📊 HR Analytics – Employee Promotion Prediction using Machine Learning

1️⃣ Business Problem

HR departments often struggle to identify employees who truly deserve promotions.
Bias, incomplete information, and manual reviews can lead to unfair decisions and lower employee satisfaction.

👉 Our task: Predict whether an employee will be promoted based on historical HR data.
This ensures data-driven fairness in promotions and helps HR managers focus on top talent.

---

2️⃣ Data Overview

📂 Dataset Files

• train.csv: 54,808 employees (with target is_promoted)
• test.csv: 23,490 employees (without target, used for predictions)

🔑 Key Columns

• employee_id: Unique identifier (not useful for prediction → dropped later)
• department, region, education, gender, recruitment_channel: Categorical attributes
• no_of_trainings, age, previous_year_rating, length_of_service, KPIs_met >80%, awards_won?, avg_training_score: Numerical attributes
• is_promoted: Target variable (0 = Not promoted, 1 = Promoted)

---

3️⃣ Methodology

🔍 Step 1: Exploratory Data Analysis (EDA)

• Checked missing values → Found missing education, previous_year_rating
• Plotted univariate distributions (age, training scores, ratings)
• Bivariate analysis → Promotions are strongly linked to:
  • High KPI (>80%)
  • High ratings (≥4)
  • Awards won
• Outlier detection on age, length_of_service
• Correlation matrix → low collinearity, so features are useful

✅ Reasoning:
EDA ensures we understand patterns before modeling.
E.g., knowing that KPI + rating correlate with promotion → feature engineering idea.

---

🧹 Step 2: Preprocessing

Script: scripts/preprocessing.py

• Missing categorical (education) → filled with "Unknown"
• Missing numeric (previous_year_rating) → filled with median
• Dropped employee_id (identifier, no predictive value)
• Label Encoding for categorical features (department, region, education, gender, recruitment_channel)

✅ Reasoning:
ML models cannot handle NaNs or text → must be numeric and clean.
Dropping IDs avoids overfitting (model memorizing instead of learning).

---

🛠 Step 3: Feature Engineering

Script: scripts/feature_engineering.py

Added new HR-relevant features:

1. age_bucket → Young / Mid / Senior
2. tenure_bucket → New / Experienced / Veteran
3. high_performance_flag → (KPI >80% + rating ≥4)

✅ Reasoning:
HR often thinks in groups (young talent, experienced hires, veterans).
Features like high_performance_flag directly map to promotion rules.

---

🤖 Step 4: Model Training

Script: scripts/model_training.py

• Train/Test Split (80/20, stratified to preserve promotion ratio)
• Compared baseline models:
  • Logistic Regression
  • Decision Tree
  • Random Forest
  • XGBoost
• Applied Cross-validation (CV=5)
• Performed Hyperparameter tuning (GridSearchCV)
  • RF → tuned n_estimators, max_depth
  • XGB → tuned n_estimators, max_depth, learning_rate
• Final evaluation on test split
• Saved best model (.pkl)

✅ Reasoning:
We start simple (LogReg), then test complex models.
RF/XGB capture non-linear relationships better → improved accuracy.

---

🔮 Step 5: Predictions

Script: scripts/predict.py

• Loaded best model
• Transformed test data (same encoders + features)
• Predicted is_promoted_predicted
• Saved results → output/predictions.csv

✅ Reasoning:
Encoders ensure consistent category handling.
Predictions are reproducible and exportable for HR review.

---

🖥️ Step 6: Streamlit Dashboard

App: app.py

Features:

1. Single Prediction Tab
   • HR inputs employee details manually
   • Model outputs promotion likelihood (✅ Likely / ❌ Unlikely) + confidence score (%)
2. Batch Prediction Tab
   • Upload CSV of employees
   • Predictions + confidence for all employees
   • Downloadable results CSV
   • Visualizations:
     • Promotion distribution by Department
     • KPI >80% pie chart
     • Education vs Promotion Rate

✅ Reasoning:
Streamlit makes ML accessible to non-technical users (HR managers).
Visuals improve interpretability of model outputs.

---

4️⃣ Logging & Monitoring

• All scripts log progress to logs/
• Examples:
  • Data loaded, shape
  • Missing values handled
  • Model accuracy scores
  • Errors in plots/file paths

✅ Reasoning:
Logs help in debugging & tracking experiments.

---

5️⃣ Results

• Best Model: Random Forest / XGBoost (depending on dataset splits)
• Accuracy: ~92%
• Key Insights:
  • Employees with high KPI + rating ≥4 have 3x higher promotion chances
  • Awards also strongly linked to promotions
  • Education has weaker impact compared to performance metrics

---

Output

App Screenshots:

---

🚀 Demo:

HR.-.Analytics.mp4

---

6️⃣ Future Work

• Integrate SHAP explainability → show why a specific employee got predicted result
• Add fairness metrics (avoid gender/region bias)
• Deploy on AWS/GCP/Streamlit Cloud
• Automate retraining with MLflow pipeline

---

7️⃣ Project Structure

HR_Analytics/ │ ├── data/ # Raw datasets (train.csv, test.csv) ├── cleaned_data/ # Processed datasets after cleaning & feature engineering ├── output/ │ ├── figures/ # EDA plots (univariate, bivariate, outliers, correlations) │ └── models/ # Trained models (.pkl) + encoders ├── logs/ # Logging outputs for debugging & monitoring ├── notebook/ # Jupyter notebooks for EDA & experimentation ├── scripts/ # Modular Python scripts for pipeline │ ├── __init.py __ │ ├── eda.py │ ├── preprocessing.py # Missing value handling, label encoding │ ├── feature_engineering.py # Age buckets, tenure buckets, performance flags │ ├── model_training.py # Model training, hyperparameter tuning, evaluation │ ├── predict.py # Load model, run predictions, save results │ ├── app.py # Streamlit dashboard for single/batch predictions ├── requirements.txt # Project dependencies └── README.md # Documentation

---

8️⃣ Setup Instructions

1. Clone the repository

   git clone <repo-url>
   cd HR_Analytics

2. Create and activate a virtual environment

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

3. Install dependencies

   pip install -r requirements.txt

4. Run the app

   streamlit run app.py

5. Usage

   • Fill in employee details in the sidebar.
   • Click "Predict Promotion Likelihood" to see the result.
   • View EDA images and insights if available.

Notes

• Ensure output/models/XGBoost_Tuned.pkl and output/models/encoders.pkl exist. If not, train the model first.
• Place EDA images in output/figures/ to display them in the app.
• For any issues, check the logs or raise an issue in the repository.

Requirements

See requirements.txt for all Python dependencies.

---

👤 Author

Ayesha Banu

• 🎓 M.Sc. Computer Science | 🏅 Gold Medalist
• 💼 Data Scientist | Data Analyst | Full-Stack Python Developer | GenAI Enthusiast
• 📫 LinkedIn
• Project: HR Analytics – Employee Promotion Prediction -- Aug/2025

---

📄 License

Distributed under the MIT License. See LICENSE file for details.