analyze_quality_models_revisions.R

   1 library(MASS)
   2 library(brms)
   3 options(mc.cores=28)
   4 library(ggplot2)
   5 library(data.table)
   6 library(arrow)
   7 library(wCorr)
   8
   9 source("RemembR/R/RemembeR.R")
  10
  11 change.remember.file("ordinal.quality.analysis_revisions.RDS")
  12
  13 #model.1 <- readRDS("models/ordinal_quality_intercept.RDS")
  14 model.main.pca <- readRDS("models/ordinal_quality_pca_revision.RDS")
  15 model.main.pca.cumulative <- readRDS("models/ordinal_quality_pca_revision.cumulative.RDS")
  16 model.qe6 <- readRDS("models/ordinal_quality_qe6_revision.RDS")
  17 df <- readRDS("data/training_quality_labels.RDS")
  18
  19 # then compare them with loo
  20 comparison.loo <- loo_compare(model.main.pca,model.qe6,model.main.pca.cumulative)
  21 #comparison.waic <- loo_compare(model.main.noC,model.main.noB,model.main.noFa,model.main.noGa,model.main.noStart,model.main.noStub,criterion='waic')
  22 print(comparison.loo,simplify=F,digits=2)
  23 remember(comparison.loo,"comparison.loo")
  24
  25 # LOO Chooses NoC
  26 best.model <- model.main.pca.cumulative
  27
  28 pca_features <- readRDS("data/ores_pca_features_revisions.RDS")
  29 pca_features_unweighted <- readRDS("data/ores_pca_features.noweights.RDS")
  30
  31 test.df <- readRDS("data/holdout_quality_labels.RDS")
  32
  33 wpca_transform <- function(wpca, x){
  34     x <- as.matrix(x)
  35     centered <- as.matrix(t(t(x) - wpca$means))
  36     return(centered %*% wpca$basis)
  37 }
  38
  39 new_pca_features <- wpca_transform(pca_features, test.df[,.(Stub, Start, C, B, GA, FA)])
  40
  41 test.df<-test.df[,":="(pca1.revision = new_pca_features[,1],
  42                        pca2.revision = new_pca_features[,2],
  43                        pca3.revision = new_pca_features[,3],
  44                        pca4.revision = new_pca_features[,4],
  45                        pca5.revision = new_pca_features[,5])]
  46
  47 draws <- as.data.table(best.model)
  48
  49 test.df <- test.df[,idx.max:=.(apply(test.df[,.(Stub,Start,C,B,GA,FA)],1,which.max))]
  50 test.df <- test.df[,MPQC:=.(apply(test.df[,.(idx.max)],1,function(idx) c("stub","start","c","b","ga","fa")[idx]))]
  51 top.preds <- test.df[,MPQC]
  52
  53 #ordinal.fitted.1 <- fitted(best.model, test.df, scale='response')
  54 ordinal.fitted <- data.table(fitted(best.model, test.df, scale='linear'))
  55 remember(ordinal.fitted,'ordinal.fitted')
  56 ordinal.pred <- ordinal.fitted$Estimate
  57
  58 quality.ordinal <- ordinal.pred
  59 quality.even6 <- apply(test.df[,.(Stub,Start,B,C,GA,FA)],1,function(r) r %*% c(0,1,2,3,4,5))
  60 quality.even5 <- apply(test.df[,.(Stub,Start,B,GA,FA)],1,function(r) r %*% c(1,2,3,4,5))
  61
  62 test.df <- test.df[,quality.ordinal := quality.ordinal]
  63 test.df <- test.df[,quality.even6 := quality.even6]
  64
  65 (spearcor <- weightedCorr(test.df$quality.ordinal, test.df$quality.even6, method='spearman', weights=test.df$revision_weight))
  66 remember(spearcor, 'spearman.corr')
  67 (pearsoncor <- weightedCorr(test.df$quality.ordinal, test.df$quality.even6, method='pearson', weights=test.df$revision_weight))
  68 remember(pearsoncor, 'pearson.corr')
  69
  70 ordinal.preds <- data.table(predict(best.model, test.df, robust=F))
  71 #names(ordinal.preds) <- c("Stub","Start","C","B","A","GA","FA")
  72 names(ordinal.preds) <- c("Stub","Start","C","B","GA","FA")
  73 ordinal.preds <- ordinal.preds[,idx.max:=.(apply(ordinal.preds[,.(Stub,Start,C,B,GA,FA)],1,which.max))]
  74 #ordinal.preds <- ordinal.preds[,predicted:=.(apply(ordinal.preds[,.(idx.max)],1,function(idx) c("stub","start","c","b",'a',"ga","fa")[idx]))]
  75 ordinal.preds <- ordinal.preds[,predicted:=.(apply(ordinal.preds[,.(idx.max)],1,function(idx) c("stub","start","c","b","ga","fa")[idx]))]
  76 pred.qe6 <- data.table(predict(model.qe6,test.df))
  77 names(pred.qe6) <- c("Stub","Start","C","B","GA","FA")
  78 pred.qe6 <- pred.qe6[,idx.max:=.(apply(pred.qe6[,.(Stub,Start,C,B,GA,FA)],1,which.max))]
  79 #pred.qe6 <- pred.qe6[,predicted:=.(apply(pred.qe6[,.(idx.max)],1,function(idx) c("stub","start","c","b",'a',"ga","fa")[idx]))]
  80 pred.qe6 <- pred.qe6[,predicted:=.(apply(pred.qe6[,.(idx.max)],1,function(idx) c("stub","start","c","b","ga","fa")[idx]))]
  81
  82 test.df <- test.df[,ordinal.pred := ordinal.preds$predicted]
  83 test.df <- test.df[,pred.qe6 := pred.qe6$predicted]
  84 test.df <- test.df[,idx.max:=.(apply(test.df[,.(Stub,Start,C,B,GA,FA)],1,which.max))]
  85 test.df <- test.df[,MPQC:=.(apply(test.df[,.(idx.max)],1,function(idx) c("stub","start","c","b","ga","fa")[idx]))]
  86
  87 (top.pred.accuracy <- weighted.mean(test.df[,.(MPQC)] == test.df[,.(wp10)],test.df$revision_weight))
  88 remember(top.pred.accuracy, "top.pred.accuracy")
  89 (ordinal.pred.accuracy <- weighted.mean(test.df[,.(ordinal.pred)] == test.df[,.(wp10)], test.df$revision_weight))
  90 remember(ordinal.pred.accuracy, "ordinal.pred.accuracy")
  91 quality.even6 <- apply(df[,.(Stub,Start,B,C,GA,FA)],1,function(r) r %*% c(1,2,3,4,5,6))
  92 (pred.qe6.accuracy <- weighted.mean(test.df[,.(pred.qe6)] == test.df[,.(wp10)], test.df$revision_weight))
  93 remember(ordinal.pred.accuracy, "ordinal.pred.accuracy")
  94 remember(best.model, "best.model")
  95
  96 remember(test.df,'test.df')
  97
  98
  99 ordinal.preds[,wp10:=test.df$wp10]
 100 ordinal.preds[,weight:=test.df$article_weight]
 101 total.weight <- sum(ordinal.preds$weight)
 102 library(modi)
 103 calibration.stats.1 <- ordinal.preds[,.(prob.predicted=apply(.SD,2,function(c) weighted.mean(c,weight)),
 104                                       var.predicted=apply(.SD,2,function(c) weighted.var(c,weight))),.SDcols=c("Stub","Start","C","B","GA","FA")]
 105
 106 calibration.stats.1[,wp10:=c("stub","start","c","b","ga","fa")]
 107 calip.data = ordinal.preds[order(wp10),.(prob.data=sum(weight)/total.weight,
 108                                          var.data=var(weight)/total.weight),by=.(wp10)]
 109
 110 calibration.stats.1 <- calibration.stats.1[calip.data,on=.(wp10)]
 111
 112 calibration.stats.1$weighttype <- 'Article weight'
 113
 114 ordinal.preds[,weight:=test.df$revision_weight]
 115 total.weight <- sum(ordinal.preds$weight)
 116
 117 calibration.stats.2 <- ordinal.preds[,.(prob.predicted=apply(.SD,2,function(c) weighted.mean(c,weight)),
 118                                         var.predicted=apply(.SD,2,function(c) weighted.var(c,weight))),.SDcols=c("Stub","Start","C","B","GA","FA")]
 119
 120
 121 calibration.stats.2[,wp10:=c("stub","start","c","b","ga","fa")]
 122 calip.data = ordinal.preds[order(wp10),.(prob.data=sum(weight)/total.weight,
 123                                          var.data=var(weight)/total.weight),by=.(wp10)]
 124
 125 calibration.stats.2 <- calibration.stats.2[calip.data,on=.(wp10)]
 126
 127 calibration.stats.2$weighttype <- 'Revision weight'
 128
 129
 130 ordinal.preds[,weight:=rep(1,nrow(ordinal.preds))]
 131 total.weight <- sum(ordinal.preds$weight)
 132
 133 calibration.stats.3 <- ordinal.preds[,.(prob.predicted=apply(.SD,2,function(c) weighted.mean(c,weight)),
 134                                         var.predicted=apply(.SD,2,function(c) weighted.var(c,weight))),.SDcols=c("Stub","Start","C","B","GA","FA")]
 135
 136
 137 calibration.stats.3[,wp10:=c("stub","start","c","b","ga","fa")]
 138 calip.data = ordinal.preds[order(wp10),.(prob.data=sum(weight)/total.weight,
 139                                          var.data=var(weight)/total.weight),by=.(wp10)]
 140
 141 calibration.stats.3 <- calibration.stats.3[calip.data,on=.(wp10)]
 142
 143 calibration.stats.3$weighttype <- 'No weight'
 144
 145 calibration.stats <- rbind(calibration.stats.1,calibration.stats.2,calibration.stats.3)
 146
 147 calibration.stats[,'calibration':=prob.data - prob.predicted]
 148
 149 remember(calibration.stats, "calibration.stats")
 150
 151
 152 ## p <- ggplot(data.frame(quality.ordinal, quality.even6, quality.even5))
 153 ## p <- p + geom_point(aes(x=quality.even6,y=quality.ordinal)) + geom_smooth(aes(x=quality.even6,y=quality.ordinal))
 154
 155 ## print(p)
 156 ## dev.off()
 157
 158 ## post.pred <- posterior_predict(model.main)
 159 ## preds <- as.character(predict(polrmodel))
 160 ## polrmodel.accuracy <- weighted.mean(preds==df$wp10,df$weight)