In drake-smu/unintCS02:

knitr::opts_chunk$set(echo = TRUE)

Executive Summary

Problem and Motivation

The cost replacing an entry-level employee is 50% of that employee's salary, and for a supervisory role, that percentage increases as high as 150%! Within this notebook we explore the most predictive variables that lead to attrition and retain talent within DDSAnalystics Talent Management dataset. At a high level, we found that the four most predictive factors that lead to attrition were: over time, monthly income, job role and business travel. The 4 most predictive factors that retain talent are: work life balance, environment satisfaction, stock options, compensation department ratio.

Variable Explination

The explanatory variable overtime in the data set is stored as a boolean categorical type with “Yes” and “No” values. The variable monthly income is provided as a raw continuous field that was bucketed into bins in order to provide more predictive power. The buckets created contain 4 levels: 0-5000, 5000-10000, 10000-20000 and 20000-max. Job role is defined in the dataset as a categorical variable with 9 levels. These 9 levels range from roles like: “Laboratory Technician” to “Sales Executive”. Business travel is a categorical variable that contains 3 levels: Frequent, Rarely, and None – thus the employee travels often on behalf of their role, rarely on behalf of their role, or never. The work life balance variable is a numerical factor in which 1 represents “Bad Work Life Balance” and 4 represents “Best Work Life Balance”. Environment satisfaction is a numerical factor with variable with 4 levels: “Low Environment Satisfaction", "Medium Environment Satisfaction", "High Environment Satisfaction" , “Very High Environment Satisfaction”. Stock option benefits for employees are represented in the dataset as a numerical factor with 4 levels: 0,1,2,3. The 0 refers to no stock options and the 3 refers to the greatest stock options. Lastly, the compensation department ratio variable is a engineered feature calculation that takes the raw monthly income of the employee and divides it by the mean monthly income of everyone in that department.

Results and Outcome

While over time, monthly income, job role, business travel lead to attrition and work life balance, employment satisfaction and stock options retain talent, we found that these specific variables only lead to an explainability of about 30%. In order to conduct this analysis, we built over 50 different models for EDA, ranging from simple linear regressions to decision trees and random forests. We created over 14 engineered features.

In conclusion, we found that employee retention is a difficult and complex problem to solve! However, there is rich and fruitful insights that can be gained from leveraging this dataset. In addition, we believe building and productionalizing an employee retention model would yield promising results and opens up opportunities for great savings. In order to improve model explainability, more data would be extremely beneficial. We also believe there is a lot of room for improvement with the current dataset.

Next Steps

For variable selection and creation, with 34 explanatory variables, 2 of which contained no variability – we believe that more explanatory variables could better represent the complex environment of employee attrition. Some potential variables include: time series employee and company data. This would allow for the ability to track changes within an employee’s lifecycle and life events at the company. An excellent practical example of this would be big life events that often evoke moving: marriage and death. In addition, detailed company demographic variables like office location, organizational changes, news and press articles could show potential predictive power. Delving more into detailed employee demographics like health benefits, and remote working opportunities could help extrapolate insight into why employees leave (and stay!). Other useful information that could prove important is any competitor information and employee feedback forms!

In terms of model selection and data exploration techniques, there are always opportunities for more advanced dimensionality reduction techniques like KNN and PCA. We could also do more feature engineering with respect to oversampling techniques. We also would like to explore insights given by building using nueral networks.

Introduction

This github projct is a complilation of extensive exploratory data analysis regarding DDSAnalytics employee retention data. Below you will find techniques accumlicated over 15 weeks in Southern Methodist University Masters in Data Science program MSDS6306 Doing Data Science class. We were provided a single excel spread sheet with attrition data and asked to glean insights leveraging all the techniques covered throughout the semester. The goal was to create a final presentation in which we would showcase the insights we found. In addition to the presentation, we were required to create an R markdown file that would allow for reproducibility in the event DDSAnalytics would like to build, or leverage any of our findings in the future. You will find an executive summary providing technical insight into the results, as well as the code used to generate said insights. Enjoy!

Analysis

library(coefplot)
library(ggthemes)
library(ggplot2)
library(grid)
library(gridExtra)
library(knitr)
library(pander)
library(randomForest)
library(readxl)
library(tidyverse)


source("R/feature-engineering.R")
source("R/model2r.R")
source("R/model2r2.R")
source("R/forests.R")
source("R/forests2.R")
read_chunk("R/feature-engineering.R");
read_chunk("R/forests.R");
read_chunk("R/linear-model1.R");
read_chunk("R/model2R.R");
read_chunk("R/model2R2.R");

Process Data

First we need to import the client excel data and convert it into more accessible R objects.

Now that we have the data loaded we can evaluate the classes, features, and cleanness.

raw_data_df <- readRDS("data/hr-data.rds");
raw_definitions_df <- readRDS("data/hr-definitions.rds")

# Raw Data Structures
str(raw_data_df);
str(raw_definitions_df)

#Reading in hr data file
raw<-readRDS("data/hr-data.rds")
pander(head(raw))

#Turning Variables into Factors
names <- c('WorkLifeBalance' ,'StockOptionLevel','PerformanceRating','JobSatisfaction',
           'RelationshipSatisfaction','JobLevel','JobInvolvement','EnvironmentSatisfaction','Education')
raw[,names] <- lapply(raw[,names] , factor)

#Looking at dataset structure
pander(str(raw))

Data Investigation

Missing Value Analysis

na_count <- data.frame(sapply(raw, function(y) sum(length(which(is.na(y))))))
pander(na_count)

Initial Modeling

Coefficients

cplo2

Random Forest

varImpPlot(VariableImportance)

Reduced Linear Regression

Feature Engineering

Age Groups

Distance Groupings

Years With Mangager

Total Years Worked

Various Income Groupings

str(engr_df)

Updated Models

Updated Coefficients

cplot3

Updated Random Forest

varImpPlot(VariableImportance2)

Updated Linear Regression

drake-smu/unintCS02 documentation built on May 31, 2019, 5:40 a.m.