Remote Work Burnout Prediction — Machine Learning Pipeline

A complete end-to-end machine learning pipeline to predict employee burnout risk in remote working environments. This project covers exploratory data analysis, class imbalance handling, multi-model classification, regression, threshold tuning, cross-validation, and feature importance analysis.

---

📋 Table of Contents

• Overview
• Dataset
• Project Structure
• Pipeline Stages
• Models Used
• Key Results
• Installation
• Usage
• Notebooks
• Technologies
• Author

---

Overview

Remote work has increased significantly in recent years, and with it the risk of employee burnout. This project builds a machine learning system that:

• Classifies employees into burnout risk categories (Low or High)
• Predicts a continuous burnout score using regression
• Identifies the behavioural features that contribute most to burnout risk

A key challenge addressed in this project is severe class imbalance — High-risk cases form a small minority of the dataset. The pipeline includes SMOTE oversampling, class weighting, and decision threshold tuning to ensure the models are genuinely useful rather than superficially accurate.

---

Dataset

File: work_from_home_burnout_dataset.csv

Feature	Type	Description
work_hours	Continuous	Daily hours worked
screen_time_hours	Continuous	Daily screen time in hours
meetings_count	Integer	Number of meetings per day
breaks_taken	Integer	Number of breaks taken per day
after_hours_work	Binary	Whether the employee worked after hours (0 or 1)
sleep_hours	Continuous	Hours of sleep per night
task_completion_rate	Continuous	Percentage of daily tasks completed
day_type	Categorical	Weekday or Weekend
burnout_score	Continuous	Target (regression): numerical burnout score
burnout_risk	Categorical	Target (classification): Low or High

1,800 records across 10 features.

---

Project Structure

remote-work-burnout-prediction/
│
├── Remote_Work_Burnout_Improved.ipynb   # ✅ Full improved pipeline (recommended)
├── Remote_Work_Burnout_Prediction.ipynb # Original notebook
│
├── work_from_home_burnout_dataset.csv   # Dataset (add your own copy here)
│
├── requirements.txt                     # Python dependencies
├── .gitignore                           # Files excluded from version control
├── LICENSE                              # MIT License
└── README.md                            # This file

---

Pipeline Stages

Raw Data
   ↓
Exploratory Data Analysis (distributions, heatmaps, box plots)
   ↓
Preprocessing (LabelEncoder, StandardScaler, null checks)
   ↓
SMOTE — Synthetic Minority Oversampling
   ↓
Model Training (RF Classifier, Logistic Regression, Gradient Boosting)
   ↓
Threshold Tuning (F1-optimal decision boundary for RF)
   ↓
Evaluation (Confusion Matrix, ROC-AUC, Classification Report)
   ↓
Cross-Validation (5-fold Stratified)
   ↓
Feature Importance Analysis
   ↓
Regression (RF Regressor → burnout score)
   ↓
Sample Prediction (new employee profile)

---

Models Used

Classification (Burnout Risk: Low / High)

Model	Key Settings
Random Forest Classifier	200 estimators, class_weight=balanced, threshold tuned
Logistic Regression	max_iter=1000, class_weight=balanced
Gradient Boosting Classifier	200 estimators, lr=0.05, max_depth=4

Regression (Burnout Score)

Model	Key Settings
Random Forest Regressor	200 estimators

---

Key Results

Note: Exact values depend on random state and SMOTE sampling. Values below are representative.

Model	Accuracy	F1 (High class)	ROC-AUC
Random Forest (tuned)	~0.92	Improved	~0.97
Logistic Regression	~0.89	Moderate	~0.94
Gradient Boosting	~0.94	Strong	~0.98

Regression (Burnout Score)

• R² score demonstrates strong predictive fit
• Visualised with predicted vs actual scatter plot and residuals plot

Critical fix applied: Without threshold tuning, the Random Forest achieved 97.5% accuracy but detected zero High-risk cases. After tuning, High-risk recall improved substantially — accuracy alone is not a valid metric for imbalanced classification.

---

Installation

1. Clone the repository

git clone https://github.com/YOUR_USERNAME/remote-work-burnout-prediction.git
cd remote-work-burnout-prediction

2. Create a virtual environment (recommended)

python -m venv venv
source venv/bin/activate        # macOS / Linux
venv\Scripts\activate           # Windows

3. Install dependencies

pip install -r requirements.txt

4. Add the dataset

Place your work_from_home_burnout_dataset.csv file in the root of the repository.

5. Launch Jupyter

jupyter notebook

---

Usage

Open Remote_Work_Burnout_Improved.ipynb and run all cells from top to bottom. The notebook is fully self-contained and will:

1. Load and explore the dataset
2. Preprocess features and handle class imbalance
3. Train and evaluate all models
4. Display visualisations for each stage
5. Predict burnout risk for a sample employee profile
6. Print a full project summary

To predict for a custom employee, edit the sample dictionary in Section 15:

sample = pd.DataFrame({
    'day_type_enc':         [1],      # 1 = Weekday, 0 = Weekend
    'work_hours':           [9.5],
    'screen_time_hours':    [12.2],
    'meetings_count':       [5],
    'breaks_taken':         [1],
    'after_hours_work':     [1],
    'sleep_hours':          [5.6],
    'task_completion_rate': [60]
})

---

Notebooks

Notebook	Description
Remote_Work_Burnout_Improved.ipynb	Full pipeline with all fixes and additions — start here
Remote_Work_Burnout_Prediction.ipynb	Original exploratory notebook

---

Technologies

Library	Purpose
pandas	Data loading and manipulation
numpy	Numerical computation
matplotlib	Visualisation
seaborn	Statistical plots and heatmaps
scikit-learn	Preprocessing, models, and evaluation
imbalanced-learn	SMOTE oversampling
jupyter	Interactive notebook environment

---

Code Screenshots

---

Author

Muhammad Sheryar Adil

---

License

This project is licensed under the MIT License. See LICENSE for details.