/* This file is part of Pazpar2.
- Copyright (C) 2006-2013 Index Data
+ Copyright (C) Index Data
Pazpar2 is free software; you can redistribute it and/or modify it under
the terms of the GNU General Public License as published by the Free
// This is the old behavior, and the default
const int scorefield_internal = -2; // use our tf/idf, but normalize it
const int scorefield_position = -3; // fake a score based on the position
+// Positive numbers indicate the field to be used for scoring.
// A structure for each (sub)record. There is one list for each client
struct norm_record
}
-// Add a record in the list for that client, for normalizing later
+// Add all records from a cluster into the list for that client, for normalizing later
static void setup_norm_record( struct relevance *rel, struct record_cluster *clust)
{
struct record *record;
{
struct record_metadata *md = record->metadata[norm->scorefield];
rp->score = md->data.fnumber;
- assert(rp->score>0); // ###
}
yaz_log(YLOG_LOG,"Got score for %d/%d : %f ",
norm->num, record->position, rp->score );
+ record -> score = rp->score;
if ( norm->count == 1 )
{
norm->max = rp->score;
} else {
if ( rp->score > norm->max )
norm->max = rp->score;
- if ( rp->score < norm->min && abs(rp->score) < 1e-6 )
- norm->min = rp->score; // skip zeroes
+ if ( rp->score < norm->min )
+ norm->min = rp->score;
}
}
}
return sum;
}
+// For each client, normalize scores
static void normalize_scores(struct relevance *rel)
{
- // For each client, normalize scores
+ const int maxiterations = 1000;
+ const double enough = 100.0; // sets the number of decimals we are happy with
+ const double stepchange = 0.5; // reduction of the step size when finding middle
+ // 0.5 sems to be magical, much better than 0.4 or 0.6
struct norm_client *norm;
for ( norm = rel->norm; norm; norm = norm->next )
{
- yaz_log(YLOG_LOG,"Normalizing client %d: scorefield=%d count=%d",
- norm->num, norm->scorefield, norm->count);
+ yaz_log(YLOG_LOG,"Normalizing client %d: scorefield=%d count=%d range=%f %f = %f",
+ norm->num, norm->scorefield, norm->count, norm->min,
+ norm->max, norm->max-norm->min);
norm->a = 1.0; // default normalizing factors, no change
norm->b = 0.0;
if ( norm->scorefield != scorefield_none &&
double a,b; // params to optimize
double as,bs; // step sizes
double chi;
+ char *branch = "?";
// initial guesses for the parameters
+ // Rmax = a * rmax + b # want to be 1.0
+ // Rmin = a * rmin + b # want to be 0.0
+ // Rmax - Rmin = a ( rmax - rmin ) # subtracting equations
+ // 1.0 - 0.0 = a ( rmax - rmin )
+ // a = 1 / range
+ // Rmin = a * rmin + b
+ // b = Rmin - a * rmin
+ // = 0.0 - 1/range * rmin
+ // = - rmin / range
+
if ( range < 1e-6 ) // practically zero
range = norm->max;
a = 1.0 / range;
- b = abs(norm->min);
- as = a / 3;
- bs = b / 3;
+ b = -1.0 * norm->min / range;
+ // b = fabs(norm->min) / range;
+ as = a / 10;
+ bs = fabs(b) / 10;
chi = squaresum( norm->records, a,b);
- while (it++ < 100) // safeguard against things not converging
+ yaz_log(YLOG_LOG,"Initial done: it=%d: a=%f / %f b=%f / %f chi = %f",
+ 0, a, as, b, bs, chi );
+ while (it++ < maxiterations) // safeguard against things not converging
{
- // optimize a
- double plus = squaresum(norm->records, a+as, b);
- double minus= squaresum(norm->records, a-as, b);
- if ( plus < chi && plus < minus )
+ double aplus = squaresum(norm->records, a+as, b);
+ double aminus= squaresum(norm->records, a-as, b);
+ double bplus = squaresum(norm->records, a, b+bs);
+ double bminus= squaresum(norm->records, a, b-bs);
+ double prevchi = chi;
+ if ( aplus < chi && aplus < aminus && aplus < bplus && aplus < bminus)
{
a = a + as;
- chi = plus;
+ chi = aplus;
+ as = as * (1.0 + stepchange);
+ branch = "aplus ";
}
- else if ( minus < chi && minus < plus )
+ else if ( aminus < chi && aminus < aplus && aminus < bplus && aminus < bminus)
{
a = a - as;
- chi = minus;
+ chi = aminus;
+ as = as * (1.0 + stepchange);
+ branch = "aminus";
}
- else
- as = as / 2;
- // optimize b
- plus = squaresum(norm->records, a, b+bs);
- minus= squaresum(norm->records, a, b-bs);
- if ( plus < chi && plus < minus )
+ else if ( bplus < chi && bplus < aplus && bplus < aminus && bplus < bminus)
{
b = b + bs;
- chi = plus;
+ chi = bplus;
+ bs = bs * (1.0 + stepchange);
+ branch = "bplus ";
}
- else if ( minus < chi && minus < plus )
+ else if ( bminus < chi && bminus < aplus && bminus < bplus && bminus < aminus)
{
b = b - bs;
- chi = minus;
+ chi = bminus;
+ branch = "bminus";
+ bs = bs * (1.0+stepchange);
}
else
- bs = bs / 2;
- yaz_log(YLOG_LOG,"Fitting it=%d: a=%f / %f b=%f / %f chi = %f",
- it, a, as, b, bs, chi );
+ { // a,b is the best so far, adjust one step size
+ // which one? The one that has the greatest effect to chi
+ // That is, the average of plus and minus is further away from chi
+ double adif = 0.5 * ( aplus + aminus ) - prevchi;
+ double bdif = 0.5 * ( bplus + bminus ) - prevchi;
+ if ( fabs(adif) > fabs(bdif) )
+ {
+ as = as * ( 1.0 - stepchange);
+ branch = "step a";
+ }
+ else
+ {
+ bs = bs * ( 1.0 - stepchange);
+ branch = "step b";
+ }
+ }
+ yaz_log(YLOG_LOG,"Fitting %s it=%d: a=%g %g b=%g %g chi=%g ap=%g am=%g, bp=%g bm=%g p=%g",
+ branch, it, a, as, b, bs, chi,
+ aplus, aminus, bplus, bminus, prevchi );
norm->a = a;
norm->b = b;
- if ( abs(as) * 1000.0 < abs(a) &&
- abs(bs) * 1000.0 < abs(b) )
+ if ( fabs(as) * enough < fabs(a) &&
+ fabs(bs) * enough < fabs(b) ) {
break; // not changing much any more
+
+ }
}
+ yaz_log(YLOG_LOG,"Fitting done: it=%d: a=%g / %g b=%g / %g chi = %g",
+ it-1, a, as, b, bs, chi );
}
-
+
if ( norm->scorefield != scorefield_none )
{ // distribute the normalized scores to the records
struct norm_record *nr = norm->records;
double r = nr->score;
r = norm->a * r + norm -> b;
nr->clust->relevance_score = 10000 * r;
+ nr->record->score = r;
yaz_log(YLOG_LOG,"Normalized %f * %f + %f = %f",
nr->score, norm->a, norm->b, r );
// TODO - This keeps overwriting the cluster score in random order!
- // Need to merge results better
+ // Need to merge results better
}
-
}
-
} // client loop
}
rel->doc_frequency_vec[i]);
}
}
- // Calculate relevance for each document
+ // Calculate relevance for each document (cluster)
while (1)
{
int relevance = 0;
// Build the normalizing structures
// List of (sub)records for each target
setup_norm_record( rel, rec );
-
- // TODO - Loop again, merge individual record scores into clusters
- // Can I reset the reclist, or can I leave and enter without race conditions?
-
+
} // cluster loop
normalize_scores(rel);
-
+
+ // TODO - Calculate the cluster scores from individual records
+ // At the moment the record scoring puts one of them in the cluster...
+ reclist_rewind(reclist);
+
reclist_leave(reclist);
xfree(idfvec);