Ashley O'Mahony | ashleyomahony.com | February 2019
This case study aims to model the probability of attrition of each employee from the HR Analytics Dataset, available on Kaggle. Its conclusions will allow the management to understand which factors urge the employees to leave the company and which changes should be made to avoid their departure.
All the files of this project are saved in a GitHub repository.
The libraries used for this project include: pandas and numpy for data manipulation, matplotlib.pyplot and seaborn for plotting, spicy for preprocessing, and scikit-learn for Machine Learning.
The dataset is stored in the GitHub repository as a CSV file: turnover.csv. The file is loaded directly from the repository.
The first stage of this analysis is to describe the dataset, understand the meaning of each variable, detect possible patterns and perform the necessary adjustments to ensure that the data will be proceeded correctly during the Machine Learning process.
The dataset consists in 14,999 rows and 10 columns. Each row represents an employee, and each column contains one employee attribute. None of these attributes contains any NA.
| Variable | Type | Range | Definition |
|---|---|---|---|
| satisfaction_level | Float | 0 to 1 | Employee satisfaction level. |
| last_evaluation | Float | 0 to 1 | Employee last evaluation score. |
| number_project | Integer | 2 to 7 | Number of projects handled by the employee. |
| average_montly_hours | Integer | 96 to 310 | Average monthly hours worked by the employee. |
| time_spend_company | Integer | 2 to 10 | Number of years spent in the company by the employee. |
| Work_accident | Boolean | 0 or 1 | Flag indicating if the employee had a work accident. |
| Left | Boolean | 0 or 1 | Flag indicating if the employee has left the company. This is the target variable of the study, the one to be modelled. |
| promotion_last_5years | Boolean | 0 or 1 | Flag indicating if the employee has been promoted within the past 5 years. |
| department | Categorical | 10 values | Initially sales, renamed as department. Department of the employee: Sales, Accounting, HR, Technical, Support, Management, IT, Product Management, Marketing, R&D. |
| salary | Categorical | 3 values | Salary level of the employee: Low, Medium, High. |
Figure 1: Variables of the HR Analytics Dataset
The objective of this study is to build a model to predict the value of the variable left, based on the other variables available. A first inspection reveals that 23.8% of the employee listed in the dataset have left the company. The dataset is not balanced, which might introduce some bias in the predictive model. The Synthetic Minority Oversampling Technique (SMOTE) has been used at the end of the study to compare the model with another one developed from an over-sampled dataset.
A closer look to the means of the variables allow to highlight the differences between the employees who left the company and those who stayed. Employees who left the company have:
- a lower satisfaction level: 0.44 vs 0.67.
- higher average monthly working hours: 207 vs 199.
- a lower work accident ratio: 0.05 vs 0.18.
- a lower promotion rate in the past 5 years: 0.01 vs 0.03.
The salary level seems to have a great impact on the employee turnover, as higher salaries tend to stay in the company (7% of turnover), whereas lower salaries tend to leave the company (30% of turnover). Departments, even with different turnover rates, don’t seem to have a significant impact on the employee departure.
Employees with very low satisfaction level (below 0.12) obviously leave the company. A risky zone is when employees rates their satisfaction just below 0.5 (between 0.36 and 0.46). Employees also tend to leave the company when they become moderately satisfied (between 0.72 and 0.92).
Employees with low evaluation scores tend to leave the company (between 0.45 and 0.57). A large number of good employees (scores higher than 0.77) leave the company, maybe to get a better opportunity. Interestingly, the ones with very low scores seem to stay.
Employees with really low numbers of hours per month (below 125) tend to stay in the company, whereas employees working too many hours (above 275 hours) have a high probability to leave the company. A safe range is between 161 and 217 hours, which seems to be ideal to keep employees in the company.
The main observation regarding the number of projects is that employees with only 2 or more than 5 projects have a higher probability to leave the company. It also seems that employees with 3-6 years of services are leaving the company. Employees with a work accident tend to stay in the company. And employees with a promotion within the past 5 years have less propensity to leave the company.
No strong correlation appears in the dataset. However, it is possible to see clear groups when looking at the relationships of pairs of variables: Number of Projects vs Average Monthly Hours, Number of Projects vs Last Evaluation, Last Evaluation vs Average Monthly Hours, Last Evaluation vs Satisfaction.
| 
|
| 
|
Figure 2: Bar Plots of interesting pairs of variables, highlighting possible groups
For the model to proceed with the data efficiently, the categorical variables salary and department have been encoded. As the values of salary have an order, they have been encoded into integers within the same variable. For department, as the values have no specific order, they have been encoded into individual variables with boolean values. Thus, the dataset has been transformed from 10 variables to 19 variables. Numerical variables average_monthly_hours, last_evaluation and satisfaction_level are scaled between 0 and 1 to remove any influence of their difference in value ranges on the model. They have also been checked for skewness, without a real change on their shape.
The dataset will be split randomly into Train and Test sets, with ratio 70|30. This method will be used at each step of the feature engineering, before the modelling steps.
A logistic regression algorithm will be used to develop this classification model. The baseline results of the model return an accuracy of 0.797, which is acceptable. However, the results regarding employees who left the company - our main objective - aren’t so satisfactory, as they present a very low recall of 0.34, which means that only 34% of the employees who left the company were detected. These results should improve significantly after the Feature Engineering phase for the model to be satisfactory.
| 
|
Figure 3: Baseline Results
The model will be cross-validated using a 10-fold cross validation method returning the average accuracy. This method will be applied at every modelling step, to ensure that the model is not biased by the training set split.
In order to improve the model results, a set of features will be created and modified to describe more accurately the characteristics and patterns of the data.
Based on the EDA, the Satisfactory Level is binned and one hot encoded in 6 bins: (0.00, 0.11], (0.11, 0.35] , (0.35, 0.46] , (0.46, 0.71] , (0.71, 0.92] , (0.92, 1.00]. The new feature is then one hot encoded. This step increases the accuracy of the model to 0.914. The feature is accepted.
| 
|
Figure 4: Satisfaction Level Bar Plot, Before and After Binning
Based on the EDA, the Last Evaluation is binned and one hot encoded in 4 bins: (0.00, 0.44], (0.44, 0.57] , (0.57, 0.76] , (0.76, 1.00]. The new feature is then one hot encoded. This step increases the accuracy of the model to 0.936. The feature is accepted.
| 
|
Figure 5: Last Evaluation Bar Plot, Before and After Binning
Based on the EDA, the Average Monthly Hours is binned and one hot encoded in 7 bins: (0, 125], (125, 131] , (131, 161] , (161, 216] , (216, 274] , (274, 287] , (287, 310]. The new feature is then one hot encoded. This step increases the accuracy of the model to 0.945. The feature is accepted.
| 
|
Figure 6: Average Monthly Hours Bar Plot, Before and After Binning
Based on the EDA, the Number of Projects can be categorized into 4 categories: too low, normal, too high, extreme. The new feature is then one hot encoded. The step increases the accuracy of the model to 0.950. The feature is accepted.
Based on the EDA, the Time Spent in Company can be categorized into 4 categories, related to the rate of departure: no departure, low departure, high departure, very high departure. The new feature is then one hot encoded. The step increases the accuracy of the model to 0.956. The feature is accepted.
Based on the EDA, the employees can be cluster by Workload, based on the Number of Projects and Average Monthly Hours, into 5 categories: very low, low, normal, high, extreme. The new feature is then one hot encoded. The step increases the accuracy of the model to 0.959. The feature is accepted.
| 
|
Figure 7: Number of Projects by Average Monthly Hours Scatter Plot, Before and After Clustering
Based on the EDA, the employees can be cluster by Project Performance, based on the Number of Projects and Last Evaluation, into 4 categories: very low, low, normal, high. The new feature is then one hot encoded. The step decreases the accuracy of the model to 0.958, but the 10-fold cross validation average accuracy increases from 0.958 to 0.960. The feature is kept, even if it is not clearly defined if it has an impact on the model accuracy. The Feature Selection phase might later clarify its importance.
| 
|
Figure 8: Number of Projects by Last Evaluation Scatter Plot, Before and After Clustering
Based on the EDA, the employees can be clustered by Efficiency, based on the Last Evaluation and the Average Monthly Hours, into 4 categories: very low, low, normal, high. The new feature is then one hot encoded. The step increases the accuracy of the model to 0.960. The feature is accepted.
| 
|
Figure 9: Last Evaluation by Average Monthly Hours Scatter Plot, Before and After Clustering
Based on the EDA, the employees can be clustered by Attitude, based on the Last Evaluation and the Satisfaction Level, into 7 categories: very unhappy, unhappy, low performance, unhappy and low performance, normal, happy and high performance, very happy. The new feature is then one hot encoded. The step increases the accuracy of the model to 0.964. The feature is accepted.
| 
|
Figure 10: Last Evaluation by Satisfaction Level Scatter Plot, Before and After Clustering
The dataset resulting from the Feature Engineering phase contains 58 features, with a model reaching the accuracy of 0.964. The Feature Selection phase aims to reduce the number of variables used by the model. The Recursive Feature Elimination (RFE) method is used to select the most relevant features for the model. It returns a list of 14 features which should be sufficient to our model.
The final model is tested with the 14 selected features and returns the accuracy of 0.967. The recall for employees who left the company now reaches 88%, which will allow the management to better predict which employees have a high probability to leave.
| 
|
Figure 11: Final Model Results
Analyzing the 14 selected features and their coefficients will help understanding the underlying reasons for an employee to want to stay or to leave the company.
| Features reducing the probability of departure | Features increasing the probability of departure | ||||
|---|---|---|---|---|---|
| Attitude | Normal | -2.936478 | Attitude | Very Unhappy | 2.582233 |
| Attitude | Very Happy | -2.423485 | |||
| Attitude | Unhappy | -2.295986 | |||
| Satisfaction | (0.92, 1.00] | -2.448987 | Satisfaction | (0.00, 0.11] | 2.582233 |
| Workload | Normal | -2.183617 | Workload | Extreme | 2.245120 |
| Efficiency | Very Low | -4.182167 | Efficiency | Low | 3.295257 |
| Time Spent | No Departure | -2.016273 | Time Spent | Very High Departure | 2.403233 |
| Av. Monthly Hours | (287, 310] | 2.245120 | |||
| Number Project | Extreme | 3.987229 | |||
| Intercept. | -0.375952 | ||||
Figure 12: Selected features and corresponding model coefficients
This model will allow the company to calculate the probability of an employee to leave the company and to act on key-factors to avoid departures. The satisfaction of employees and the amount of workload they have to bear seem to be important causes of withdrawals. A particular attention on the work-life balance would be crucial to improve the turnover rate.
The high accuracy is driven by the binned features tailored to the dataset. If they work really well for this data, it might not be the case for another dataset. The features should instead be set using standard binning approach, which wouldn't fit as well the data but which would be adaptable to any dataset. That solution would be recommended if the model has to be run in production.