💼 Employee Attrition Predictor

A machine learning web app that predicts whether an employee is likely to leave a company, built with Streamlit, XGBoost and Random Forest.

Live demo: https://employee-attrition-classifier.streamlit.app

How we built it

1. Dataset: We used the IBM HR Analytics Employee Attrition dataset: 1,470 employees with 35 attributes such as age, income, overtime, job satisfaction, work-life balance and years at the company. Only about 16% of employees in it left, so the data is heavily imbalanced.
2. Preprocessing: Categorical columns (department, job role, marital status, etc.) were label-encoded, and all features were scaled with a RobustScaler, which is less sensitive to outliers such as very high salaries.
3. Feature engineering: We created 10 new features that capture relationships between attributes, for example overtime × job satisfaction, distance from home × overtime, income per year of experience and years since last promotion relative to total experience.
4. Handling imbalance: The data was split 80/20 into training and test sets (keeping the same attrition ratio in both). On the training set we applied SMOTE-Tomek, which creates synthetic examples of employees who left and removes borderline cases, taking the ratio from 5.2 : 1 to 1 : 1.
5. Training and tuning: We trained a Random Forest and an XGBoost model, testing 10 hyperparameter combinations for each, and tuned the decision threshold to get the best F1 score.
6. Model selection: XGBoost performed best (F1 = 0.56 vs. 0.55 for Random Forest) and is the recommended model in the app, but both are available for comparison.
7. Business thresholds: Instead of a single yes/no cutoff, the app offers three strategies so HR can choose between catching more at-risk employees or raising fewer false alarms.
8. Web app: The trained models were saved and loaded into a Streamlit app where you can test example employees, enter your own data, and explore the model's performance.

Features

• Example profiles: 10 real employees from the test set, with their actual outcome and the model's prediction
• Manual input: enter any employee's details and get an attrition risk score
• Model choice: XGBoost or Random Forest
• Threshold strategies: High Recall, Balanced or High Precision
• Performance metrics: accuracy, F1, ROC-AUC, confusion matrices, classification reports and feature importance

Results (test set)

Model	Accuracy	F1	ROC-AUC
XGBoost	86.7%	0.562	0.809
Random Forest	80.6%	0.551	0.803

XGBoost with different thresholds:

Strategy	Threshold	Precision	Recall
High Recall	0.10	0.35	0.74
Balanced	0.33	0.60	0.53
High Precision	0.60	0.71	0.32

What's in each folder

• main.py: the Streamlit app
• dataset/: the HR employee attrition dataset (CSV)
• model/: the training notebook (AI_Project_Code.ipynb) and the saved models, scaler and encoders used by the app
• documentation/: the project proposal, final report and presentation (Word, PowerPoint and PDF)

Run it locally

You need Python 3.11 or newer.

# 1. Check your Python version
python3 --version
# If it's older than 3.11, check whether you have a newer one installed,
# e.g. python3.12 --version, and use that name instead of python3 below.

# 2. Download the project
git clone https://github.com/violazeneli31/EmployeeAttritionClassifier.git
cd EmployeeAttritionClassifier

# 3. Create a virtual environment and install the libraries
python3 -m venv .venv            # or: python3.12 -m venv .venv
source .venv/bin/activate        # Windows: .venv\Scripts\activate
pip install -r requirements.txt

# 4. Start the app
streamlit run main.py

The app opens at http://localhost:8501.

Mac users: XGBoost needs the OpenMP library, which macOS doesn't include. If you get an error saying libomp.dylib could not be loaded, install it with Homebrew and start the app again:

brew install libomp

The library versions in requirements.txt are pinned to the ones the models were trained with. Other versions of scikit-learn or XGBoost may fail to load the saved models.

Dataset

IBM HR Analytics Employee Attrition & Performance (Kaggle)

Authors

Sidrit Rrustemi and Viola Zeneli: CEN 352 Artificial Intelligence, SWE 3C