How does the binomial regression perform on the "Meh" and "Stans only" classes?

[wkumler]

library(tidyverse)

# dataset_version <- "FT350"
dataset_version <- "FT2040"
output_folder <- paste0("made_data_", dataset_version, "/")

features_extracted <- read_csv(paste0(output_folder, "features_extracted.csv")) %>%
  mutate(sn=ifelse(is.infinite(sn), 0, sn))

set.seed(123)
traintestlist <- features_extracted %>%
  # select(-med_SNR, -med_cor, -med_missed_scans, -shape_cor, -area_cor) %>%
  filter(feat_class%in%c("Good", "Bad")) %>%
  mutate(feat_class=ifelse(feat_class=="Good", TRUE, FALSE)) %>%
  slice_sample(n = nrow(.)) %>%
  split(rbernoulli(nrow(.), 0.2)) %>%
  setNames(c("train", "test"))

full_model <- traintestlist$train %>% 
  select(-feat_id, -blank_found) %>%
  glm(formula=feat_class~., family = binomial)
pred_prob_vec <- features_extracted %>%
  select(-blank_found, -feat_id) %>%
  predict(object=full_model, type = "response")

features_extracted %>%
  mutate(pred_prob=cut(pred_prob_vec, pretty(pred_prob_vec, n = 10))) %>%
  ggplot() +
  geom_bar(aes(x=pred_prob, fill=feat_class)) +
  facet_wrap(~feat_class, ncol = 1, scales = "free_y") +
  theme(axis.text.x = element_text(angle=90, hjust=1, vjust=0.5))

Unsurprisingly, we get a much "flatter" distribution among the meh and stans features. It is little surprising to see so much separation in the "stans only" but this also might just be because they're misclassified - sometimes a good peak hides underneath the super-tall standards.