
return cat->df_sum;
}
+/* Returns true iff categoricals_done() has been called for CAT. */
bool
categoricals_is_complete (const struct categoricals *cat)
{
- return (NULL != cat->reverse_variable_map_short);
+ return (NULL != cat->df_to_iact);
}
-/* This function must be called *before* any call to categoricals_get_*_by subscript and
- *after* all calls to categoricals_update */
+/* This function must be called (once) before any call to the *_by_subscript or
+ *_by_category functions, but AFTER any calls to categoricals_update. If this
+ function returns false, then no calls to _by_subscript or *_by_category are
+ allowed. */
void
categoricals_done (const struct categoricals *cat_)
{
struct categoricals *cat = CONST_CAST (struct categoricals *, cat_);
int v;
int i;
- int idx_short = 0;
- int idx_long = 0;
+ int idx_df = 0;
+ int idx_cat = 0;
if (NULL == cat)
return;
cat->df_sum = 0;
cat->n_cats_total = 0;
+ /* Assign 'index' to each variables' value_nodes, counting up from 0 in
+ ascending order by value. */
+ struct variable_node *vn;
+ HMAP_FOR_EACH (vn, struct variable_node, node, &cat->varmap)
+ {
+ size_t n_vals = hmap_count (&vn->valmap);
+ if (!n_vals)
+ {
+ cat->sane = false;
+ return;
+ }
+
+ struct value_node **nodes = xcalloc (sizeof *nodes, n_vals);
+ int x = 0;
+ struct value_node *valnd;
+ HMAP_FOR_EACH (valnd, struct value_node, node, &vn->valmap)
+ nodes[x++] = valnd;
+ sort (nodes, n_vals, sizeof *nodes, compare_value_node_3way, vn);
+ for (x = 0; x < n_vals; ++x)
+ nodes[x]->index = x;
+ free (nodes);
+ }
+
/* Calculate the degrees of freedom, and the number of categories */
for (i = 0 ; i < cat->n_iap; ++i)
{
cat->iap[i].df_prod = iact->n_vars ? xcalloc (iact->n_vars, sizeof (int)) : NULL;
cat->iap[i].n_cats = 1;
-
+
for (v = 0 ; v < iact->n_vars; ++v)
{
- int x;
- const struct variable *var = iact->vars[v];
-
- struct variable_node *vn = lookup_variable (&cat->varmap, var, hash_pointer (var, 0));
-
- struct value_node *valnd = NULL;
- struct value_node **array ;
-
- assert (vn->n_vals == hmap_count (&vn->valmap));
-
- if (vn->n_vals == 0)
- {
- cat->sane = false;
- return;
- }
-
- vn->indirection = pool_calloc (cat->pool, vn->n_vals, sizeof *vn->indirection);
-
- /* Sort the VALMAP here */
- array = xcalloc (sizeof *array, vn->n_vals);
- HMAP_FOR_EACH (valnd, struct value_node, node, &vn->valmap)
- {
- /* Note: This loop is probably superfluous, it could be done in the
- update stage (at the expense of a realloc) */
- array[valnd->index] = valnd;
- }
-
- sort (array, vn->n_vals, sizeof (*array),
- compare_value_node_3way, vn);
-
- for (x = 0; x < vn->n_vals; ++x)
- {
- struct value_node *vvv = array[x];
- vn->indirection[vn->n_vals - x - 1] = vvv->index;
- }
- free (array);
-
- cat->iap[i].df_prod[v] = df * (vn->n_vals - 1);
+ struct variable_node *vn = cat->iap[i].varnodes[v];
+ size_t n_vals = hmap_count (&vn->valmap);
+ cat->iap[i].df_prod[v] = df * (n_vals - 1);
df = cat->iap[i].df_prod[v];
- cat->iap[i].n_cats *= vn->n_vals;
+ cat->iap[i].n_cats *= n_vals;
}
if (v > 0)
}
- cat->reverse_variable_map_short = pool_calloc (cat->pool,
- cat->df_sum,
- sizeof *cat->reverse_variable_map_short);
+ cat->df_to_iact = pool_calloc (cat->pool, cat->df_sum,
+ sizeof *cat->df_to_iact);
- cat->reverse_variable_map_long = pool_calloc (cat->pool,
- cat->n_cats_total,
- sizeof *cat->reverse_variable_map_long);
+ cat->cat_to_iact = pool_calloc (cat->pool, cat->n_cats_total,
+ sizeof *cat->cat_to_iact);
for (i = 0 ; i < cat->n_iap; ++i)
{
int ii;
struct interact_params *iap = &cat->iap[i];
- iap->base_subscript_short = idx_short;
- iap->base_subscript_long = idx_long;
+ iap->base_df = idx_df;
+ iap->base_cats = idx_cat;
- iap->reverse_interaction_value_map = pool_calloc (cat->pool, iap->n_cats,
- sizeof *iap->reverse_interaction_value_map);
+ iap->ivs = pool_calloc (cat->pool, iap->n_cats, sizeof *iap->ivs);
HMAP_FOR_EACH (ivn, struct interaction_value, node, &iap->ivmap)
{
- iap->reverse_interaction_value_map[x++] = ivn;
+ iap->ivs[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);
+ sort (iap->ivs, x, sizeof *iap->ivs,
+ 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;
+ iap->ivs[ii] = NULL;
- /* Populate the reverse variable maps. */
+ /* Populate the variable maps. */
if (iap->df_prod)
{
for (ii = 0; ii < iap->df_prod [iap->iact->n_vars - 1]; ++ii)
- cat->reverse_variable_map_short[idx_short++] = i;
+ cat->df_to_iact[idx_df++] = i;
}
for (ii = 0; ii < iap->n_cats; ++ii)
- cat->reverse_variable_map_long[idx_long++] = i;
+ cat->cat_to_iact[idx_cat++] = i;
}
assert (cat->n_vars <= cat->n_iap);
for (y = 0; y < hmap_count (&iap->ivmap); ++y)
{
- struct interaction_value *iv = iap->reverse_interaction_value_map[y];
- for (x = iap->base_subscript_short; x < iap->base_subscript_short + df ;++x)
+ struct interaction_value *iv = iap->ivs[y];
+ for (x = iap->base_df; x < iap->base_df + df ;++x)
{
const double bin = categoricals_get_effects_code_for_case (cat, x, iv->ccase);
- iap->enc_sum [x - iap->base_subscript_short] += bin * iv->cc;
+ iap->enc_sum [x - iap->base_df] += bin * iv->cc;
}
if (cat->payload && cat->payload->calculate)
cat->payload->calculate (cat->aux1, cat->aux2, iv->user_data);
static int
-reverse_variable_lookup_short (const struct categoricals *cat, int subscript)
+df_to_iap (const struct categoricals *cat, int subscript)
{
- assert (cat->reverse_variable_map_short);
+ assert (cat->df_to_iact);
assert (subscript >= 0);
assert (subscript < cat->df_sum);
- return cat->reverse_variable_map_short[subscript];
+ return cat->df_to_iact[subscript];
}
static int
-reverse_variable_lookup_long (const struct categoricals *cat, int subscript)
+cat_index_to_iap (const struct categoricals *cat, int subscript)
{
- assert (cat->reverse_variable_map_long);
+ assert (cat->cat_to_iact);
assert (subscript >= 0);
assert (subscript < cat->n_cats_total);
- return cat->reverse_variable_map_long[subscript];
+ return cat->cat_to_iact[subscript];
}
const struct interaction *
categoricals_get_interaction_by_subscript (const struct categoricals *cat, int subscript)
{
- int index = reverse_variable_lookup_short (cat, subscript);
+ int index = df_to_iap (cat, subscript);
return cat->iap[index].iact;
}
double
categoricals_get_weight_by_subscript (const struct categoricals *cat, int subscript)
{
- int vindex = reverse_variable_lookup_short (cat, subscript);
+ int vindex = df_to_iap (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);
+ int vindex = df_to_iap (cat, subscript);
const struct interact_params *vp = &cat->iap[vindex];
- return vp->enc_sum[subscript - vp->base_subscript_short];
+ return vp->enc_sum[subscript - vp->base_df];
}
{
const struct interaction *iact = categoricals_get_interaction_by_subscript (cat, subscript);
- const int i = reverse_variable_lookup_short (cat, subscript);
+ const int i = df_to_iap (cat, subscript);
- const int base_index = cat->iap[i].base_subscript_short;
+ const int base_index = cat->iap[i].base_df;
int v;
double result = 1.0;
const union value *val = case_data (c, var);
const int width = var_get_width (var);
- const struct variable_node *vn = lookup_variable (&cat->varmap, var, hash_pointer (var, 0));
+ const struct variable_node *vn = iap->varnodes[v];
const unsigned int hash = value_hash (val, width, 0);
const struct value_node *valn = lookup_value (&vn->valmap, val, hash, width);
const int index = ((subscript - base_index) % iap->df_prod[v] ) / dfp;
dfp = iap->df_prod [v];
- if (effects_coding && vn->indirection [valn->index] == df )
+ if (effects_coding && valn->index == df )
bin = -1.0;
- else if ( vn->indirection [valn->index] != index )
+ else if ( valn->index != index )
bin = 0;
-
+
result *= bin;
}
}
/* Returns unity if the value in case C at SUBSCRIPT is equal to the category
- for that subscript.
+ for that subscript.
Else if it is the last category, return -1.
Otherwise return 0.
*/
return categoricals_get_code_for_case (cat, subscript, c, true);
}
-
-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
+/* Return a case containing the set of values corresponding to
the Nth Category of the IACTth interaction */
const struct ccase *
categoricals_get_case_by_category_real (const struct categoricals *cat, int iact, int n)
if ( n >= hmap_count (&vp->ivmap))
return NULL;
- vn = vp->reverse_interaction_value_map [n];
+ vn = vp->ivs [n];
return vn->ccase;
}
if ( n >= hmap_count (&vp->ivmap))
return NULL;
- iv = vp->reverse_interaction_value_map [n];
+ iv = vp->ivs [n];
return iv->user_data;
}
const struct ccase *
categoricals_get_case_by_category (const struct categoricals *cat, int subscript)
{
- int vindex = reverse_variable_lookup_long (cat, subscript);
+ int vindex = cat_index_to_iap (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];
+ const struct interaction_value *vn = vp->ivs [subscript - vp->base_cats];
return vn->ccase;
}
void *
categoricals_get_user_data_by_category (const struct categoricals *cat, int subscript)
{
- int vindex = reverse_variable_lookup_long (cat, subscript);
+ int vindex = cat_index_to_iap (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];
+ const struct interaction_value *iv = vp->ivs [subscript - vp->base_cats];
return iv->user_data;
}
projects / pspp / blobdiff