Employee Performance Analysis for INX Future Inc.

Machine learning multiclass classification model for predicting employee performance ratings. Achieves 92.5% test accuracy and 93.3% cross-validation F1-score with XGBoost.

Project Code: 10281 | Document Code: CDS_Project_2_INX_Future_Emp_Data_V1.6

---

Overview

Analyzes employee performance for INX Future Inc., a data analytics and automation solutions provider facing declining employee performance. Predicts performance ratings from 11 key features including environment satisfaction, salary hikes, and promotion history.

Final Model: XGBoost Classifier (multiclass)

• Test: Accuracy = 92.5%, F1-Score = 92.42%, ROC AUC = 0.976
• Cross-validation: F1-Score = 93.32% +/- 0.39%
• Features: 11 predictors (satisfaction scores, tenure, compensation)
• Overfitting gap: 0.83%
• Class weights: {0: 2.065, 1: 0.458, 2: 3.019} for imbalance handling

See reports/Analysis/ for detailed business insights and recommendations.

---

Quick Start

Installation

pip install -r requirements.txt

Load Trained Model

import joblib
model = joblib.load('models/XGBClassifier_model_multiclass.pkl')

# Predict performance ratings (0=Low, 1=Average, 2=Outstanding)
predictions = model.predict(X_new)  # Requires 11 engineered features

Run Full Pipeline

Execute notebooks in sequence:

1. notebooks/data_exploratory_analysis.ipynb - EDA and feature engineering
2. notebooks/data_processing.ipynb - Encoding and dataset preparation
3. notebooks/train_model.ipynb - Model training and evaluation
4. notebooks/predict_model.ipynb - Inference on new data

Dataset

Property	Details
Source	INX Future Inc. Employee Database
Samples	1,200 employees
Features	28 attributes (27 predictors + PerformanceRating target)
Final Features	11 selected via importance analysis
Split	960 train / 240 test (stratified, random_state=42)
Classes	3 ratings (Low: 16.15%, Average: 72.92%, Outstanding: 10.94%)

See data/raw/Data dictionaries.md for full feature descriptions.

Project Structure

Core directories:

• data/raw/ - Original dataset (INX_Future_Inc_Employee_Performance_CDS_Project2_Data_V1.8.xls)
• data/processed/ - Train/test splits (X_train.csv, X_test.csv, y_train.csv, y_test.csv, prepared_data.csv)
• notebooks/ - Sequential analysis pipeline (4 notebooks)
• src/ - Reusable modules (statistical_analysis.py, model_evaluation.py, utils.py)
• models/ - Trained model (XGBClassifier_model_multiclass.pkl)
• reports/Analysis/ - Business insights and recommendations
• results/figures/ - Visualizations and plots

Working with the Notebooks

Import pattern used: The notebooks import functions from src/ modules using:

from src.utils import memory_usage, dataframe_memory_usage, garbage_collection
from src.statistical_analysis import normality_test_with_skew_kurt, spearman_correlation, calculate_vif
from src.model_evaluation import evaluate_model, mean_absolute_shap, cross_validation_analysis_table, plot_multiclass_evaluation

Pipeline execution order:

1. data_exploratory_analysis.ipynb - EDA, outlier treatment, multicollinearity detection, feature engineering
2. data_processing.ipynb - Encoding strategies, feature selection, dataset preparation
3. train_model.ipynb - Model comparison, hyperparameter tuning, SHAP analysis
4. predict_model.ipynb - Model inference and evaluation

Model Training Workflow

Base model evaluation:

from src.model_evaluation import evaluate_model

metrics = evaluate_model(model, X_train, y_train, X_test, y_test)
# Returns: Accuracy, Precision, Recall, F1-Score, ROC AUC, CV F1, Train Accuracy, Overfit Score, Training Time

Cross-validation analysis:

from src.model_evaluation import cross_validation_analysis_table

cv_results = cross_validation_analysis_table(
    model=model,
    X_train=X_train,
    y_train=y_train,
    cv_folds=5,
    scoring_metric='f1_weighted'
)

Multiclass evaluation plots:

from src.model_evaluation import plot_multiclass_evaluation

plot_multiclass_evaluation(
    model=model,
    X_test=X_test,
    y_test=y_test,
    class_labels=['Low', 'Average', 'Outstanding']
)
# Generates: ROC curve, Precision-Recall curve, Confusion matrices

Statistical Analysis Functions

Normality testing:

from src.statistical_analysis import normality_test_with_skew_kurt

normal_df, not_normal_df = normality_test_with_skew_kurt(df)
# Uses Shapiro-Wilk (n<=5000) or Kolmogorov-Smirnov (n>5000)
# Returns DataFrames with p-values, statistics, skewness, kurtosis

Multicollinearity detection:

from src.statistical_analysis import calculate_vif

vif_data, high_vif_features = calculate_vif(
    data=df,
    exclude_target='PerformanceRating',
    multicollinearity_threshold=5.0
)
# Returns VIF scores and features exceeding threshold

Spearman correlation:

from src.statistical_analysis import spearman_correlation

corr_matrix = spearman_correlation(
    data=df,
    non_normal_cols=['col1', 'col2'],
    exclude_target='PerformanceRating',
    multicollinearity_threshold=0.8
)
# Generates correlation heatmap and identifies multicollinear pairs

Model Persistence

Loading the final model:

import joblib

# Load trained XGBoost model
model = joblib.load('models/XGBClassifier_model_multiclass.pkl')

# Make predictions (requires 11 features in correct order)
predictions = model.predict(X_new)

# Get probability scores
probabilities = model.predict_proba(X_new)

Required features for inference (in order):

1. EmpLastSalaryHikePercent
2. EmpEnvironmentSatisfaction
3. YearsSinceLastPromotion
4. ExperienceYearsAtThisCompany
5. EmpDepartment_Development
6. EmpJobRole_freq
7. EmpWorkLifeBalance
8. OverTime
9. EmpJobSatisfaction
10. TotalWorkExperienceInYears
11. DistanceFromHome

Key Findings

Department-Wise Performance

Department	% Excellent/Outstanding	Attrition Rate
Development & Data Science	96.4%	5.1%
Sales	76.7%	18.6%
Finance	69.4%	15.3%

Analysis:

• Development teams show higher performance with frequent promotions (>20% salary hikes)
• Sales/Finance departments show low satisfaction (40-50% "Low" scores) and longer promotion gaps (3.7 years average)

Top 3 Performance Drivers (SHAP Analysis)

1. EmpEnvironmentSatisfaction (SHAP: 0.4945)
   • Low satisfaction employees 3x more likely to underperform
2. EmpLastSalaryHikePercent (SHAP: 0.372)
   • Employees with <15% hikes show 40% lower performance
3. YearsSinceLastPromotion (SHAP: 0.2631)
   • Delays >3 years lead to 35% lower performance

Model Comparison

Top Models After Hyperparameter Tuning

Model	Accuracy	F1-Score	ROC AUC	Training Time	Overfit	CV F1
XGBoost	92.50%	92.42%	0.9756	1.013s	0.0083	0.9332
Random Forest	92.50%	92.34%	0.9731	1.000s	0.0531	0.9273
HistGradientBoosting	91.25%	91.24%	0.9616	1.615s	0.0750	0.9156
CatBoost	90.83%	90.80%	0.9700	1.612s	0.0865	0.9172
LightGBM	89.17%	89.19%	0.9595	1.233s	0.1083	0.9147

XGBoost Class-Wise Performance

Rating	Precision	Recall	F1-Score
Low (0)	82%	85%	84%
Average (1)	94%	97%	95%
Outstanding (2)	100%	77%	87%

Why XGBoost was selected:

• Minimal overfitting (0.83% gap vs 5.31% for Random Forest)
• Highest ROC AUC (0.9756) and CV F1 stability (93.32% +/- 0.39%)
• Fast training (1.013s) with strong generalization
• SHAP explainability for business insights

Key Design Decisions

Feature engineering:

• YearsSinceFirstJob = Age - EducationYears (based on education level mapping)
• TotalCompensation = EmpHourlyRate * 2080 (annual working hours)

Outlier treatment:

• Winsorization at 95th percentile: TotalWorkExperienceInYears, ExperienceYearsAtThisCompany
• Domain-informed capping: YearsSinceLastPromotion (10), NumCompaniesWorked (8)

Encoding strategies:

• One-hot encoding (cardinality <= 6): Gender, MaritalStatus, EmpDepartment, etc.
• Frequency encoding (cardinality > 6): EmpJobRole
• Ordinal encoding: Satisfaction scores (1-4), EmpEducationLevel
• Binary encoding: OverTime, Attrition

Multicollinearity resolution:

• Dropped 4 features: Age (VIF=23.3), EmpHourlyRate, ExperienceYearsInCurrentRole (r=0.87), YearsWithCurrManager

Reports

Detailed analysis in reports/Analysis/:

• Employee Performance Analysis Report for INX Future Inc..md - Full methodology and results
• Top 3 Factors Affecting Employee Performance.md - SHAP analysis and insights
• Department-Wise Performance Analysis.md - Departmental comparisons
• Employee Performance Prediction Model.md - Model development details
• Recommendations to Improve Employee Performance at INX Future Inc..md - Strategic recommendations
• GALLERY.md - Visualizations (44 figures)

Development

Code Quality

# Install development dependencies
pip install -r requirements-dev.txt

# Setup pre-commit hooks
pre-commit install

# Format code
black .
isort .

# Lint code
flake8 .

# Format notebooks
nbqa black notebooks/

# Run all pre-commit checks
pre-commit run --all-files

Pre-commit Hooks

Configuration (.pre-commit-config.yaml):

• black (88-char lines)
• isort (black-compatible)
• nbqa-black (notebooks)
• Validation (YAML, trailing whitespace, file size checks)

Configuration Files

• .flake8: Max line length 88, max complexity 10, numpy docstring convention
• pyproject.toml: Black and isort settings (line length 88, Black profile)

---

MIT License - Copyright (c) 2025 Dhanesh B. B.