Omnibox - Boost Frequency Scores Based on Number of Matching Pages

components/omnibox/browser/scored_history_match.cc

 // Copyright (c) 2012 The Chromium Authors. All rights reserved.
 // Use of this source code is governed by a BSD-style license that can be
 // found in the LICENSE file.
 
 #include "components/omnibox/browser/scored_history_match.h"
 
 #include <math.h>
 
 #include <algorithm>
 #include <vector>
@@ skipping 99 matching lines @@
 float ScoredHistoryMatch::typed_value_;
 bool ScoredHistoryMatch::fix_few_visits_bug_;
 bool ScoredHistoryMatch::frequency_uses_sum_;
 size_t ScoredHistoryMatch::max_visits_to_score_;
 bool ScoredHistoryMatch::allow_tld_matches_;
 bool ScoredHistoryMatch::allow_scheme_matches_;
 size_t ScoredHistoryMatch::num_title_words_to_allow_;
 float ScoredHistoryMatch::topicality_threshold_;
 ScoredHistoryMatch::ScoreMaxRelevances*
     ScoredHistoryMatch::relevance_buckets_override_ = nullptr;
+OmniboxFieldTrial::NumMatchesScores*
+    ScoredHistoryMatch::matches_to_specificity_override_ = nullptr;
 
 ScoredHistoryMatch::ScoredHistoryMatch()
     : ScoredHistoryMatch(history::URLRow(),
                          VisitInfoVector(),
                          base::string16(),
                          String16Vector(),
                          WordStarts(),
                          RowWordStarts(),
                          false,
-                         base::Time::Max()) {
-}
+                         1,
+                         base::Time::Max()) {}
 
 ScoredHistoryMatch::ScoredHistoryMatch(
     const history::URLRow& row,
     const VisitInfoVector& visits,
     const base::string16& lower_string,
     const String16Vector& terms_vector,
     const WordStarts& terms_to_word_starts_offsets,
     const RowWordStarts& word_starts,
     bool is_url_bookmarked,
+    size_t num_matching_pages,
     base::Time now)
     : HistoryMatch(row, 0, false, false), raw_score(0) {
   // NOTE: Call Init() before doing any validity checking to ensure that the
   // class is always initialized after an instance has been constructed. In
   // particular, this ensures that the class is initialized after an instance
   // has been constructed via the no-args constructor.
   ScoredHistoryMatch::Init();
 
   // Figure out where each search term appears in the URL and/or page title
   // so that we can score as well as provide autocomplete highlighting.
@@ skipping 97 matching lines @@
         likely_can_inline = true;
         innermost_match = (best_inlineable_prefix->num_components ==
                            best_prefix->num_components);
       }
     }
   }
 
   const float topicality_score = GetTopicalityScore(
       terms_vector.size(), url, terms_to_word_starts_offsets, word_starts);
   const float frequency_score = GetFrequency(now, is_url_bookmarked, visits);
-  raw_score = base::saturated_cast<int>(
-      GetFinalRelevancyScore(topicality_score, frequency_score));
+  const float specificity_score =
+      GetDocumentSpecificityScore(num_matching_pages);
+  raw_score = base::saturated_cast<int>(GetFinalRelevancyScore(
+      topicality_score, frequency_score, specificity_score));
 
   if (also_do_hup_like_scoring_ && likely_can_inline) {
     // HistoryURL-provider-like scoring gives any match that is
     // capable of being inlined a certain minimum score.  Some of these
     // are given a higher score that lets them be shown in inline.
     // This test here derives from the test in
     // HistoryURLProvider::PromoteMatchForInlineAutocomplete().
     const bool promote_to_inline =
         (row.typed_count() > 1) || (IsHostOnly() && (row.typed_count() == 1));
     int hup_like_score =
@@ skipping 334 matching lines @@
   // Compute the average weighted value_of_transition and return it.
   // Use |max_visits_to_score_| as the denominator for the average regardless of
   // how many visits there were in order to penalize a match that has
   // fewer visits than kMaxVisitsToScore.
   if (fix_few_visits_bug_)
     return summed_visit_points / ScoredHistoryMatch::max_visits_to_score_;
   return visits.size() * summed_visit_points /
          ScoredHistoryMatch::max_visits_to_score_;
 }
 
+float ScoredHistoryMatch::GetDocumentSpecificityScore(
+    size_t num_matching_pages) const {
+  // A mapping from the number of matching pages to their associated document
+  // specificity scores.  See omnibox_field_trial.h for more details.
+  CR_DEFINE_STATIC_LOCAL(OmniboxFieldTrial::NumMatchesScores,
+                         default_matches_to_specificity,
+                         (OmniboxFieldTrial::HQPNumMatchesScores()));
+  OmniboxFieldTrial::NumMatchesScores* matches_to_specificity =
+      matches_to_specificity_override_ ? matches_to_specificity_override_
+                                       : &default_matches_to_specificity;
+
+  // The floating point value below must be less than the lowest score the
+  // server would send down.
+  OmniboxFieldTrial::NumMatchesScores::const_iterator it = std::upper_bound(
+      matches_to_specificity->begin(), matches_to_specificity->end(),
+      std::pair<size_t, double>{num_matching_pages, -1});
+  return (it != matches_to_specificity->end()) ? it->second : 1.0;
+};
+
 // static
 float ScoredHistoryMatch::GetFinalRelevancyScore(float topicality_score,
-                                                 float frequency_score) {
+                                                 float frequency_score,
+                                                 float specificity_score) {
   // |relevance_buckets| gives a mapping from intemerdiate score to the final
   // relevance score.
   CR_DEFINE_STATIC_LOCAL(ScoreMaxRelevances, default_relevance_buckets,
                          (GetHQPBuckets()));
   ScoreMaxRelevances* relevance_buckets = relevance_buckets_override_
                                               ? relevance_buckets_override_
                                               : &default_relevance_buckets;
   DCHECK(!relevance_buckets->empty());
   DCHECK_EQ(0.0, (*relevance_buckets)[0].first);
 
   if (topicality_score == 0)
     return 0;
-  // Here's how to interpret intermediate_score: Suppose the omnibox
-  // has one input term.  Suppose we have a URL for which the omnibox
+  // Here's how to interpret intermediate_score: Suppose the omnibox has one
+  // input term.  Suppose the input matches many documents.  (This implies
+  // specificity_score == 1.0.)  Suppose we have a URL for which the omnibox
   // input term has a single URL hostname hit at a word boundary.  (This
   // implies topicality_score = 1.0.).  Then the intermediate_score for
   // this URL will depend entirely on the frequency_score with
   // this interpretation:
   // - a single typed visit more than three months ago, no other visits -> 0.2
   // - a visit every three days, no typed visits -> 0.706
   // - a visit every day, no typed visits -> 0.916
   // - a single typed visit yesterday, no other visits -> 2.0
   // - a typed visit once a week -> 11.77
   // - a typed visit every three days -> 14.12
   // - at least ten typed visits today -> 20.0 (maximum score)
   //
   // The below code maps intermediate_score to the range [0, 1399].
   // For example:
   // The default scoring buckets: "0.0:400,1.5:600,12.0:1300,20.0:1399"
   // We will linearly interpolate the scores between:
   //      0 to 1.5    --> 400 to 600
   //    1.5 to 12.0   --> 600 to 1300
   //    12.0 to 20.0  --> 1300 to 1399
   //       >= 20.0    --> 1399
   //
   // The score maxes out at 1399 (i.e., cannot beat a good inlineable result
   // from HistoryURL provider).
-  const float intermediate_score = topicality_score * frequency_score;
+  const float intermediate_score =
+      topicality_score * frequency_score * specificity_score;
 
   // Find the threshold where intermediate score is greater than bucket.
   size_t i = 1;
   for (; i < relevance_buckets->size(); ++i) {
     const ScoreMaxRelevance& bucket = (*relevance_buckets)[i];
     if (intermediate_score >= bucket.first) {
       continue;
     }
     const ScoreMaxRelevance& previous_bucket = (*relevance_buckets)[i - 1];
     const float slope = ((bucket.second - previous_bucket.second) /
@@ skipping 32 matching lines @@
     ScoreMaxRelevance bucket;
     bool is_valid_intermediate_score =
         base::StringToDouble(it->first, &bucket.first);
     DCHECK(is_valid_intermediate_score);
     bool is_valid_hqp_score = base::StringToInt(it->second, &bucket.second);
     DCHECK(is_valid_hqp_score);
     hqp_buckets.push_back(bucket);
   }
   return hqp_buckets;
 }