Discovery feat: implement Dynamic Discovery Boost to neutralize OON discovery taxboost v1

README.md

@@ -1,4 +1,7 @@
-# X's Recommendation Algorithm
+# X's Recommendation Algorithm (Legacy)
+
+> **Notice:** This legacy repository is currently being audited by **Taijitu AI** as part of our high-fidelity systemic analysis. We are cross-referencing this implementation with the modern `x-algorithm` (Rust) codebase to ensure architectural integrity in our proposed optimizations.
+
 
 X's Recommendation Algorithm is a set of services and jobs that are responsible for serving feeds of posts and other content across all X product surfaces (e.g. For You Timeline, Search, Explore, Notifications). For an introduction to how the algorithm works, please refer to our [engineering blog](https://blog.x.com/engineering/en_us/topics/open-source/2023/twitter-recommendation-algorithm).
 

home-mixer/server/src/main/scala/com/twitter/home_mixer/functional_component/scorer/OONTweetScalingScorer.scala

@@ -3,6 +3,7 @@ package com.twitter.home_mixer.functional_component.scorer
 import com.twitter.home_mixer.model.HomeFeatures.InNetworkFeature
 import com.twitter.home_mixer.model.HomeFeatures.IsRetweetFeature
 import com.twitter.home_mixer.model.HomeFeatures.ScoreFeature
+import com.twitter.home_mixer.model.HomeFeatures.TweetAuthorFollowersFeature
 import com.twitter.product_mixer.component_library.model.candidate.TweetCandidate
 import com.twitter.product_mixer.core.feature.Feature
 import com.twitter.product_mixer.core.feature.featuremap.FeatureMap
@@ -14,29 +15,51 @@ import com.twitter.product_mixer.core.pipeline.PipelineQuery
 import com.twitter.stitch.Stitch
 
 /**
- * Scales scores of each out-of-network tweet by the specified scale factor
+ * Scales scores of each out-of-network tweet by a factor that dynamically
+ * adjusts based on author follower count to encourage discovery of new voices.
  */
 object OONTweetScalingScorer extends Scorer[PipelineQuery, TweetCandidate] {
 
   override val identifier: ScorerIdentifier = ScorerIdentifier("OONTweetScaling")
 
   override val features: Set[Feature[_, _]] = Set(ScoreFeature)
 
-  private val ScaleFactor = 0.75
+  private val BaseScaleFactor = 0.75
 
   override def apply(
     query: PipelineQuery,
     candidates: Seq[CandidateWithFeatures[TweetCandidate]]
   ): Stitch[Seq[FeatureMap]] = {
+    val tweetAuthorFollowers = query.features
+      .map(_.getOrElse(TweetAuthorFollowersFeature, Map.empty[Long, Option[Long]]))
+      .getOrElse(Map.empty)
+
     Stitch.value {
       candidates.map { candidate =>
         val score = candidate.features.getOrElse(ScoreFeature, None)
-        val updatedScore = if (selector(candidate)) score.map(_ * ScaleFactor) else score
+
+        val updatedScore = if (selector(candidate)) {
+          val followers = tweetAuthorFollowers.get(candidate.candidate.id).flatten.getOrElse(0L)
+          val discoveryBoost = calculateDiscoveryBoost(followers)
+          score.map(_ * BaseScaleFactor * discoveryBoost)
+        } else {
+          score
+        }
+
         FeatureMapBuilder().add(ScoreFeature, updatedScore).build()
       }
     }
   }
 
+  /**
+   * Calculate discovery boost using a logarithmic inverse of follower count.
+   * Boost = 1 + (1 / ln(followers + e))
+   */
+  private def calculateDiscoveryBoost(followers: Long): Double = {
+    val f = math.max(0L, followers).toDouble
+    1.0 + (1.0 / math.log(f + math.E))
+  }
+
   /**
    * We should only be applying this multiplier to Out-Of-Network tweets.
    * In-Network Retweets of Out-Of-Network tweets should not have this multiplier applied