Retaining employees using data analytics¶
Classification and regression trees will be applied to a simulated HR dataset.
The business case is formed around the question: Can we predict those employees who are likely to leave the organization?¶
To answer this question, we will:
- Perform any initial data preparations.
- Split the data into training and validation.
- Develop an initial model.
- Which are the most important variables.
- Report on the accuracy of the model.
- Interpret two (2) complete paths.
This is a simulated dataset that uses several measures, such as customer satisfaction ratings, employee evaluation, average number of projects and so forth, to predict which employees are at a risk to leave the company.¶
Imports¶
In [1]:
# install and load libraries
# for data import
library(readr)
library(readr)
# for data wrangling
library(tidyverse)
library(tidyr)
# for visualization
library(gmodels)
library(ggplot2)
library(reshape2)
library(scales)
library(ggthemes)
library(ggthemes)
library(dplyr)
# for building classification and regression trees
library(rpart)
library(rpart.plot)
Explore¶
In [2]:
# load data into dataframe
HR_comma_sep <- read_csv("/home/jovyan/HR Analytics+CART/HR_comma_sep.csv")
In [3]:
# first rows of dataset
head(HR_comma_sep, 10)
In [4]:
# visualizing continous variables
theme_set(theme_gray(base_size = 10))
d <- melt(HR_comma_sep)
ggplot(d,aes(x = value)) +
facet_wrap(~variable,scales = "free_x") +
geom_histogram(bins = 10, fill = "lightblue", colour = "darkblue", size = .1) +
scale_y_continuous( name = "Number of employees" ) +
scale_x_continuous( name = "Frequency" )
- Satisfaction level: Most employees are highly satisfied.
- Last Evaluation: Most employees are good performers with 75% of the data set being evaluated between 56 percent - 87 percent.
- Number of Projects: Most employees do a reasonable number of projects.
- Average Monthly Hours: Most employees spend, fairly, a higher number of hours at work.
- Time Spent in Company: Fewer employees stay beyond 4 years.
In [5]:
# second glance at the binary continous variable: Work_accident
wa <- ggplot(data=HR_comma_sep, aes(x=Work_accident, fill=as.factor(Work_accident))) +
scale_x_continuous(breaks=0:1,
labels = c("No accident" , "Yes accident")) +
geom_bar(width=0.5,
(aes(y = (..count..)/sum(..count..))))+
scale_y_continuous(labels = scales::percent) +
geom_text(aes(label = scales::percent((..count..)/sum(..count..)),
y= (..count..)/sum(..count..)), stat= "count", vjust =.5, hjust= 1.2, size=3, color='white') +
labs(title = "Frequency of accidents at work",
y = "Percentage of employees") +
theme(aspect.ratio = .3) +
coord_flip()
wa
- Most employees did not have accidents at work.
In [6]:
# second glance at the binary continous variable: resigned
re <- ggplot(data=HR_comma_sep, aes(x = resigned , fill=as.factor(resigned))) +
scale_x_continuous(breaks=0:1,
labels = c("Stayed" , "Resigned")) +
geom_bar(width=0.5,
(aes(y = (..count..)/sum(..count..))))+
scale_y_continuous(labels = scales::percent) +
geom_text(aes(label = scales::percent((..count..)/sum(..count..)),
y= (..count..)/sum(..count..)), stat= "count", vjust =.5, hjust= 1.2, size=3, color='white') +
labs(title = "Frequency of resignations",
y = "Percentage of employees") +
theme(aspect.ratio = .3) +
coord_flip()
re
- Most employees stayed with the organization and did not leave.
In [7]:
# second glance at the binary continous variable: promotion in last 5 years
pro <- ggplot(data=HR_comma_sep, aes(x = promotion_last_5years , fill=as.factor(promotion_last_5years))) +
scale_x_continuous(breaks=0:1,
labels = c("No promotion" , "Yes promotion")) +
geom_bar(width=0.5,
(aes(y = (..count..)/sum(..count..))))+
scale_y_continuous(labels = scales::percent) +
geom_text(aes(label = scales::percent((..count..)/sum(..count..)),
y= (..count..)/sum(..count..)), stat= "count", vjust =.5, hjust= .7, size=2, color='black') +
labs(title = "Frequency of promotions in last 5 years",
y = "Percentage of employees") +
theme(aspect.ratio = .3) +
coord_flip()
pro
- Most people have not recieved a promotion in the last five years.
In [8]:
# visualizing categorical variables
# salary_grade
sg <- ggplot(data=HR_comma_sep, aes(x = salary_grade , fill=as.factor(salary_grade))) +
geom_bar(width=0.9,
(aes(y = (..count..)/sum(..count..))))+
scale_y_continuous(labels = scales::percent) +
geom_text(aes(label = scales::percent((..count..)/sum(..count..)),
y= (..count..)/sum(..count..)), stat= "count", vjust =.5, hjust= 1.2, size=3, color='white') +
labs(title = "Salary grade of employees",
y = "Percentage of employees") +
theme(aspect.ratio = .4) +
coord_flip()
sg
In [9]:
# department
dp <- ggplot(data=HR_comma_sep, aes(x = department , fill=as.factor(department))) +
geom_bar(width=0.5,
(aes(y = (..count..)/sum(..count..))))+
scale_y_continuous(labels = scales::percent) +
geom_text(aes(label = scales::percent((..count..)/sum(..count..)),
y= (..count..)/sum(..count..)), stat= "count", vjust = 0.5, hjust= 1.5, size=3, color='white') +
labs(title = "Departments of the organization",
y = "Percentage of employees") +
theme(aspect.ratio = 1.2) +
coord_flip()
dp
In [10]:
# Creating factors of categorical variables and binary, continous variable resigned
hr <- HR_comma_sep %>%
mutate(salary_grade = as.factor(salary_grade) ,
department = as.factor(department),
resigned = as.factor(resigned),
random = runif(14999))
In [11]:
# rename factor levels of variable 'resigned'
levels(hr$resigned) <- c("stayed", "resigned")
In [12]:
print((summary(hr$resigned)))
In [13]:
print((summary(hr$salary_grade)))
In [14]:
print((summary(hr$department)))
Splitting the data into training and validation sets:¶
In [15]:
set.seed(123)
train <- hr %>%
filter(random < .7) %>%
select(-random)
val <- hr %>%
filter(random >= .7) %>%
select(-random)
In [16]:
ct1 <- rpart(resigned ~ . , data = train, method = 'class')
In [17]:
# plotting the model
rpart.plot(ct1)
In [18]:
# complexity parameter table
ct1$cptable
Interpreting Two Complete Paths:¶
- Path 1: Will Not Leave (Stayed/loyal)
- First condition: satisfaction_level >= 47 percent.
- Second condition: time_spend_company < 5 years.
- Third condition: last_evaluation < 81 percent.
Hence, those who stayed are highly satisfied, have spent at least 4 years in the organization, and are good performers with an evaluation of at least 80 percent.¶
Path 2: Will Leave (resign)
- First condition: satisfaction_level < 47 percent.
- Second condition: number_project >= 3 projects.
- Third condition: last_evaluation >= 58 percent.
Hence, those who leave are lowly or moderately satisfied, have a workload of 3 or more projects with their performance being evaluated at least 58 percent.¶
In [19]:
print(var_importance <- data.frame(ct1$variable.importance))
In [20]:
# how good is our initial model(ct1)?
val$resign_predicted <- predict(ct1, val, type = 'class')
In [21]:
print(summary(val))
In [22]:
# rename factor levels of variable 'resigned_predicted'
levels(val$resign_predicted) <- c("predicted_stay", "predicted_resign")
print(summary(val$resign_predicted))
In [23]:
# Crosstable of actual resignations versus predicted resignation values by initial model
CrossTable(val$resigned , val$resign_predicted)
In [24]:
# creating a second model for comparision
ct2 <- rpart(resigned ~ . ,
data = train ,
method = 'class' ,
cp = .052)
rpart.plot(ct2)
In [25]:
### how good is our second model(ct2)?
val$resign_predicted2 <- predict(ct2, val, type = 'class')
print(summary(val))
In [26]:
# Crosstable of actual resignations versus predicted resignation values by second model
CrossTable(val$resigned , val$resign_predicted2)