Update harassment-abuse tag-consensus model penalty, update user enrollment, and re-tune diligence model

sourcecode/scoring/constants.py

@@ -47,6 +47,7 @@
 authorTopNotHelpfulTagValues = "authorTopNotHelpfulTagValues"
 modelingPopulationKey = "modelingPopulation"
 modelingGroupKey = "modelingGroup"
+numberOfTimesEarnedOutKey = "numberOfTimesEarnedOut"
 
 # TSV Values
 notHelpfulValueTsv = "NOT_HELPFUL"
@@ -416,10 +417,16 @@ def rater_factor_key(i):
   (timestampOfLastEarnOut, np.double),  # double because nullable.
   (modelingPopulationKey, str),
   (modelingGroupKey, np.float64),
+  (numberOfTimesEarnedOutKey, np.int64),
 ]
 userEnrollmentTSVColumns = [col for (col, _) in userEnrollmentTSVColumnsAndTypes]
 userEnrollmentTSVTypes = [dtype for (_, dtype) in userEnrollmentTSVColumnsAndTypes]
 userEnrollmentTSVTypeMapping = {col: dtype for (col, dtype) in userEnrollmentTSVColumnsAndTypes}
+# TODO: Remove the "old" user enrollment schemas below once numberOfTimesEarnedOut is in production
+userEnrollmentTSVColumnsOld = [col for (col, _) in userEnrollmentTSVColumnsAndTypes[:7]]
+userEnrollmentTSVTypeMappingOld = {
+  col: dtype for (col, dtype) in userEnrollmentTSVColumnsAndTypes[:7]
+}
 
 noteInterceptMaxKey = "internalNoteIntercept_max"
 noteInterceptMinKey = "internalNoteIntercept_min"
@@ -564,6 +571,7 @@ def rater_factor_key(i):
   (groupRaterFactor1Key, np.double),
   (modelingGroupKey, np.float64),
   (raterHelpfulnessReputationKey, np.double),
+  (numberOfTimesEarnedOutKey, np.int64),
 ]
 raterModelOutputTSVColumns = [col for (col, dtype) in raterModelOutputTSVColumnsAndTypes]
 raterModelOutputTSVTypeMapping = {col: dtype for (col, dtype) in raterModelOutputTSVColumnsAndTypes}

sourcecode/scoring/contributor_state.py

@@ -56,7 +56,6 @@ def is_earned_out(authorEnrollmentCounts: pd.DataFrame):
   return (
     (authorEnrollmentCounts[c.enrollmentState] != c.newUser)
     & (authorEnrollmentCounts[c.enrollmentState] != c.earnedOutAcknowledged)
-    & (authorEnrollmentCounts[c.enrollmentState] != c.earnedOutNoAcknowledge)
     & (authorEnrollmentCounts[c.notesCurrentlyRatedNotHelpful] > c.isAtRiskCRNHCount)
   )
 
@@ -401,9 +400,27 @@ def get_contributor_state(
     contributorScoresWithEnrollment.loc[
       is_at_risk(contributorScoresWithEnrollment), c.enrollmentState
     ] = c.enrollmentStateToThrift[c.atRisk]
+
+    # for earned out users, first increment the number of times they have earned out,
+    # use this to overwrite successful rating needed to earn in,
+    # then set new state
+    earnedOutUsers = is_earned_out(contributorScoresWithEnrollment)
+    contributorScoresWithEnrollment.loc[earnedOutUsers, c.numberOfTimesEarnedOutKey] = (
+      contributorScoresWithEnrollment.loc[earnedOutUsers, c.numberOfTimesEarnedOutKey] + 1
+    )
+
+    contributorScoresWithEnrollment.loc[
+      earnedOutUsers, c.successfulRatingNeededToEarnIn
+    ] = contributorScoresWithEnrollment.loc[earnedOutUsers].apply(
+      lambda row: c.ratingImpactForEarnIn
+      + max([row[c.ratingImpact], 0])
+      + (c.ratingImpactForEarnIn * row[c.numberOfTimesEarnedOutKey]),
+      axis=1,
+    )
+
     contributorScoresWithEnrollment.loc[
-      is_earned_out(contributorScoresWithEnrollment), c.enrollmentState
-    ] = c.enrollmentStateToThrift[c.earnedOutNoAcknowledge]
+      earnedOutUsers, c.enrollmentState
+    ] = c.enrollmentStateToThrift[c.earnedOutAcknowledged]
 
     contributorScoresWithEnrollment.loc[
       is_earned_in(contributorScoresWithEnrollment), c.enrollmentState

sourcecode/scoring/helpfulness_scores.py

@@ -75,8 +75,8 @@ def compute_general_helpfulness_scores(
   ratings: Optional[pd.DataFrame] = None,
   tagConsensusHarassmentAbuseNotes: Optional[pd.DataFrame] = None,
   tagConsensusHarassmentHelpfulRatingPenalty=10,
-  multiplyPenaltyByHarassmentScore: bool = False,
-  minimumHarassmentScoreToPenalize: float = 2.5,
+  multiplyPenaltyByHarassmentScore: bool = True,
+  minimumHarassmentScoreToPenalize: float = 2.0,
 ) -> pd.DataFrame:
   """Given notes scored by matrix factorization, compute helpfulness scores.
   Author helpfulness scores are based on the scores of the notes you wrote.

sourcecode/scoring/mf_base_scorer.py

@@ -102,7 +102,7 @@ def __init__(
     crnhThresholdNMIntercept: float = -0.15,
     crnhThresholdUCBIntercept: float = -0.04,
     crhSuperThreshold: float = 0.5,
-    lowDiligenceThreshold: float = 0.217,
+    lowDiligenceThreshold: float = 0.263,
     factorThreshold: float = 0.5,
     inertiaDelta: float = 0.01,
     useStableInitialization: bool = True,
@@ -117,6 +117,9 @@ def __init__(
     globalInterceptLambda=None,
     diamondLambda=None,
     normalizedLossHyperparameters=None,
+    multiplyPenaltyByHarassmentScore: bool = True,
+    minimumHarassmentScoreToPenalize: float = 2.0,
+    tagConsensusHarassmentHelpfulRatingPenalty: int = 10,
   ):
     """Configure MatrixFactorizationScorer object.
 
@@ -174,6 +177,9 @@ def __init__(
     self._maxFinalMFTrainError = maxFinalMFTrainError
     self._lowDiligenceThreshold = lowDiligenceThreshold
     self._factorThreshold = factorThreshold
+    self.multiplyPenaltyByHarassmentScore = multiplyPenaltyByHarassmentScore
+    self.minimumHarassmentScoreToPenalize = minimumHarassmentScoreToPenalize
+    self.tagConsensusHarassmentHelpfulRatingPenalty = tagConsensusHarassmentHelpfulRatingPenalty
     mfArgs = dict(
       [
         pair
@@ -460,6 +466,7 @@ def _score_notes_and_users(
         c.notHelpfulSpamHarassmentOrAbuseTagKey,
         noteParamsUnfiltered,
         raterParamsUnfiltered,
+        name="harassment",
       )
 
     # Assigns contributor (author & rater) helpfulness bit based on (1) performance
@@ -481,6 +488,9 @@ def _score_notes_and_users(
         self._minRaterAgreeRatio,
         ratings=ratingsForTraining,
         tagConsensusHarassmentAbuseNotes=harassmentAbuseNoteParams,
+        tagConsensusHarassmentHelpfulRatingPenalty=self.tagConsensusHarassmentHelpfulRatingPenalty,
+        multiplyPenaltyByHarassmentScore=self.multiplyPenaltyByHarassmentScore,
+        minimumHarassmentScoreToPenalize=self.minimumHarassmentScoreToPenalize,
       )
 
       # Filters ratings matrix to include only rows (ratings) where the rater was

sourcecode/scoring/mf_group_scorer.py

@@ -119,8 +119,11 @@ def __init__(
     crnhThresholdNoteFactorMultiplier: float = -0.8,
     crnhThresholdNMIntercept: float = -0.15,
     crhSuperThreshold: float = 0.5,
-    lowDiligenceThreshold: float = 0.217,
+    lowDiligenceThreshold: float = 0.263,
     factorThreshold: float = 0.5,
+    multiplyPenaltyByHarassmentScore: bool = True,
+    minimumHarassmentScoreToPenalize: float = 2.0,
+    tagConsensusHarassmentHelpfulRatingPenalty: int = 10,
   ) -> None:
     """Configure MFGroupScorer object.
 
@@ -161,6 +164,9 @@ def __init__(
       crhSuperThreshold=crhSuperThreshold,
       lowDiligenceThreshold=lowDiligenceThreshold,
       factorThreshold=factorThreshold,
+      multiplyPenaltyByHarassmentScore=multiplyPenaltyByHarassmentScore,
+      minimumHarassmentScoreToPenalize=minimumHarassmentScoreToPenalize,
+      tagConsensusHarassmentHelpfulRatingPenalty=tagConsensusHarassmentHelpfulRatingPenalty,
     )
     assert groupNumber > 0, "groupNumber must be positive.  0 is reserved for unassigned."
     assert groupNumber <= groupScorerCount, "groupNumber exceeds maximum expected groups."

sourcecode/scoring/note_ratings.py

@@ -366,7 +366,7 @@ def compute_scored_notes(
   is_crh_function: Callable[..., pd.Series] = is_crh,
   is_crnh_diamond_function: Callable[..., pd.Series] = is_crnh_diamond,
   is_crnh_ucb_function: Callable[..., pd.Series] = is_crnh_ucb,
-  lowDiligenceThreshold: float = 0.217,
+  lowDiligenceThreshold: float = 0.263,
   factorThreshold: float = 0.5,
 ) -> pd.DataFrame:
   """

sourcecode/scoring/process_data.py

@@ -147,15 +147,28 @@ def read_from_tsv(
   if userEnrollmentPath is None:
     userEnrollment = None
   else:
-    userEnrollment = tsv_reader(
-      userEnrollmentPath, c.userEnrollmentTSVTypeMapping, c.userEnrollmentTSVColumns, header=headers
-    )
-    assert len(userEnrollment.columns.values) == len(c.userEnrollmentTSVColumns) and all(
-      userEnrollment.columns == c.userEnrollmentTSVColumns
-    ), (
-      f"userEnrollment columns don't match: \n{[col for col in userEnrollment.columns if not col in c.userEnrollmentTSVColumns]} are extra columns, "
-      + f"\n{[col for col in c.userEnrollmentTSVColumns if not col in userEnrollment.columns]} are missing."
-    )
+    try:
+      userEnrollment = tsv_reader(
+        userEnrollmentPath,
+        c.userEnrollmentTSVTypeMapping,
+        c.userEnrollmentTSVColumns,
+        header=headers,
+      )
+      assert len(userEnrollment.columns.values) == len(c.userEnrollmentTSVColumns) and all(
+        userEnrollment.columns == c.userEnrollmentTSVColumns
+      ), (
+        f"userEnrollment columns don't match: \n{[col for col in userEnrollment.columns if not col in c.userEnrollmentTSVColumns]} are extra columns, "
+        + f"\n{[col for col in c.userEnrollmentTSVColumns if not col in userEnrollment.columns]} are missing."
+      )
+    except ValueError:
+      # TODO: clean up fallback for old mappings once numberOfTimesEarnedOut column is in production
+      userEnrollment = tsv_reader(
+        userEnrollmentPath,
+        c.userEnrollmentTSVTypeMappingOld,
+        c.userEnrollmentTSVColumnsOld,
+        header=headers,
+      )
+      userEnrollment[c.numberOfTimesEarnedOutKey] = 0
 
   return notes, ratings, noteStatusHistory, userEnrollment
 

sourcecode/scoring/run_scoring.py

@@ -104,6 +104,9 @@ def _get_scorers(
         crnhThresholdNMIntercept=-0.02,
         lowDiligenceThreshold=1000,
         factorThreshold=0.4,
+        multiplyPenaltyByHarassmentScore=False,
+        minimumHarassmentScoreToPenalize=2.5,
+        tagConsensusHarassmentHelpfulRatingPenalty=10,
       )
     )
 
@@ -537,6 +540,7 @@ def _compute_helpfulness_scores(
           c.successfulRatingNeededToEarnIn,
           c.authorTopNotHelpfulTagValues,
           c.isEmergingWriterKey,
+          c.numberOfTimesEarnedOutKey,
         ]
       ],
       on=c.raterParticipantIdKey,

sourcecode/scoring/tag_consensus.py

@@ -1,3 +1,5 @@
+from typing import Optional
+
 from . import constants as c, process_data
 from .matrix_factorization.matrix_factorization import MatrixFactorization
 
@@ -10,7 +12,7 @@ def train_tag_model(
   helpfulModelNoteParams: pd.DataFrame = None,
   helpfulModelRaterParams: pd.DataFrame = None,
   useSigmoidCrossEntropy: bool = True,
-  name: str = "harassment",
+  name: Optional[str] = None,
 ):
   print(f"-------------------Training for tag {tag}-------------------")
   ratingDataForTag, labelColName = prepare_tag_data(ratings, tag)
@@ -62,6 +64,8 @@ def train_tag_model(
     noteInit=helpfulModelNoteParams,
   )
 
+  if name is None:
+    name = tag.split("elpful")[-1]
   noteParams.columns = [col.replace("internal", name) for col in noteParams.columns]
   raterParams.columns = [col.replace("internal", name) for col in raterParams.columns]
   return noteParams, raterParams, globalBias