Add a bias term

In effect, a bias value allows you to shift the activation function
to the left or right, which may be critical for successful learning.

xapian-letor/api/featurelist.cc

@@ -99,7 +99,8 @@ FeatureList::normalise(std::vector<FeatureVector>& fvec) const
 std::vector<FeatureVector>
 FeatureList::create_feature_vectors(const Xapian::MSet & mset,
 				    const Xapian::Query & letor_query,
-				    const Xapian::Database & letor_db) const
+				    const Xapian::Database& letor_db,
+				    bool flag, double bias) const
 {
     LOGCALL(API, std::vector<FeatureVector>, "FeatureList::create_feature_vectors", mset | letor_query | letor_db);
     if (mset.empty())
@@ -121,6 +122,7 @@ FeatureList::create_feature_vectors(const Xapian::MSet & mset,
 	    // Append feature values
 	    fvals.insert(fvals.end(), values.begin(), values.end());
 	}
+	if (flag) fvals.push_back(bias);
 	double wt = i.get_weight();
 	// Weight is added as a feature by default.
 	fvals.push_back(wt);

xapian-letor/include/xapian-letor/featurelist.h

@@ -88,7 +88,8 @@ class XAPIAN_VISIBILITY_DEFAULT FeatureList {
     std::vector<Xapian::FeatureVector>
     create_feature_vectors(const Xapian::MSet & mset,
 			   const Xapian::Query & letor_query,
-			   const Xapian::Database & letor_db) const;
+			   const Xapian::Database& letor_db, bool flag = false,
+			   double bias = 1.0) const;
 
   private:
     /// Perform query-level normalisation of FeatureVectors.

xapian-letor/include/xapian-letor/ranker.h

@@ -88,7 +88,8 @@ prepare_training_file(const std::string & db_path,
 		      const std::string & qrel_file,
 		      Xapian::doccount msetsize,
 		      const std::string & filename,
-		      const Xapian::FeatureList & flist = FeatureList());
+		      const Xapian::FeatureList& flist = FeatureList(),
+		      bool flag = false, double bias = 1.0);
 
 class XAPIAN_VISIBILITY_DEFAULT Ranker : public Xapian::Internal::intrusive_base {
     /// Path to Xapian::Database instance to be used.

xapian-letor/ranker/ranker.cc

@@ -254,8 +254,11 @@ initialise_queryparser(const Xapian::Database & db)
 
 void
 Xapian::prepare_training_file(const string & db_path, const string & queryfile,
-		      const string & qrel_file, Xapian::doccount msetsize,
-		      const string & filename, const FeatureList & flist)
+			      const string& qrel_file,
+			      Xapian::doccount msetsize,
+			      const string& filename,
+			      const FeatureList& flist,
+			      bool flag, double bias)
 {
     // Set db
     Xapian::Database letor_db(db_path);
@@ -298,7 +301,9 @@ Xapian::prepare_training_file(const string & db_path, const string & queryfile,
 
 	vector<FeatureVector> fvv_mset = flist.create_feature_vectors(mset,
 								      query,
-								      letor_db);
+								      letor_db,
+								      flag,
+								      bias);
 	vector<FeatureVector> fvv_qrel;
 	int k = 0;
 	for (Xapian::MSetIterator i = mset.begin(); i != mset.end(); ++i) {

xapian-letor/tests/.gitignore

@@ -31,4 +31,5 @@
 /ndcg_score_output.txt
 /training_output.txt
 /training_output1.txt
+/training_output_data_bias.txt
 /training_output_three_correct.txt

xapian-letor/tests/Makefile.am

@@ -97,6 +97,7 @@ EXTRA_DIST +=\
 	testdata/querythree.txt \
 	testdata/score_qrel.txt \
 	testdata/training_data.txt \
+	testdata/training_data_bias.txt \
 	testdata/training_data_ndcg.txt \
 	testdata/training_data_one_document.txt \
 	testdata/training_data_three_correct.txt \
@@ -117,6 +118,7 @@ CLEANFILES +=\
 	ndcg_output_listnet_3.txt \
 	ndcg_score_output.txt \
 	ndcg_score_test.txt \
+	training_output_data_bias.txt \
 	training_output_data_one_doc.txt \
 	training_output_empty.txt \
 	training_output_three_correct.txt \

xapian-letor/tests/api_letor.cc

@@ -352,6 +352,59 @@ DEFINE_TESTCASE(preparetrainingfileonedb, generated && path && writable)
     unlink("training_output_data_one_doc.txt");
 }
 
+DEFINE_TESTCASE(preparetrainingfile_with_bias, generated && path && writable)
+{
+    vector<Xapian::Feature*> flist;
+    flist.push_back(new Xapian::TfFeature());
+    flist.push_back(new Xapian::TfDoclenFeature());
+    flist.push_back(new Xapian::IdfFeature());
+    flist.push_back(new Xapian::CollTfCollLenFeature());
+    flist.push_back(new Xapian::TfIdfDoclenFeature());
+    flist.push_back(new Xapian::TfDoclenCollTfCollLenFeature());
+    string db_path = get_database_path("apitest_listnet_ranker1",
+				       db_index_one_document);
+    string data_directory = test_driver::get_srcdir() + "/testdata/";
+    string query = data_directory + "queryone.txt";
+    string qrel = data_directory + "qrelone.txt";
+    string training_data = data_directory + "training_data_bias.txt";
+    unlink("training_output_data_bias.txt");
+    Xapian::prepare_training_file(db_path, query, qrel, 10,
+				  "training_output_data_bias.txt", flist, true);
+    TEST(file_exists("training_output_data_bias.txt"));
+    ifstream if1(training_data);
+    ifstream if2("training_output_data_bias.txt");
+    string line1;
+    string line2;
+    while (getline(if1, line1)) {
+	TEST(getline(if2, line2));
+	istringstream iss1(line1);
+	istringstream iss2(line2);
+	string temp1;
+	string temp2;
+	int i = 0;
+	while ((iss1 >> temp1) && (iss2 >> temp2)) {
+	    // The 0th, 1st and 22nd literals taken as input, are strings,
+	    // and can be compared directly, They are: For example(test):
+	    // ("1", "qid:20001" and "#docid=1") at 0th, 1st, and 22nd pos
+	    // respectively. Whereas the other values are doubles which
+	    // would have to tested under TEST_DOUBLE() against precision.
+	    if (i == 0 || i == 1 || i == 22) {
+		TEST_EQUAL(temp1, temp2);
+	    } else {
+		size_t t1 = temp1.find_first_of(':');
+		size_t t2 = temp2.find_first_of(':');
+		TEST_EQUAL_DOUBLE(stod(temp1.substr(t1 + 1)),
+				  stod(temp2.substr(t2 + 1)));
+	    }
+	    i++;
+	}
+	TEST_REL(i, ==, 23);
+	TEST(!(iss2 >> temp2));
+    }
+    TEST(!getline(if2, line2));
+    unlink("training_output_data_bias.txt");
+}
+
 #define TEST_PARSE_EXCEPTION(TESTFILE) TEST_EXCEPTION(Xapian::LetorParseError,\
 	Xapian::prepare_training_file(db_path,\
 				      data_directory + TESTFILE, qrel, 10,\

xapian-letor/tests/testdata/training_data_bias.txt

@@ -0,0 +1 @@
+4 qid:110001 1:1 2:1 3:1 4:1 5:1 6:1 7:-0.155541 8:-0.311083 9:-0.466624 10:1 11:1 12:1 13:-0.0269673 14:-0.00573887 15:-0.0063153 16:1 17:1 18:1 19:1 20:1 #docid=1