[search] Prefer objects with higher popularity

5 files changed, 14 insertions, 5 deletions

diff --git a/search/intermediate_result.cpp b/search/intermediate_result.cpp
index b0e62b6d3b..0c0d9f1326 100644
--- a/search/intermediate_result.cpp
+++ b/search/intermediate_result.cpp
@@ -69,10 +69,12 @@ PreRankerResult::PreRankerResult(FeatureID const & id, PreRankingInfo const & in
 }
 
 // static
-bool PreRankerResult::LessRank(PreRankerResult const & r1, PreRankerResult const & r2)
+bool PreRankerResult::LessRankAndPopularity(PreRankerResult const & r1, PreRankerResult const & r2)
 {
   if (r1.m_info.m_rank != r2.m_info.m_rank)
     return r1.m_info.m_rank > r2.m_info.m_rank;
+  if (r1.m_info.m_popularity != r2.m_info.m_popularity)
+    return r1.m_info.m_popularity > r2.m_info.m_popularity;
   return r1.m_info.m_distanceToPivot < r2.m_info.m_distanceToPivot;
 }
 
diff --git a/search/intermediate_result.hpp b/search/intermediate_result.hpp
index a2cd42ff20..e509b1ba0f 100644
--- a/search/intermediate_result.hpp
+++ b/search/intermediate_result.hpp
@@ -31,7 +31,7 @@ class PreRankerResult
 public:
   PreRankerResult(FeatureID const & id, PreRankingInfo const & info);
 
-  static bool LessRank(PreRankerResult const & r1, PreRankerResult const & r2);
+  static bool LessRankAndPopularity(PreRankerResult const & r1, PreRankerResult const & r2);
   static bool LessDistance(PreRankerResult const & r1, PreRankerResult const & r2);
 
   FeatureID const & GetId() const { return m_id; }
diff --git a/search/pre_ranker.cpp b/search/pre_ranker.cpp
index 4ed40dea2b..4db663baaf 100644
--- a/search/pre_ranker.cpp
+++ b/search/pre_ranker.cpp
@@ -30,8 +30,9 @@ void SweepNearbyResults(double eps, set<FeatureID> const & prevEmit, vector<PreR
   {
     auto const & p = results[i].GetInfo().m_center;
     uint8_t const rank = results[i].GetInfo().m_rank;
+    uint8_t const popularity = results[i].GetInfo().m_popularity;
     uint8_t const prevCount = prevEmit.count(results[i].GetId()) ? 1 : 0;
-    uint8_t const priority = max(rank, prevCount);
+    uint8_t const priority = max({rank, prevCount, popularity});
     sweeper.Add(p.x, p.y, i, priority);
   }
 
@@ -195,7 +196,7 @@ void PreRanker::Filter(bool viewportSearch)
   {
     size_t n = min(m_results.size(), BatchSize());
     nth_element(m_results.begin(), m_results.begin() + n, m_results.end(),
-                &PreRankerResult::LessRank);
+                &PreRankerResult::LessRankAndPopularity);
     filtered.insert(m_results.begin(), m_results.begin() + n);
   }
 
diff --git a/search/ranking_info.cpp b/search/ranking_info.cpp
index 37b526de3f..b87cb72616 100644
--- a/search/ranking_info.cpp
+++ b/search/ranking_info.cpp
@@ -14,6 +14,8 @@ namespace
 // these coeffs correspond to coeffs in a linear model.
 double const kDistanceToPivot = -1.0000000;
 double const kRank = 1.0000000;
+// todo: (@t.yan) Adjust.
+double const kPopularity = 1.0000000;
 double const kFalseCats = -0.0839847;
 double const kErrorsMade = 0.0066984;
 double const kAllTokensUsed = 0.0000000;
@@ -96,6 +98,7 @@ double RankingInfo::GetLinearModelRank() const
   // integrated in the build system.
   double const distanceToPivot = TransformDistance(m_distanceToPivot);
   double const rank = static_cast<double>(m_rank) / numeric_limits<uint8_t>::max();
+  double const popularity = static_cast<double>(m_popularity) / numeric_limits<uint8_t>::max();
 
   auto nameScore = m_nameScore;
   if (m_pureCats || m_falseCats)
@@ -112,6 +115,7 @@ double RankingInfo::GetLinearModelRank() const
   double result = 0.0;
   result += kDistanceToPivot * distanceToPivot;
   result += kRank * rank;
+  result += kPopularity * popularity;
   result += kNameScore[nameScore];
   result += kErrorsMade * GetErrorsMade();
   result += kType[m_type];
diff --git a/search/search_quality/scoring_model.py b/search/search_quality/scoring_model.py
index aa333ea52e..9ccc36bf4e 100755
--- a/search/search_quality/scoring_model.py
+++ b/search/search_quality/scoring_model.py
@@ -16,10 +16,11 @@ import sys
 
 MAX_DISTANCE_METERS = 2e6
 MAX_RANK = 255
+MAX_POPULARITY = 255
 RELEVANCES = {'Irrelevant': 0, 'Relevant': 1, 'Vital': 3}
 NAME_SCORES = ['Zero', 'Substring', 'Prefix', 'Full Match']
 SEARCH_TYPES = ['POI', 'Building', 'Street', 'Unclassified', 'Village', 'City', 'State', 'Country']
-FEATURES = ['DistanceToPivot', 'Rank', 'FalseCats', 'ErrorsMade', 'AllTokensUsed'] + NAME_SCORES + SEARCH_TYPES
+FEATURES = ['DistanceToPivot', 'Rank', 'Popularity', 'FalseCats', 'ErrorsMade', 'AllTokensUsed'] + NAME_SCORES + SEARCH_TYPES
 
 BOOTSTRAP_ITERATIONS = 10000
 
@@ -36,6 +37,7 @@ def normalize_data(data):
 
     data['DistanceToPivot'] = data['DistanceToPivot'].apply(transform_distance)
     data['Rank'] = data['Rank'].apply(lambda v: v / MAX_RANK)
+    data['Popularity'] = data['Popularity'].apply(lambda v: v / MAX_POPULARITY)
     data['Relevance'] = data['Relevance'].apply(lambda v: RELEVANCES[v])
 
     cats = data['PureCats'].combine(data['FalseCats'], max)