HR Attrition Intelligence Hub

This project transforms attrition prediction into a decision support system for HR strategy.

Problem & Value

Employee attrition is costly and difficult to manage proactively.

Most solutions stop at prediction — this system goes further by:

• Explaining why employees are at risk
• Prioritizing who to retain
• Recommending what actions to take

---

Live Demo: Attrition Decision Support System

Experience the live strategic dashboard here: https://employee-retention-ai.streamlit.app/

---

⚡ Quick Start

git clone https://github.com/Andrew-Hany/employee-attrition-decision-system.git
cd employee-attrition-decision-system
pip install -r requirements.txt
streamlit run app.py

---

---

📊 Phase 1: Exploratory Data Analysis (EDA)

Our initial discovery (detailed in EDA.ipynb) revealed that attrition is driven by four core strategic drivers. By analyzing these sub-metrics, we can move from raw data to a strategic "Employee Value" classification.

The Attrition Driver Tree

graph LR
    D[Attrition Drivers] --> W[Workload Stress]
    D --> C[Seniority & Comp.]
    D --> L[Life Stage Mobility]
    D --> R[Role Design]
    
    W --> W1[Overtime]
    W --> W2[Job Involvement]
    
    C --> C1[Job Level / Salary]
    C --> C2[Income Efficiency]
    C --> C3[Promotion Velocity]
    
    L --> L1[Age / Experience]
    L --> L2[Marital Status]
    L --> L3[Jump History]
    
    R --> R1[Travel Load]
    R --> R2[Job Satisfaction]
    R --> R3[Environment Fit]

    style D fill:#f8fafc,stroke:#334155,stroke-width:2px
    classDef pillar fill:#eff6ff,stroke:#3b82f6,stroke-width:1px;
    class W,C,L,R pillar

Loading

Strategic Pillar Details

1. Workload: Impact of overtime vs. role involvement.
2. Seniority & Compensation: Evaluation of salary competitiveness (Income Efficiency) and growth rate (Promotion Velocity).
3. Life Stage: Factors in external tenure stability and professional mobility.
4. Role Structure: Focuses on travel burden and environmental satisfaction.

---

⚙️ Phase 2: Model Training Pipeline

The technical development of our predictive engine is detailed in attrition_model.ipynb. This stage focuses on transforming raw behavior features into a high-recall retention model.

The Training Workflow

graph TD
    subgraph "Training Lifecycle"
        D1[Raw IBM HR Data] --> P1[Data Pre-processing]
        P1 --> F1[Feature Engineering]
        F1 --> M1[Model Selection & Tuning]
        M1 --> T1[Final Model Training]
    end
    
    style T1 fill:#f0fdf4,stroke:#16a34a,stroke-width:2px

Loading

Model Selection & Performance Results

We prioritized Recall (catching as many potential leavers as possible) while maintaining a balanced Precision. As shown below, SMOTE (Oversampling) was the key breakthrough in our modeling strategy.

Experiment	Model	Threshold	Test ROC AUC	Recall (Leave)	Precision	F1-Score
I. Baseline	Logistic Regression	0.35	0.810	0.489	0.535	0.511
(No SMOTE)	CatBoost	0.35	0.780	0.511	0.453	0.480
	Random Forest	0.40	0.779	0.170	0.444	0.246
II. SMOTE	✨ Logistic Regression	0.40	0.789	🚀 0.766	0.353	0.483
(Over-sampled)	Random Forest	0.35	0.771	0.723	0.333	0.456
	CatBoost	0.35	0.758	0.681	0.327	0.441

Tip

Key Insight: Moving from Baseline to SMOTE increased our ability to catch leavers (Recall) from 48.9% to 76.6%, which is critical for a proactive retention strategy.

Threshold Optimization

The system uses a Logistic Regression model optimized at a 0.35 - 0.40 threshold.

• Lower Threshold (0.35): Maximizes Recall (Catching more potential leavers).
• Selection: We chose 0.40 for the baseline to achieve a 76.6% Recall rate.

---

🎯 Phase 3: Inference Strategy (Intelligence Hub)

The final system takes the trained model and wraps it in a Human-Centric Intelligence Layer. This hub transforms raw numerical predictions into a structured strategic narrative.

The Inference Pipeline

graph TD
    subgraph "Live Strategy Engine"
        D2[New Employee Data] --> P2[Model Prediction]
        P2 --> S2[Employee Value Scoring]
        S2 --> R2[Risk Driver Identification]
        R2 --> F2[Problem Framing]
        F2 --> O([Retention Actions / Decisions])
    end
    
    style O fill:#f0fdf4,stroke:#16a34a,stroke-width:3px

Loading

---

1. Strategic Value Scoring Logic

The system prioritizes retention for high-impact talent by processing model outputs through our Value Scoring Engine.

graph TD
    Start([Start: Base Score = 0]) --> Seniority[Seniority & Role]
    
    Seniority --> JL["+2: Job Level >= 3"]
    Seniority --> TC["+2: Tenure >= 5y"]
    Seniority --> MT["+1: High Mgr Tenure"]
    
    JL --> Perf{Performance}
    TC --> Perf
    MT --> Perf
    
    Perf --> PR3["+1: Rating == 3"]
    Perf --> PR4["+2: Rating >= 4"]
    Perf --> JI["+1: Involvement >= 3"]
    Perf --> IE["+1: Income per Level > 3000"]
    Perf --> PV["+1: Promotion Velocity > 2.5"]
    
    PR3 & PR4 & JI & IE & PV --> Status{Status Factors}
    
    Status --> ST["-2: Tenure < 2y"]
    Status --> JH["-1: Job Hopping (5+)"]
    Status --> LS["-1: Low Satisfaction (<= 2)"]
    
    ST & JH & LS --> Score[Final Score Calculation]
    
    Score --> HV([High Value Employee])
    Score --> V([Valuable Employee])
    Score --> A([Average Employee])
    Score --> LV([Low Value Employee])

    %% Design Styles
    style Start fill:#f8fafc,stroke:#64748b,stroke-width:1px
    style Seniority fill:#f8fafc,stroke:#64748b,stroke-width:1px
    style Perf fill:#eff6ff,stroke:#3b82f6,stroke-width:2px
    style Status fill:#eff6ff,stroke:#3b82f6,stroke-width:2px
    style Score fill:#f8fafc,stroke:#64748b,stroke-width:1px
    
    classDef booster fill:#eff6ff,stroke:#3b82f6,stroke-width:1px,color:#1e40af;
    class JL,TC,MT,PR3,PR4,JI,IE,PV booster
    
    classDef risk fill:#fff1f2,stroke:#f43f5e,stroke-width:1px,color:#9f1239;
    class ST,JH,LS risk

    style HV fill:#fef3c7,stroke:#fbbf24,stroke-width:3px
    style V fill:#dbeafe,stroke:#60a5fa,stroke-width:3px
    style A fill:#f1f5f9,stroke:#94a3b8,stroke-width:3px
    style LV fill:#fee2e2,stroke:#f87171,stroke-width:3px

Loading

---

2. Risk Identification & Problem Framing

This layer translates complex feature correlations into clear, human-readable Business Problem Frames.

graph TD
    A[Employee Features] --> B[Rule-Based Conditions]
    B --> C[Triggered Signals]
    C --> D{Driver Categories}
    
    D --> W[Workload Stress]
    D --> Co[Seniority & Comp.]
    D --> Ro[Role Structure]
    D --> Li[Life Stage]
    
    W & Co & Ro & Li --> SD[Structured Drivers]
    SD --> PF[Problem Framing]
    PF --> RA([Retention Actions / Decisions])
    
    style PF fill:#eff6ff,stroke:#3b82f6,stroke-width:2px
    style RA fill:#f0fdf4,stroke:#16a34a,stroke-width:3px

Loading

---

3. Strategic Action Mapping

Identified problems are linked directly to targeted HR interventions to ensure immediate, data-backed response.

graph TD
    PF[Problem Framing] --> BR[Burnout Risk]
    PF --> CM[Career Misalignment]
    PF --> TR[Travel Burden]
    PF --> PR[Personal Mobility Risk]
    
    BR --> A1[Reduce Workload]
    BR --> A2[Improve Work-Life Balance]
    CM --> A3[Review Compensation]
    CM --> A4[Define Promotion Path]
    TR --> A5[Adjust Role Design]
    TR --> A6[Reduce Travel]
    PR --> A7[Remote Work]
    PR --> A8[Relocation Support]

    classDef action fill:#f0fdf4,stroke:#16a34a,stroke-width:1px;
    class A1,A2,A3,A4,A5,A6,A7,A8 action
    classDef prob fill:#fff1f2,stroke:#f43f5e,stroke-width:1px;
    class BR,CM,TR,PR prob

Loading

---

🛠️ Technical Stack

• Model: Logistic Regression (optimized with SMOTE)
• Explainability Layer: Rule-based driver identification
• Decision Engine: Value scoring + problem framing
• Interface: Streamlit
• Visualization: Matplotlib, Seaborn, Mermaid

📂 Project Structure

• EDA.ipynb → Exploratory analysis
• attrition_model.ipynb → Model training & evaluation
• app.py → Streamlit application
• utils/ → Logic (scoring, drivers, recommendations)
• assets/ → Images & visuals

---

Developed by Andrew Zaki.