#include <config.h>
-#include "roc.h"
#include <data/procedure.h>
#include <language/lexer/variable-parser.h>
#include <language/lexer/value-parser.h>
+#include <language/command.h>
#include <language/lexer/lexer.h>
#include <data/casegrouper.h>
{
size_t n_vars;
const struct variable **vars;
+ const struct dictionary *dict;
- struct variable *state_var ;
+ const struct variable *state_var ;
union value state_value;
/* Plot the roc curve */
bool invert ; /* True iff a smaller test result variable indicates
a positive result */
+
+ double pos;
+ double neg;
+ double pos_weighted;
+ double neg_weighted;
};
static int run_roc (struct dataset *ds, struct cmd_roc *roc);
roc.ci = 95;
roc.bi_neg_exp = false;
roc.invert = false;
+ roc.pos = roc.pos_weighted = 0;
+ roc.neg = roc.neg_weighted = 0;
+ roc.dict = dataset_dict (ds);
if (!parse_variables_const (lexer, dict, &roc.vars, &roc.n_vars,
PV_APPEND | PV_NO_DUPLICATE | PV_NUMERIC))
- return 2;
+ goto error;;
if ( ! lex_force_match (lexer, T_BY))
{
- return 2;
+ goto error;;
}
roc.state_var = parse_variable (lexer, dict);
if ( !lex_force_match (lexer, '('))
{
- return 2;
+ goto error;;
}
parse_value (lexer, &roc.state_value, var_get_width (roc.state_var));
if ( !lex_force_match (lexer, ')'))
{
- return 2;
+ goto error;;
}
else
{
lex_error (lexer, NULL);
- return 2;
+ goto error;;
}
}
}
else
{
lex_error (lexer, NULL);
- return 2;
+ goto error;;
}
}
else if (lex_match_id (lexer, "PRINT"))
else
{
lex_error (lexer, NULL);
- return 2;
+ goto error;;
}
}
}
else
{
lex_error (lexer, NULL);
- return 2;
+ goto error;;
}
lex_force_match (lexer, ')');
}
else
{
lex_error (lexer, NULL);
- return 2;
+ goto error;;
}
lex_force_match (lexer, ')');
}
else
{
lex_error (lexer, NULL);
- return 2;
+ goto error;;
}
lex_force_match (lexer, ')');
}
else
{
lex_error (lexer, NULL);
- return 2;
+ goto error;;
}
}
}
}
}
- run_roc (ds, &roc);
+ if ( ! run_roc (ds, &roc))
+ goto error;;
+
+ return CMD_SUCCESS;
- return 1;
+ error:
+ free (roc.vars);
+ return CMD_FAILURE;
}
return ok;
}
-
+#if 0
static void
dump_casereader (struct casereader *reader)
{
casereader_destroy (r);
}
+#endif
+
+/*
+ Return true iff the state variable indicates that C has positive actual state.
+
+ As a side effect, this function also accumulates the roc->{pos,neg} and
+ roc->{pos,neg}_weighted counts.
+ */
static bool
match_positives (const struct ccase *c, void *aux)
{
struct cmd_roc *roc = aux;
+ const struct variable *wv = dict_get_weight (roc->dict);
+ const double weight = wv ? case_data (c, wv)->f : 1.0;
+
+ const bool positive =
+ ( 0 == value_compare_3way (case_data (c, roc->state_var), &roc->state_value,
+ var_get_width (roc->state_var)));
+
+ if ( positive )
+ {
+ roc->pos++;
+ roc->pos_weighted += weight;
+ }
+ else
+ {
+ roc->neg++;
+ roc->neg_weighted += weight;
+ }
- return 0 == value_compare_3way (case_data (c, roc->state_var),
- &roc->state_value,
- var_get_width (roc->state_var));
+ return positive;
}
#define N_EQ 1
#define N_PRED 2
+/* Some intermediate state for calculating the cutpoints and the
+ standard error values */
struct roc_state
{
double auc;
double max;
};
-
-
#define CUTPOINT 0
#define TP 1
#define FN 2
#define FP 4
+/*
+ Return a new casereader based upon CUTPOINT_RDR.
+ The number of "positive" cases are placed into
+ the position TRUE_INDEX, and the number of "negative" cases
+ into FALSE_INDEX.
+ POS_COND and RESULT determine the semantics of what is
+ "positive".
+ WEIGHT is the value of a single count.
+ */
static struct casereader *
accumulate_counts (struct casereader *cutpoint_rdr,
double result, double weight,
struct ccase *cpc;
double prev_cp = SYSMIS;
-
for ( ; (cpc = casereader_read (r) ); case_unref (cpc))
{
struct ccase *new_case;
const double cp = case_data_idx (cpc, CUTPOINT)->f;
+ assert (cp != SYSMIS);
+
/* We don't want duplicates here */
if ( cp == prev_cp )
continue;
new_case = case_clone (cpc);
if ( pos_cond (result, cp))
- {
- case_data_rw_idx (new_case, true_index)->f += weight;
- }
+ case_data_rw_idx (new_case, true_index)->f += weight;
else
- {
- case_data_rw_idx (new_case, false_index)->f += weight;
- }
+ case_data_rw_idx (new_case, false_index)->f += weight;
prev_cp = cp;
struct casereader **cutpoint_rdr,
bool (*pos_cond) (double, double),
int true_index,
- int false_index
- )
+ int false_index)
{
const struct variable *w = dict_get_weight (dict);
+
struct casereader *r1 =
casereader_create_distinct (sort_execute_1var (reader, var), var, w);
for ( ; (c1 = casereader_read (r1) ); case_unref (c1))
{
+ struct ccase *new_case = case_create (proto);
struct ccase *c2;
struct casereader *r2 = casereader_clone (rclone);
double n_pred = 0.0;
*cutpoint_rdr = accumulate_counts (*cutpoint_rdr, d1, weight1,
- pos_cond,
- true_index, false_index);
-
- struct ccase *new_case = case_create (proto);
+ pos_cond,
+ true_index, false_index);
*cc += weight1;
return casewriter_make_reader (wtr);
}
+/* Some more indeces into case data */
+#define N_POS_EQ 1 /* number of positive cases with values equal to n */
+#define N_POS_GT 2 /* number of postive cases with values greater than n */
+#define N_NEG_EQ 3 /* number of negative cases with values equal to n */
+#define N_NEG_LT 4 /* number of negative cases with values less than n */
+
static bool
gt (double d1, double d2)
{
return d1 < d2;
}
+
+/*
+ Return a casereader with width 3,
+ populated with cases based upon READER.
+ The cases will have the values:
+ (N, number of cases equal to N, number of cases greater than N)
+ As a side effect, update RS->n1 with the number of positive cases.
+*/
static struct casereader *
process_positive_group (const struct variable *var, struct casereader *reader,
const struct dictionary *dict,
TP, FN);
}
-
+/*
+ Return a casereader with width 3,
+ populated with cases based upon READER.
+ The cases will have the values:
+ (N, number of cases equal to N, number of cases less than N)
+ As a side effect, update RS->n2 with the number of negative cases.
+*/
static struct casereader *
process_negative_group (const struct variable *var, struct casereader *reader,
const struct dictionary *dict,
case_data_rw_idx (cc, TN)->f = 0;
case_data_rw_idx (cc, FP)->f = 0;
-
casewriter_write (writer, cc);
}
+/*
+ Create and initialise the rs[x].cutpoint_rdr casereaders. That is, the readers will
+ be created with width 5, ready to take the values (cutpoint, TP, FN, TN, FP), and the
+ reader will be populated with its final number of cases.
+ However on exit from this function, only CUTPOINT entries will be set to their final
+ value. The other entries will be initialised to zero.
+*/
static void
-do_roc (struct cmd_roc *roc, struct casereader *input, struct dictionary *dict)
+prepare_cutpoints (struct cmd_roc *roc, struct roc_state *rs, struct casereader *input)
{
int i;
+ struct casereader *r = casereader_clone (input);
+ struct ccase *c;
+ struct caseproto *proto = caseproto_create ();
- struct roc_state *rs = xcalloc (roc->n_vars, sizeof *rs);
+ struct subcase ordering;
+ subcase_init (&ordering, CUTPOINT, 0, SC_ASCEND);
- struct casewriter *neg_wtr = autopaging_writer_create (casereader_get_proto (input));
+ proto = caseproto_add_width (proto, 0); /* cutpoint */
+ proto = caseproto_add_width (proto, 0); /* TP */
+ proto = caseproto_add_width (proto, 0); /* FN */
+ proto = caseproto_add_width (proto, 0); /* TN */
+ proto = caseproto_add_width (proto, 0); /* FP */
- struct casereader *negatives = NULL;
- struct casereader *positives = NULL;
+ for (i = 0 ; i < roc->n_vars; ++i)
+ {
+ rs[i].cutpoint_wtr = sort_create_writer (&ordering, proto);
+ rs[i].prev_result = SYSMIS;
+ rs[i].max = -DBL_MAX;
+ rs[i].min = DBL_MAX;
+ }
+ for (; (c = casereader_read (r)) != NULL; case_unref (c))
+ {
+ for (i = 0 ; i < roc->n_vars; ++i)
+ {
+ const union value *v = case_data (c, roc->vars[i]);
+ const double result = v->f;
- /* Prepare the cutpoints */
- {
- struct casereader *r = casereader_clone (input);
- struct ccase *c;
- struct caseproto *proto = caseproto_create ();
+ if ( mv_is_value_missing (var_get_missing_values (roc->vars[i]), v, roc->exclude))
+ continue;
- struct subcase ordering;
- subcase_init (&ordering, CUTPOINT, 0, SC_ASCEND);
+ minimize (&rs[i].min, result);
+ maximize (&rs[i].max, result);
+ if ( rs[i].prev_result != SYSMIS && rs[i].prev_result != result )
+ {
+ const double mean = (result + rs[i].prev_result ) / 2.0;
+ append_cutpoint (rs[i].cutpoint_wtr, mean);
+ }
- proto = caseproto_add_width (proto, 0); /* cutpoint */
- proto = caseproto_add_width (proto, 0); /* TP */
- proto = caseproto_add_width (proto, 0); /* FN */
- proto = caseproto_add_width (proto, 0); /* TN */
- proto = caseproto_add_width (proto, 0); /* FP */
+ rs[i].prev_result = result;
+ }
+ }
+ casereader_destroy (r);
- for (i = 0 ; i < roc->n_vars; ++i)
- {
- rs[i].cutpoint_wtr = sort_create_writer (&ordering, proto);
- rs[i].prev_result = SYSMIS;
- rs[i].max = -DBL_MAX;
- rs[i].min = DBL_MAX;
- }
+ /* Append the min and max cutpoints */
+ for (i = 0 ; i < roc->n_vars; ++i)
+ {
+ append_cutpoint (rs[i].cutpoint_wtr, rs[i].min - 1);
+ append_cutpoint (rs[i].cutpoint_wtr, rs[i].max + 1);
- for (; (c = casereader_read (r)) != NULL; case_unref (c))
- {
- const double weight = dict_get_case_weight (dict, c, NULL);
- for (i = 0 ; i < roc->n_vars; ++i)
- {
- const double result = case_data (c, roc->vars[i])->f;
+ rs[i].cutpoint_rdr = casewriter_make_reader (rs[i].cutpoint_wtr);
+ }
+}
- minimize (&rs[i].min, result);
- maximize (&rs[i].max, result);
+static void
+do_roc (struct cmd_roc *roc, struct casereader *reader, struct dictionary *dict)
+{
+ int i;
- if ( rs[i].prev_result != SYSMIS && rs[i].prev_result != result )
- {
- const double mean = (result + rs[i].prev_result ) / 2.0;
- append_cutpoint (rs[i].cutpoint_wtr, mean);
- }
+ struct roc_state *rs = xcalloc (roc->n_vars, sizeof *rs);
- rs[i].prev_result = result;
- }
- }
- casereader_destroy (r);
+ struct casereader *negatives = NULL;
+ struct casereader *positives = NULL;
+ struct caseproto *n_proto = caseproto_create ();
- /* Append the min and max cutpoints */
- for (i = 0 ; i < roc->n_vars; ++i)
- {
- append_cutpoint (rs[i].cutpoint_wtr, rs[i].min - 1);
- append_cutpoint (rs[i].cutpoint_wtr, rs[i].max + 1);
+ struct subcase up_ordering;
+ struct subcase down_ordering;
- rs[i].cutpoint_rdr = casewriter_make_reader (rs[i].cutpoint_wtr);
- }
- }
+ struct casewriter *neg_wtr = NULL;
+
+ struct casereader *input = casereader_create_filter_missing (reader,
+ roc->vars, roc->n_vars,
+ roc->exclude,
+ NULL,
+ NULL);
+
+ input = casereader_create_filter_missing (input,
+ &roc->state_var, 1,
+ roc->exclude,
+ NULL,
+ NULL);
+
+ neg_wtr = autopaging_writer_create (casereader_get_proto (input));
+
+ prepare_cutpoints (roc, rs, input);
- positives =
+
+ /* Separate the positive actual state cases from the negative ones */
+ positives =
casereader_create_filter_func (input,
match_positives,
NULL,
roc,
neg_wtr);
+ n_proto = caseproto_create ();
+
+ n_proto = caseproto_add_width (n_proto, 0);
+ n_proto = caseproto_add_width (n_proto, 0);
+ n_proto = caseproto_add_width (n_proto, 0);
+ n_proto = caseproto_add_width (n_proto, 0);
+ n_proto = caseproto_add_width (n_proto, 0);
+
+ subcase_init (&up_ordering, VALUE, 0, SC_ASCEND);
+ subcase_init (&down_ordering, VALUE, 0, SC_DESCEND);
for (i = 0 ; i < roc->n_vars; ++i)
{
+ struct casewriter *w = NULL;
+ struct casereader *r = NULL;
+
+ struct ccase *c;
+
struct ccase *cpos;
struct casereader *n_neg ;
const struct variable *var = roc->vars[i];
struct casereader *neg ;
struct casereader *pos = casereader_clone (positives);
+
struct casereader *n_pos =
process_positive_group (var, pos, dict, &rs[i]);
n_neg = process_negative_group (var, neg, dict, &rs[i]);
- printf ("Positives:\n");
- dump_casereader (n_pos);
-
- printf ("Negatives:\n");
- dump_casereader (n_neg);
-
-#if 0
- /* Simple join on VALUE */
+ /* Merge the n_pos and n_neg casereaders */
+ w = sort_create_writer (&up_ordering, n_proto);
for ( ; (cpos = casereader_read (n_pos) ); case_unref (cpos))
{
- struct ccase *cneg = NULL;
- double dneg = -DBL_MAX;
- const double dpos = case_data_idx (cpos, VALUE)->f;
- while (dneg < dpos)
+ struct ccase *pos_case = case_create (n_proto);
+ struct ccase *cneg;
+ const double jpos = case_data_idx (cpos, VALUE)->f;
+
+ while ((cneg = casereader_read (n_neg)))
{
- if ( cneg )
- case_unref (cneg);
+ struct ccase *nc = case_create (n_proto);
- cneg = casereader_read (n_neg);
- if ( ! cneg )
+ const double jneg = case_data_idx (cneg, VALUE)->f;
+
+ case_data_rw_idx (nc, VALUE)->f = jneg;
+ case_data_rw_idx (nc, N_POS_EQ)->f = 0;
+
+ case_data_rw_idx (nc, N_POS_GT)->f = SYSMIS;
+
+ *case_data_rw_idx (nc, N_NEG_EQ) = *case_data_idx (cneg, N_EQ);
+ *case_data_rw_idx (nc, N_NEG_LT) = *case_data_idx (cneg, N_PRED);
+
+ casewriter_write (w, nc);
+
+ case_unref (cneg);
+ if ( jneg > jpos)
break;
- dneg = case_data_idx (cneg, VALUE)->f;
- }
-
- if ( dpos == dneg )
- {
- double n_pos_eq = case_data_idx (cpos, N_EQ)->f;
- double n_neg_eq = case_data_idx (cneg, N_EQ)->f;
- double n_pos_gt = case_data_idx (cpos, N_PRED)->f;
- double n_neg_lt = case_data_idx (cneg, N_PRED)->f;
-
- rs[i].auc += n_pos_gt * n_neg_eq + (n_pos_eq * n_neg_eq) / 2.0;
- rs[i].q1hat +=
- n_neg_eq * ( pow2 (n_pos_gt) + n_pos_gt * n_pos_eq + pow2 (n_pos_eq) / 3.0);
- rs[i].q2hat +=
- n_pos_eq * ( pow2 (n_neg_lt) + n_neg_lt * n_neg_eq + pow2 (n_neg_eq) / 3.0);
}
- if ( cneg )
- case_unref (cneg);
+ case_data_rw_idx (pos_case, VALUE)->f = jpos;
+ *case_data_rw_idx (pos_case, N_POS_EQ) = *case_data_idx (cpos, N_EQ);
+ *case_data_rw_idx (pos_case, N_POS_GT) = *case_data_idx (cpos, N_PRED);
+ case_data_rw_idx (pos_case, N_NEG_EQ)->f = 0;
+ case_data_rw_idx (pos_case, N_NEG_LT)->f = SYSMIS;
+
+ casewriter_write (w, pos_case);
}
- rs[i].auc /= rs[i].n1 * rs[i].n2;
- if ( roc->invert )
- rs[i].auc = 1 - rs[i].auc;
+/* These aren't used anymore */
+#undef N_EQ
+#undef N_PRED
- if ( roc->bi_neg_exp )
- {
- rs[i].q1hat = rs[i].auc / ( 2 - rs[i].auc);
- rs[i].q2hat = 2 * pow2 (rs[i].auc) / ( 1 + rs[i].auc);
- }
- else
- {
- rs[i].q1hat /= rs[i].n2 * pow2 (rs[i].n1);
- rs[i].q2hat /= rs[i].n1 * pow2 (rs[i].n2);
- }
+ r = casewriter_make_reader (w);
+
+ /* Propagate the N_POS_GT values from the positive cases
+ to the negative ones */
+ {
+ double prev_pos_gt = rs[i].n1;
+ w = sort_create_writer (&down_ordering, n_proto);
+
+ for ( ; (c = casereader_read (r) ); case_unref (c))
+ {
+ double n_pos_gt = case_data_idx (c, N_POS_GT)->f;
+ struct ccase *nc = case_clone (c);
+
+ if ( n_pos_gt == SYSMIS)
+ {
+ n_pos_gt = prev_pos_gt;
+ case_data_rw_idx (nc, N_POS_GT)->f = n_pos_gt;
+ }
+
+ casewriter_write (w, nc);
+ prev_pos_gt = n_pos_gt;
+ }
+
+ r = casewriter_make_reader (w);
+ }
+
+ /* Propagate the N_NEG_LT values from the negative cases
+ to the positive ones */
+ {
+ double prev_neg_lt = rs[i].n2;
+ w = sort_create_writer (&up_ordering, n_proto);
+
+ for ( ; (c = casereader_read (r) ); case_unref (c))
+ {
+ double n_neg_lt = case_data_idx (c, N_NEG_LT)->f;
+ struct ccase *nc = case_clone (c);
+
+ if ( n_neg_lt == SYSMIS)
+ {
+ n_neg_lt = prev_neg_lt;
+ case_data_rw_idx (nc, N_NEG_LT)->f = n_neg_lt;
+ }
+
+ casewriter_write (w, nc);
+ prev_neg_lt = n_neg_lt;
+ }
+
+ r = casewriter_make_reader (w);
+ }
+
+ {
+ struct ccase *prev_case = NULL;
+ for ( ; (c = casereader_read (r) ); case_unref (c))
+ {
+ const struct ccase *next_case = casereader_peek (r, 0);
+
+ const double j = case_data_idx (c, VALUE)->f;
+ double n_pos_eq = case_data_idx (c, N_POS_EQ)->f;
+ double n_pos_gt = case_data_idx (c, N_POS_GT)->f;
+ double n_neg_eq = case_data_idx (c, N_NEG_EQ)->f;
+ double n_neg_lt = case_data_idx (c, N_NEG_LT)->f;
+
+ if ( prev_case && j == case_data_idx (prev_case, VALUE)->f)
+ {
+ if ( 0 == case_data_idx (c, N_POS_EQ)->f)
+ {
+ n_pos_eq = case_data_idx (prev_case, N_POS_EQ)->f;
+ n_pos_gt = case_data_idx (prev_case, N_POS_GT)->f;
+ }
+
+ if ( 0 == case_data_idx (c, N_NEG_EQ)->f)
+ {
+ n_neg_eq = case_data_idx (prev_case, N_NEG_EQ)->f;
+ n_neg_lt = case_data_idx (prev_case, N_NEG_LT)->f;
+ }
+ }
+
+ if ( NULL == next_case || j != case_data_idx (next_case, VALUE)->f)
+ {
+ rs[i].auc += n_pos_gt * n_neg_eq + (n_pos_eq * n_neg_eq) / 2.0;
+
+ rs[i].q1hat +=
+ n_neg_eq * ( pow2 (n_pos_gt) + n_pos_gt * n_pos_eq + pow2 (n_pos_eq) / 3.0);
+ rs[i].q2hat +=
+ n_pos_eq * ( pow2 (n_neg_lt) + n_neg_lt * n_neg_eq + pow2 (n_neg_eq) / 3.0);
+
+ }
+
+ case_unref (prev_case);
+ prev_case = case_clone (c);
+ }
+
+ rs[i].auc /= rs[i].n1 * rs[i].n2;
+ if ( roc->invert )
+ rs[i].auc = 1 - rs[i].auc;
+
+ if ( roc->bi_neg_exp )
+ {
+ rs[i].q1hat = rs[i].auc / ( 2 - rs[i].auc);
+ rs[i].q2hat = 2 * pow2 (rs[i].auc) / ( 1 + rs[i].auc);
+ }
+ else
+ {
+ rs[i].q1hat /= rs[i].n2 * pow2 (rs[i].n1);
+ rs[i].q2hat /= rs[i].n1 * pow2 (rs[i].n2);
+ }
+ }
}
-#endif
casereader_destroy (positives);
casereader_destroy (negatives);
free (rs);
}
-
-
-
static void
show_auc (struct roc_state *rs, const struct cmd_roc *roc)
{
tab_text (tbl, 0, 3, TAB_LEFT, _("Negative"));
-#if 0
tab_double (tbl, 1, 2, 0, roc->pos, &F_8_0);
tab_double (tbl, 1, 3, 0, roc->neg, &F_8_0);
tab_double (tbl, 2, 2, 0, roc->pos_weighted, 0);
tab_double (tbl, 2, 3, 0, roc->neg_weighted, 0);
-#endif
tab_submit (tbl);
}
projects / pspp-builds.git / blobdiff