]> code.communitydata.science - ml_measurement_error_public.git/blobdiff - civil_comments/design_example.R
update real data examples code and rerun project.
[ml_measurement_error_public.git] / civil_comments / design_example.R
index 1a83a81a791adf9f167723d885ded05dd3d1d0eb..49076880fb5f68ff362dc6730fdec8ac99f542b1 100644 (file)
@@ -5,95 +5,6 @@ source('load_perspective_data.R')
 ## the API claims that these scores are "probabilities"
 ## say we care about the model of the classification, not the probability
 
-F1 <- function(y, predictions){
-    tp <- sum( (predictions == y) & (predictions==1))
-    fn <- sum( (predictions != y) & (predictions!=1))
-    fp <- sum( (predictions != y) & (predictions==1))
-    precision <- tp / (tp + fp)
-    recall <- tp / (tp + fn)
-    return (2 * precision * recall ) / (precision + recall)
-}
-
-
-## toxicity is about 93% accurate, with an f1 of 0.8
-## identity_attack has high accuracy 97%, but an unfortunant f1 of 0.5.
-## threat has high accuracy 99%, but a really bad looking f1 of 0.48.
-accuracies <- df[,.(identity_attack_acc = mean(identity_attack_pred == identity_attack_coded),
-                    insult_pred_acc = mean(insult_pred == insult_coded),
-                    profanity_acc = mean(profanity_pred == profanity_coded),
-                    severe_toxicity_acc = mean(severe_toxicity_pred == severe_toxicity_coded),
-                    theat_acc = mean(threat_pred == threat_coded),
-                    toxicity_acc = mean(toxicity_pred == toxicity_coded))]
-
-f1s <- df[,.(identity_attack_f1 = F1(identity_attack_coded,identity_attack_pred),
-                    insult_f1 = F1(insult_coded,insult_pred),
-                    profanity_f1 = F1(profanity_coded,profanity_pred),
-                    severe_toxicity_f1 = F1(severe_toxicity_coded,severe_toxicity_pred),
-                    theat_f1 = F1(threat_coded,threat_pred),
-                    toxicity_f1 = F1(toxicity_coded,toxicity_pred))]
-
-positive_cases <- df[,.(identity_attacks = sum(identity_attack_coded),
-                        insults = sum(insult_coded),
-                        profanities = sum(profanity_coded),
-                        severe_toxic_comments = sum(severe_toxicity_coded),
-                        threats = sum(threat_coded),
-                        toxic_comments = sum(toxicity_coded))]
-
-## there are 50,000 toxic comments, 13000 identity attacks, 30000 insults, 3000 profanities, 8 severe toxic, and 1000 threats.
-
-proportions_cases <- df[,.(prop_identity = mean(identity_attack_coded),
-                           prop_insults = mean(insult_coded),
-                           prop_profanity = mean(profanity_coded),
-                           prop_severe = mean(severe_toxicity_coded),
-                           prop_threats = mean(threat_coded),
-                           prop_toxic = mean(toxicity_coded))]
-
-## at 11% of comments, "toxicity" seems not so badly skewed. Try toxicity first, and if it doesn't work out try insults.
-
-## now look for an example where differential error affects an identity, or a reaction.
-df <- df[,":="(identity_error = identity_attack_coded - identity_attack_pred,
-               insult_error = insult_coded - insult_pred,
-               profanity_error = profanity_coded - profanity_pred,
-               severe_toxic_error = severe_toxicity_coded - severe_toxicity_pred,
-               threat_error = threat_coded - threat_pred,
-               toxicity_error = toxicity_coded - toxicity_pred)]
-
-## what's correlated with toxicity_error ?
-df <- df[,approved := rating == "approved"]
-df <- df[,white := white > 0.5]
-
-cortab <- cor(df[,.(toxicity_error,
-                    identity_error,
-                    toxicity_coded,
-                    funny,
-                    approved,
-                    sad,
-                    wow,
-                    likes,
-                    disagree,
-                    male,
-                    female,
-                    transgender,
-                    other_gender,
-                    heterosexual,
-                    bisexual,
-                    other_sexual_orientation,
-                    christian,
-                    jewish,
-                    hindu,
-                    buddhist,
-                    atheist,
-                    other_religion,
-                    black,
-                    white,
-                    asian,
-                    latino,
-                    other_race_or_ethnicity,
-                    physical_disability,
-                    intellectual_or_learning_disability,
-                    psychiatric_or_mental_illness,
-                    other_disability)])
-
 ## toxicity error is weakly correlated pearson's R = 0.1 with both "white" and "black".
 
 ## compare regressions with "white" or "black" as the outcome and "toxicity_coded" or "toxicity_pred" as a predictor.
@@ -107,22 +18,6 @@ cortab['toxicity_coded',]
 cortab['identity_error',]
 cortab['white',]
 
-cortab <- cor(df[,.(toxicity_error,
-                    identity_error,
-                    toxicity_coded,
-                    funny,
-                    approved,
-                    sad,
-                    wow,
-                    likes,
-                    disagree,
-                    gender_disclosed,
-                    sexuality_disclosed,
-                    religion_disclosed,
-                    race_disclosed,
-                    disability_disclosed)])
-
-
 ## here's a simple example, is P(white | toxic and mentally ill) > P(white | toxic or mentally ill). Are people who discuss their mental illness in a toxic way more likely to be white compared to those who just talk about their mental illness or are toxic? 
 summary(glm(white ~ toxicity_coded*psychiatric_or_mental_illness, data = df, family=binomial(link='logit')))
 

Community Data Science Collective || Want to submit a patch?