Messing around with predicting college major from vocational interest data.

twitter_vi_prediction.R

# Multinomial Logit Model vs Random Forest
# Predicting College Major from Items
library(data.table) # For fread function (v fast!)
library(dplyr) # For pre-processing
library(tm) # For cleaning text / pre-processing
library(nnet) #for MNL
library(randomForest) #For rf

# Helper function for making training / testing sets.
outersect <- function(x, y) {
  sort(c(setdiff(x, y),
         setdiff(y, x)))
}

# Normalizing majors (entered as free response data)
majors <- data.table::fread(input = paste(getwd(),'/data.csv',sep=''), sep = '\t',stringsAsFactors = F)$major %>%
  toupper() %>% 
  trimws() %>% 
  removePunctuation() %>% 
  removeNumbers() %>% 
  table() %>% 
  melt() %>%
  arrange(-value) %>% 
  .[-1,]

# Final data set
vi_data <- data.table::fread(input = paste(getwd(),'/data.csv',sep=''), sep = '\t',stringsAsFactors = F) %>% 
  mutate(MAJOR = toupper(major)) %>% 
  select(-major) %>% 
  filter(MAJOR %in% as.character(majors$.[2:10])) %>% 
  filter(!is.na(MAJOR)) %>% 
  select(R1:C8, MAJOR) %>% 
  mutate(R = rowSums(select(., grep("R[0-9]", names(.)))),
         I = rowSums(select(., grep("I[0-9]", names(.)))),
         A = rowSums(select(., grep("A[0-9]", names(.)))),
         S = rowSums(select(., grep("S[0-9]", names(.)))),
         E = rowSums(select(., grep("E[0-9]", names(.)))),
         C = rowSums(select(., grep("C[0-9]", names(.))))) %>% 
  mutate(MAJOR = as.factor(MAJOR)) %>% 
  select(R,I,A,S,E,C,MAJOR)

test_rows  <- sample(1:nrow(vi_data), size = 5000, replace = F)
train_rows <- outersect(x = 1:nrow(vi_data), y = test_rows)

mnl_vi <- multinom(MAJOR ~ R+I+A+S+E+C, data=vi_data[train_rows,])

rf_vi <- randomForest::randomForest(MAJOR ~ ., data=vi_data[train_rows,])

list(
`Multi-Nomial Logit` = data.frame(vi_data[test_rows,], 
           prediction = predict(mnl_vi, vi_data[test_rows,-ncol(vi_data)])) %>% 
  select(MAJOR, prediction) %>% 
  table(.)
,
`Random Forest` = data.frame(vi_data[test_rows,], 
           prediction = predict(rf_vi, vi_data[test_rows,-ncol(vi_data)])) %>% 
  select(MAJOR, prediction) %>% 
  table(.)
)