From: Nathan TeBlunthuis Date: Tue, 29 Nov 2022 08:29:42 +0000 (-0800) Subject: Add exploratory data analysis to come up with a real-data example. X-Git-Url: https://code.communitydata.science/ml_measurement_error_public.git/commitdiff_plain/3d1964b806106d76f13301f0cf6dccf35cd7d66c?ds=sidebyside Add exploratory data analysis to come up with a real-data example. --- diff --git a/civil_comments/Makefile b/civil_comments/Makefile new file mode 100644 index 0000000..3c331c2 --- /dev/null +++ b/civil_comments/Makefile @@ -0,0 +1,6 @@ +srun_1core=srun -A comdata -p compute-bigmem --mem=4G --time=12:00:00 -c 1 --pty /usr/bin/bash -l +perspective_scores.csv: perspective_json_to_csv.sh perspective_results.json + $(srun_1core) ./$^ $@ + + + diff --git a/civil_comments/design_example.R b/civil_comments/design_example.R new file mode 100644 index 0000000..5991334 --- /dev/null +++ b/civil_comments/design_example.R @@ -0,0 +1,156 @@ +library(data.table) +library(MASS) + +scores <- fread("perspective_scores.csv") +scores <- scores[,id:=as.character(id)] + +df <- fread("all_data.csv") + +# only use the data that has identity annotations +df <- df[identity_annotator_count > 0] + +(df[!(df$id %in% scores$id)]) + +df <- df[scores,on='id',nomatch=NULL] + +## 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) +} + +df[, ":="(identity_attack_pred = identity_attack_prob >=0.5, + insult_pred = insult_prob >= 0.5, + profanity_pred = profanity_prob >= 0.5, + severe_toxicity_pred = severe_toxicity_prob >= 0.5, + threat_pred = threat_prob >= 0.5, + toxicity_pred = toxicity_prob >= 0.5, + identity_attack_coded = identity_attack >= 0.5, + insult_coded = insult >= 0.5, + profanity_coded = obscene >= 0.5, + severe_toxicity_coded = severe_toxicity >= 0.5, + threat_coded = threat >= 0.5, + toxicity_coded = toxicity >= 0.5 + )] + + + +## 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"] + +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',] + +glm(white ~ toxicity_coded + psychiatric_or_mental_illness, data = df, family=binomial(link='logit')) + +glm(white ~ toxicity_pred + psychiatric_or_mental_illness, data = df, family=binomial(link='logit')) + +m1 <- 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_coded, data = df) + +m2 <- 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')) diff --git a/civil_comments/perspective_json_to_csv.jq b/civil_comments/perspective_json_to_csv.jq new file mode 100755 index 0000000..0688d21 --- /dev/null +++ b/civil_comments/perspective_json_to_csv.jq @@ -0,0 +1,2 @@ +#!/usr/bin/bash +cat $1 | jq "[.attributeScores.IDENTITY_ATTACK.summaryScore.value, .attributeScores.INSULT.summaryScore.value, .attributeScores.PROFANITY.summaryScore.value,.attributeScores.SEVERE_TOXICITY.summaryScore.value, .attributeScores.THREAT.summaryScore.value, .attributeScores.TOXICITY.summaryScore.value] | @csv" > $2 diff --git a/civil_comments/perspective_json_to_csv.sh b/civil_comments/perspective_json_to_csv.sh new file mode 100755 index 0000000..8e2aa13 --- /dev/null +++ b/civil_comments/perspective_json_to_csv.sh @@ -0,0 +1,4 @@ +#!/usr/bin/bash +header=id,identity_attack_prob,insult_prob,profanity_prob,severe_toxicity_prob,threat_prob,toxicity_prob +echo "$header" > $2 +cat $1 | jq -r '[.id, .attributeScores.IDENTITY_ATTACK.summaryScore.value, .attributeScores.INSULT.summaryScore.value, .attributeScores.PROFANITY.summaryScore.value,.attributeScores.SEVERE_TOXICITY.summaryScore.value, .attributeScores.THREAT.summaryScore.value, .attributeScores.TOXICITY.summaryScore.value] | @csv' >> $2