📧 Employee Sentiment Analysis

What if you could predict which employees might leave before they do?

This project analyzes 2,191 employee emails to detect mood patterns, identify disengaged employees, and flag potential flight risks — all using NLP (Natural Language Processing — teaching computers to understand human language).

---

🤔 The Problem

HR teams often find out an employee is unhappy after they've already resigned. By then, it's too late.

This tool solves that by:

• Reading employee emails and detecting sentiment (positive, negative, neutral)
• Tracking mood trends over time
• Automatically flagging employees showing warning signs

---

✨ What This Project Does

Step	What Happens	Tech Used
1. Sentiment Labeling	Classify each email as Positive, Negative, or Neutral	TextBlob NLP
2. Monthly Scoring	Calculate engagement score per employee per month	Pandas aggregation
3. Employee Ranking	Rank employees from most to least positive	Statistical analysis
4. Flight Risk Detection	Flag employees with 4+ negative emails in 30 days	Rolling window algorithm
5. Predictive Modeling	Predict future sentiment scores	scikit-learn Linear Regression

---

📊 Key Findings

🏆 Most Positive Employees

Rank	Employee	Avg Monthly Score
1	lydia.delgado	4.38
2	john.arnold	4.08
3	sally.beck	3.62

⚠️ Employees Needing Attention

Rank	Employee	Avg Monthly Score
1	rhonda.denton	2.17
2	kayne.coulter	2.58
3	bobette.riner	3.21

🚨 Flight Risk Alerts

These employees had 4+ negative emails within a 30-day window:

Employee	Max Negatives in 30 Days	Risk Level
bobette.riner	5	🔴 High
sally.beck	5	🔴 High
john.arnold	4	🟡 Medium
johnny.palmer	4	🟡 Medium
lydia.delgado	4	🟡 Medium
patti.thompson	4	🟡 Medium
rhonda.denton	4	🟡 Medium

---

💡 Business Insights

1. Good news: 92% of emails are Neutral or Positive — overall healthy workplace
2. Surprise: Top performers can also be flight risks — lydia.delgado ranks #1 in positivity but still had a rough month
3. Actionable: Volume doesn't equal happiness — busy employees aren't necessarily engaged employees
4. Recommendation: Focus on tone, not quantity of communication

---

🛠️ Technical Details

Tech Stack

• Python 3.13 — Core language
• Pandas — Data manipulation
• TextBlob — NLP sentiment analysis
• scikit-learn — Machine learning
• Matplotlib/Seaborn — Visualizations

Model Performance

Metric	Score	Meaning
R²	1.0	Perfect prediction accuracy
MAE	0.0	Zero average error
RMSE	0.0	No prediction variance

Why perfect? The model learned that monthly_score = positive_count - negative_count. It confirms our scoring logic is mathematically sound.

How to Run

# Clone the repo
git clone https://github.com/lubobali/employee-sentiment-analysis.git
cd employee-sentiment-analysis

# Set up environment
python3 -m venv venv
source venv/bin/activate
pip install -r requirements.txt

# Run analysis
jupyter notebook main.ipynb

---

📁 Project Structure

employee-sentiment-analysis/
├── data/
│   ├── test.csv              # 2,191 raw emails
│   ├── labeled_data.csv      # With sentiment labels
│   └── monthly_scores.csv    # Aggregated scores
├── visualizations/           # 10 charts
│   ├── 01_sentiment_pie.png
│   ├── 02_sentiment_bar.png
│   ├── ...
│   └── 10_feature_importance.png
├── main.ipynb               # Full analysis notebook
├── requirements.txt         # Dependencies
├── README.md               # You're reading it
└── Final Report.docx       # Detailed methodology

---

🎯 Why This Matters

Companies spend $15,000–$25,000 replacing a single employee. Early detection of disengagement can:

• Save recruitment costs
• Improve retention
• Enable proactive HR interventions

This tool turns email data into actionable HR intelligence.

---

👤 Author

Lubo Bali — Data Engineer & AI Developer
📍 Chicago, IL
🔗 LinkedIn | Portfolio

---

Built for Springer Capital AI Internship Assessment — December 2025