opt: Only mark spam at a high confidence to reduce false positive

Author: ramsayleung

app/services/spam_classifier_service.rb

@@ -89,7 +89,16 @@ def classify(message_text)
       ham_score += Math.log(ham_count.to_f / (@classifier_state.total_ham_words + vocab_size))
     end
 
-    is_spam = spam_score > ham_score
+    diff = spam_score - ham_score
+    # stable logistic conversion
+    p_spam = if diff.abs > 700
+               diff > 0 ? 1.0 : 0.0
+    else
+               1.0 / (1.0 + Math.exp(-diff))
+    end
+
+    confidence_threshold = Rails.application.config.probability_threshold
+    is_spam = p_spam >= confidence_threshold
     [ is_spam, spam_score, ham_score ]
   end
 

config/application.rb

@@ -29,5 +29,7 @@ class Application < Rails::Application
     config.spam_ban_threshold =  3
     # Delete the warning message in x minutes to keep chat clean
     config.delete_message_delay = 5
+    # Spam blocked probability threshold
+    config.probability_threshold = 0.95
   end
 end