(SANDBOX) LPA Final Project

Experiment setup and predictive modeling for AnCo

Purpose: This notebook [Explain why we're doing this]

Data source: https://www.kaggle.com/akbaldawa/hr-analytics?select=Train.csvVB

Actually is from this challenge, which we will use for our test model verification: https://datahack.analyticsvidhya.com/contest/wns-analytics-hackathon-2018-1/

Section 1: Load data and packages

#Read data df_train <- read.csv("Train.csv")

#Install Libraries (to be extended) #install.packages("tidyverse") library(tidyverse) library(jtools) library(stargazer) library(sjPlot) #Set plot size options(repr.plot.width=6, repr.plot.height=5)

── Attaching packages ─────────────────────────────────────── tidyverse 1.3.0 ──

✔ ggplot2 3.3.3     ✔ purrr   0.3.4
✔ tibble  3.0.5     ✔ dplyr   1.0.3
✔ tidyr   1.1.2     ✔ stringr 1.4.0
✔ readr   1.4.0     ✔ forcats 0.5.1

── Conflicts ────────────────────────────────────────── tidyverse_conflicts() ──
✖ dplyr::filter() masks stats::filter()
✖ dplyr::lag()    masks stats::lag()


Please cite as: 


 Hlavac, Marek (2018). stargazer: Well-Formatted Regression and Summary Statistics Tables.

 R package version 5.2.2. https://CRAN.R-project.org/package=stargazer

Section 2: Prepare data

#let's have an initial look head(df_train)

# Several variables are stored as characters (<chr>) but will actually act as factors # (department, region, education, gender, recruitment_channel, previous_year_rating should all be factors) # we should have a look at the observations, and then convert them into factors df_train %>% count(department) df_train %>% count(region) df_train %>% count(education) df_train %>% count(gender) df_train %>% count(recruitment_channel) df_train %>% count(previous_year_rating)

#seems like there are som NA's in the "previous year rating" column. These are likely to be new employees #let's see how many there are table(is.na(df_train$previous_year_rating))

#only 8% of observations lack this data #intuitively we probably are not promoting people within their first year #it makes sense to delete these observations df_train <- filter(df_train, previous_year_rating >= 1)

#make sure the NA rows are now deleted table(is.na(df_train$previous_year_rating))

#similarly, 2409 have no education on file. let's remove these as well and make sure the rows are deleted df_train <- filter(df_train, education != "") df_train <- filter(df_train, is.na(education)==FALSE)

#now we can transform them into factors #department df_train$department <- parse_factor(df_train$department, levels=NULL, ordered=FALSE) class(df_train$department)

#region df_train$region <- parse_factor(df_train$region, levels=NULL, ordered=FALSE) class(df_train$region)

#gender df_train$gender <- parse_factor(df_train$gender, levels=NULL, ordered=FALSE) class(df_train$gender)

#recruitment channel df_train$recruitment_channel <- parse_factor(df_train$recruitment_channel, levels=NULL, ordered=FALSE) class(df_train$recruitment_channel)

#education (we will sort) ed_levels <- c("Below Secondary","Bachelor's","Master's & above") df_train$education <- parse_factor(df_train$education, levels=ed_levels, ordered=TRUE) class(df_train$education)

#previous_year_rating (we will sort) - different formula as the previous works for characters only rating_levels <- c("1","2","3","4","5") df_train$previous_year_rating <- as.factor(df_train$previous_year_rating) df_train$previous_year_rating <- factor(df_train$previous_year_rating, levels=rating_levels, ordered = TRUE) class(df_train$previous_year_rating)

#few more transformations needed #employee_id should be character and not int, as the numbers do not represent values; df_train$employee_id <- as.character(df_train$employee_id) #is_promoted,awards_won,KPIs_met should be booleans, as they are binary options df_train$is_promoted <- as.logical(df_train$is_promoted) df_train$awards_won. <- as.logical(df_train$awards_won.) df_train$KPIs_met..80. <- as.logical(df_train$KPIs_met..80.)

Section 3: Visualize and understand data

Section 4: Running the models

#our initial hypothesis is that promotion decisions could be linked to previous ratings, meeting KPIs, average training scores, and education levels #let's test a model with these predictors model_berten = glm(is_promoted ~ education + previous_year_rating + KPIs_met..80. + avg_training_score, data = df_train)

summary(model_berten)

model_cem_step1 = glm(is_promoted ~ . - employee_id, data = df_train) model_cem_step2 = step(model_cem_step1, direction = "backward", trace = FALSE) summary(model_cem_step2)

# we will need some new libraries if we want to try machine learning algorithms install.packages("randomForest") library(rpart.plot) library(lubridate) library(randomForest)

Installing package into ‘/work/.R/library’
(as ‘lib’ is unspecified)

Warning message in install.packages("randomForest"):
“installation of package ‘randomForest’ had non-zero exit status”
Loading required package: rpart


Attaching package: ‘lubridate’


The following objects are masked from ‘package:base’:

    date, intersect, setdiff, union


randomForest 4.6-14

Type rfNews() to see new features/changes/bug fixes.


Attaching package: ‘randomForest’


The following object is masked from ‘package:dplyr’:

    combine


The following object is masked from ‘package:ggplot2’:

    margin

#first we need to turn the outcome variable into a factor df_train$is_promoted = as.factor(df_train$is_promoted)

#now we can fit the random forest model (m=4 as rule of thumb of m= variables/3) model_rf = randomForest(is_promoted ~ department + region + education + gender + recruitment_channel + no_of_trainings + age + previous_year_rating + length_of_service + KPIs_met..80. + awards_won. + avg_training_score, data = df_train, importance = TRUE, mtry = 4)

model_rf

#now we can check the importance assigned by the algorithm to each variable varImpPlot(model_rf, type=1)