set.seed(1111) source('load_perspective_data.R') ## how accurate are the classifiers? ## 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. ## here's a great example with presumambly non-differential error: about what identities is toxicity found humorous? ## a bunch of stars reappear when you used the ground-truth data instead of the predictions. ## pro/con of this example: will have to implement family='poisson'. ## shouldn't be that bad though haha. cortab['toxicity_error',] cortab['toxicity_error','funny'] 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'))) summary(glm(white ~ toxicity_pred*psychiatric_or_mental_illness, data = df, family=binomial(link='logit'))) summary(glm(white ~ toxicity_coded*male, data = df, family=binomial(link='logit'))) summary(glm(white ~ toxicity_pred*male, data = df, family=binomial(link='logit'))) summary(glm(toxicity_coded ~ white*psychiatric_or_mental_illness, data = df, family=binomial(link='logit'))) summary(glm(toxicity_pred ~ white*psychiatric_or_mental_illness, data = df, family=binomial(link='logit'))) ## another simple enough example: is P(toxic | funny and white) > P(toxic | funny nand white)? Or, are funny comments more toxic when people disclose that they are white? summary(glm(toxicity_pred ~ funny*white, data=df, family=binomial(link='logit'))) summary(glm(toxicity_coded ~ funny*white, data=df, family=binomial(link='logit'))) source("../simulations/measerr_methods.R") saved_model_file <- "measerr_model_tox.eq.funny.cross.white.RDS" overwrite_model <- TRUE # it works so far with a 20% and 15% sample. Smaller is better. let's try a 10% sample again. It didn't work out. We'll go forward with a 15% sample. df_measerr_method <- copy(df)[sample(1:.N, 0.05 * .N), toxicity_coded_1 := toxicity_coded] df_measerr_method <- df_measerr_method[,toxicity_coded := toxicity_coded_1] summary(glm(toxicity_coded ~ funny*white, data=df_measerr_method[!is.na(toxicity_coded)], family=binomial(link='logit'))) if(!file.exists(saved_model_file) || (overwrite_model == TRUE)){ measerr_model <- measerr_mle_dv(df_measerr_method,toxicity_coded ~ funny*white,outcome_family=binomial(link='logit'), proxy_formula=toxicity_pred ~ toxicity_coded*funny*white) saveRDS(measerr_model, saved_model_file) } else { measerr_model <- readRDS(saved_model_file) } inv_hessian <- solve(measerr_model$hessian) se <- diag(inv_hessian) lm2 <- glm.nb(funny ~ (male + female + transgender + other_gender + heterosexual + bisexual + other_sexual_orientation + christian + jewish + hindu + buddhist + atheist + other_religion + asian + latino + other_race_or_ethnicity + physical_disability + intellectual_or_learning_disability + white + black + psychiatric_or_mental_illness)*toxicity_pred, data = df) m3 <- glm.nb(funny ~ (male + female + transgender + other_gender + heterosexual + bisexual + other_sexual_orientation + christian + jewish + hindu + buddhist + atheist + other_religion + asian + latino + other_race_or_ethnicity + physical_disability + intellectual_or_learning_disability + white + black + psychiatric_or_mental_illness)*toxicity, data = df) glm(white ~ disagree, data = df, family=binomial(link='logit')) ## example with differential error glm(white ~ toxicity_coded + toxicity_error, data=df,family=binomial(link='logit')) glm(toxicity_coded ~ white, data = df, family=binomial(link='logit')) glm(toxicity_pred ~ white, data = df, family=binomial(link='logit'))