Salifort Motors Attrition Rates

Project Overview

As the capstone for the Google Advanced Data Analytics certificate, this project uses a fictional human-resources dataset from Kaggle to investigate attrition at Salifort Motors.

With recent turnover at Salifort Motors, I was tasked with investigating the drivers of attrition and providing data-driven recommendations using machine-learning models to help mitigate future departures. If Salifort could predict whether an employee will leave the company, and discover the reasons behind their departure, they could better understand the problem and develop a solution. The goal is to understand what makes an employee likely to leave.

Executive Summary

I analyzed a 15k-row human resources dataset for Salifort Motors to understand drivers of attrition. Data visualisation showed associations between key working environment variables and employee departure. I produced an Executive Summary for stakeholder communication to summarize the project's discoveries.

Key Findings

Employees with more projects, higher monthly hours, longer tenure, and lower satisfaction were much more likely to leave. This was first identified through visualisations of associated influencing variables, followed by machine learning models to predict employee attrition.

Key Observations

• Three identifiable clusters of employees who departed Salifort Motors: a Low Satisfaction Cluster (x ≤ 0.12), a Mid Satisfaction Cluster (0.31 ≤ x ≤ 0.48), and a High Satisfaction Cluster (x ≥ 0.70).
• Majority of all Low Satisfaction Level Employees who departed were on 6 or 7 projects.
• Employee satisfaction level has visible associations with the number of projects; strong association with assignment to 6 and 7 different projects and moderate association with assignment to 2 different projects.
• Additional factor tenure has moderate associations with employee satisfaction level for duration of 3 and 4 years.

Prediction Models

Applying Logistic Regression and XGBoost models, I constructed a best-fitting prediction model to assess employee attrition, validating visualised variable associations. Using logistic regression as a baseline and a tuned XGBoost model for comparison.

Impact & Recommendations

Employees working on higher number of projects, a workload with higher monthly hours, longer tenure, lower satisfaction, are much more likely to leave.

Departed employees with the lowest satisfaction levels were assigned to 6 or 7 different projects. Examining all employees, past and present, from data provided, employees assigned to 6 or 7 projects are associated with the highest average monthly hours and are tenured employees around ≈4 years.

Recommendations

Reflecting upon the above visualisations and numbers presented in Project Statistics, this investigation recommends targeting the variables most strongly associated with employee departure.

• Rebalance project assignment. Employees with the highest levels of mean satisfaction are assigned to 4 or 5 projects. Restrict employees to a maximum of 5 projects.
• Restrict employee hours. Restrict average monthly hours for employees to the ranges presented by most satisfied employees, a maximum of 230 hours.

Limitations & Next Steps

Further steps to improve this investigation may include:

• Sourcing information on type of employee departure (fired, redundant, resignation). The data shows significant quantities of employees departed with medium and high satisfaction levels, primarily from the collection of employees assigned to 2 projects and in the departments of sales and human resources. Distinguishing between how employees left Salifort may assist in analyzing where employee resources may be reassigned to compensate over-worked employees working on 6 or 7 projects.
• Assessing employee previous satisfaction levels. The XGBoost model identifies previous satisfaction levels as a key indicator towards predicting an employee's eventual departure. Analysis of thresholds on satisfaction could provide potential risk management of employee attrition to monitor work environment and maintain sufficient satisfaction levels.

Repository & Technical Details

For those interested, the GitHub repo includes Jupyter notebooks for EDA, prediction models, and figure generation.