Add confidence to AutocompleteMatch.

chrome/browser/autocomplete/history_url_provider.cc

Index: chrome/browser/autocomplete/history_url_provider.cc
diff --git a/chrome/browser/autocomplete/history_url_provider.cc b/chrome/browser/autocomplete/history_url_provider.cc
index b8581a09c9f26c5ff40c4c5605c6de768df346ed..2a15cf108be07a0867abd925f4f0a57c81f45c9a 100644
--- a/chrome/browser/autocomplete/history_url_provider.cc
+++ b/chrome/browser/autocomplete/history_url_provider.cc
@@ -245,7 +245,8 @@ void HistoryURLProvider::DoAutocomplete(history::HistoryBackend* backend,
     params->matches.push_back(what_you_typed_match);
   } else if (params->prevent_inline_autocomplete ||
       history_matches.empty() ||
-      !PromoteMatchForInlineAutocomplete(params, history_matches.front())) {
+      !PromoteMatchForInlineAutocomplete(params, history_matches.front(),
+                                         history_matches)) {
     // Failed to promote any URLs for inline autocompletion.  Use the What You
     // Typed match, if we have it.
     first_match = 0;
@@ -268,8 +269,13 @@ void HistoryURLProvider::DoAutocomplete(history::HistoryBackend* backend,
     DCHECK(!have_what_you_typed_match ||
            (match.url_info.url() !=
             GURL(params->matches.front().destination_url)));
-    params->matches.push_back(HistoryMatchToACMatch(params, match, NORMAL,
-        history_matches.size() - 1 - i));
+    // TODO(dominich): Should this be history_matches.size() - first_match - i?

[Peter Kasting, 2011/08/09 20:53:00]

No, it should not.

[dominich, 2011/08/09 21:43:43]

Done.

+    AutocompleteMatch ac_match =
+        HistoryMatchToACMatch(params, match, history_matches, NORMAL,
+                              history_matches.size() - 1 - i);
+    UMA_HISTOGRAM_COUNTS_100("Autocomplete.Confidence_HistoryUrl",
+                             ac_match.confidence * 100);
+    params->matches.push_back(ac_match);
   }
 }
 
@@ -302,9 +308,12 @@ void HistoryURLProvider::QueryComplete(
 AutocompleteMatch HistoryURLProvider::SuggestExactInput(
     const AutocompleteInput& input,
     bool trim_http) {
+  // TODO(dominich): Check that 100% confidence makes sense for exact input.

[Peter Kasting, 2011/08/09 20:53:00]

It doesn't.  This will mean the what-you-typed match is always scored as 100%,
which is definitely wrong sometimes.  For example, if we promote a match for
inline autocompletion, that match will outrank this one, which suggests it
should have higher confidence as well.  Or if the user types some characters and
then pastes, then the "ctrl" keydown during pasting will add ".com" to the end
of the current typing and this will then score that at 100%.

[dominich, 2011/08/09 21:43:43]

You're mixing up the notion of rank and confidence. While they can be mixed it's
not clear to me that they should be. The rank is a cross-provider measure while
the confidence is a per-provider measure. It also is a suggestion of what we
should do with the result.

100% confident here means that /if/ the what-you-type match is the top result
then we should definitely apply instant or prerender to it.

The issue of the pasting is an interesting one but I'm not sure how we can avoid
that and still keep the confidence here relatively high.

I will set the confidence to 0 here for the first pass of stats so it doesn't
pollute them.

   AutocompleteMatch match(this,
-      CalculateRelevance(input.type(), WHAT_YOU_TYPED, 0), false,
+      CalculateRelevance(input.type(), WHAT_YOU_TYPED, 0), 1.0f, false,
       AutocompleteMatch::URL_WHAT_YOU_TYPED);
+  UMA_HISTOGRAM_COUNTS_100("Autocomplete.Confidence_HistoryUrl",
+                           match.confidence * 100);
 
   const GURL& url = input.canonicalized_url();
   if (url.is_valid()) {
@@ -411,7 +420,8 @@ bool HistoryURLProvider::FixupExactSuggestion(history::URLDatabase* db,
 
 bool HistoryURLProvider::PromoteMatchForInlineAutocomplete(
     HistoryURLProviderParams* params,
-    const HistoryMatch& match) {
+    const HistoryMatch& match,
+    const HistoryMatches& matches) {
   // Promote the first match if it's been typed at least n times, where n == 1
   // for "simple" (host-only) URLs and n == 2 for others.  We set a higher bar
   // for these long URLs because it's less likely that users will want to visit
@@ -430,7 +440,7 @@ bool HistoryURLProvider::PromoteMatchForInlineAutocomplete(
   // there's no way to know about "foo/", make reaching this point prevent any
   // future pass from suggesting the exact input as a better match.
   params->dont_suggest_exact_input = true;
-  params->matches.push_back(HistoryMatchToACMatch(params, match,
+  params->matches.push_back(HistoryMatchToACMatch(params, match, matches,
                                                   INLINE_AUTOCOMPLETE, 0));
   return true;
 }
@@ -471,6 +481,22 @@ int HistoryURLProvider::CalculateRelevance(AutocompleteInput::Type input_type,
 }
 
 // static
+float HistoryURLProvider::CalculateConfidence(
+    const history::HistoryMatch& match,
+    const history::HistoryMatches& matches) {
+  // TODO(dominich): Take into account typed count?

[Peter Kasting, 2011/08/09 20:53:00]

See comments in HCP for why typed_count is better than visit_count.

You also don't check some important clues like IsHostOnly().

[dominich, 2011/08/09 21:43:43]

It looks like CompareHistoryMatch is a good starting point for a more inclusive
measure.

I would like to get in a simple (naive) version with histograms and then try
different calculations for each provider to get the best results, ie increase
the proportion of successful prerenders.

+  // TODO(dominich): Take into account bookmarked page?
+  float numerator = match.url_info.visit_count();
+  float denominator = 0.0f;
+  for (history::HistoryMatches::const_iterator it = matches.begin();
+       it != matches.end(); ++it) {

[Peter Kasting, 2011/08/09 20:53:00]

Nit: No need for {}

[dominich, 2011/08/09 21:43:43]

See comment on HCP as to why I prefer to keep this.

+    denominator += it->url_info.visit_count();
+  }
+  DCHECK(denominator > 0);
+  return numerator / denominator;
+}
+
+// static
 void HistoryURLProvider::PromoteOrCreateShorterSuggestion(
     history::URLDatabase* db,
     const HistoryURLProviderParams& params,
@@ -774,11 +800,13 @@ size_t HistoryURLProvider::RemoveSubsequentMatchesOf(
 AutocompleteMatch HistoryURLProvider::HistoryMatchToACMatch(
     HistoryURLProviderParams* params,
     const HistoryMatch& history_match,
+    const HistoryMatches& history_matches,
     MatchType match_type,
     size_t match_number) {
   const history::URLRow& info = history_match.url_info;
   AutocompleteMatch match(this,
       CalculateRelevance(params->input.type(), match_type, match_number),
+      CalculateConfidence(history_match, history_matches),
       !!info.visit_count(), AutocompleteMatch::HISTORY_URL);
   match.destination_url = info.url();
   DCHECK(match.destination_url.is_valid());