Evaluating ORES Models

The Objective Revision Evaluation Service (ORES) is a web service running in Wikimedia Labs that provides machine learning as a service for Wikimedia Projects. (https://ores.wmflabs.org/). You can read more about it here.

We explored the bias that ORES’ damage detection models have against anonymous editors in Wikipedia. We did a statistical analysis to show that the bias is extreme and then changed our modeling strategy to minimize the bias re-measured the effects.

Thanks Aaron for the supervision!

V2 Model - Damaging

Example: https://ores.wmflabs.org/v2/scores/enwiki/damaging/623818349/

# Download revs_id
library(jsonlite)

document <- fromJSON("http://labels.wmflabs.org/campaigns/enwiki/4/?tasks")

# Find revs_is.labels.damaging
revs_id<- document$tasks$data$rev_id
damagings<- sapply(1:20000, function(x) document$tasks$labels[[x]]$data$damaging)

# Find damaging predictions
df<-as.data.frame(cbind(revs_id, damagings))
dfFalse<-select(filter(df, damagings!="TRUE"), revs_id)
dfTrue<-select(filter(df, damagings=="TRUE"), revs_id)

# Get Damaging Probabilities
getStats <- function(id) {
  rdTrue <- fromJSON(paste0("https://ores.wmflabs.org/v2/scores/enwiki/damaging/", id, "/?feature.is_anon=true"))
  rdFalse<-fromJSON(paste0("https://ores.wmflabs.org/v2/scores/enwiki/damaging/", id, "/?feature.is_anon=false"))
  
  rdTruePredictionTrue<-rdTrue$scores$enwiki$damaging$scores[[1]]$probability$true
  rdTruePredictionFalse<-rdFalse$scores$enwiki$damaging$scores[[1]]$probability$true
  
  return(list(rdTruePredictionTrue, rdTruePredictionFalse))
}

# Test  Model - Prediction is False
anonFalseModel<-sapply(1:500, function(x) getStats(dfFalse[x,]))

### Manipulating
library(dplyr)
library(reshape2)

anonFalseModelDF<-t(anonFalseModel)
colnames(anonFalseModelDF)<-c("anon_true", "anon_false")
anonFalseModelDF<-as.data.frame(anonFalseModelDF)
anonFalseModelDFManipulated<-melt(lapply(anonFalseModelDF, unlist))

### PlotFalse
library(ggplot2)

ggplot(anonFalseModelDFManipulated, aes(x=value)) + 
  geom_density(aes(group=L1, colour=L1, fill=L1), alpha=0.3) +
  labs(title="ORES damaging prediction is false", x="N=500") + 
  geom_vline(xintercept = c(0.8, 0.87, 0.94),  colour="green", linetype = "longdash")

###Calculate the difference
anonFalseModelDF$difference<-as.numeric(as.character(anonFalseModelDF$anon_true))-as.numeric(as.character(anonFalseModelDF$anon_false))
anonFalseModelDF<-anonFalseModelDF[, c("anon_true", "anon_false", "difference")]
anonFalseModelDFManipulated2<-melt(lapply(anonFalseModelDF, unlist))

### PlotTrue2
ggplot(anonFalseModelDFManipulated2, aes(x=value)) + 
  geom_density(aes(group=L1, colour=L1, fill=L1), alpha=0.3) +
  labs(title="ORES damaging prediction is false+difference", x="N=500") + 
  geom_vline(xintercept = c(0.8, 0.87, 0.94),  colour="green", linetype = "longdash")

Now, let’s calculate the true probabilities in the same way.

The New Model - Damaging

Example: https://ores-newmodels.wmflabs.org/v2/scores/enwiki/damaging/623818349/