app.py

import gradio as gr
import xgboost as xgb
import numpy as np
import pandas as pd
import joblib
import os
import warnings
import shap
import matplotlib.pyplot as plt
from sklearn.metrics import roc_curve, precision_recall_curve

# Suppress XGBoost warnings
warnings.filterwarnings("ignore", category=UserWarning, message=".*WARNING.*")

# Load your model (automatically detect XGBoost or joblib model)
def load_model():
    model_path = "xgboost_model.json"  # Ensure this matches your file name
    if os.path.exists(model_path):  
        model = xgb.Booster()
        model.load_model(model_path)
        print("✅ Model loaded successfully.")
        return model
    else:
        print("❌ Model file not found.")
        return None

model = load_model()

# Automatically find the best threshold using ROC
def optimize_threshold(X_test, y_test):
    dmatrix = xgb.DMatrix(X_test)
    y_prob = model.predict(dmatrix)
    
    fpr, tpr, thresholds = roc_curve(y_test, y_prob)
    optimal_idx = np.argmax(tpr - fpr)
    optimal_threshold = thresholds[optimal_idx]
    return optimal_threshold

# Prediction function with dynamic threshold
def predict_employee_status(satisfaction_level, last_evaluation, number_project,
                            average_monthly_hours, time_spent_company, 
                            work_accident, promotion_last_5years, salary, department, threshold=0.5):
    
    # One-hot encode the department
    departments = [
        'RandD', 'accounting', 'hr', 'management', 'marketing',
        'product_mng', 'sales', 'support', 'technical'
    ]
    department_features = {f"department_{dept}": 0 for dept in departments}
    if department in departments:
        department_features[f"department_{department}"] = 1

    # Prepare the input with all 17 features as a DataFrame with column names
    input_data = {
        "satisfaction_level": [satisfaction_level],
        "last_evaluation": [last_evaluation],
        "number_project": [number_project],
        "average_monthly_hours": [average_monthly_hours],
        "time_spent_company": [time_spent_company],
        "Work_accident": [work_accident],
        "promotion_last_5years": [promotion_last_5years],
        "salary": [salary],
        **department_features
    }

    input_df = pd.DataFrame(input_data)

    # Predict using the model
    if model is None:
        return "❌ No model found. Please upload the model file."

    try:
        dmatrix = xgb.DMatrix(input_df)
        prediction = model.predict(dmatrix)
        prediction_prob = prediction[0]

        # Apply the dynamic threshold
        result = "✅ Employee is likely to quit." if prediction_prob >= threshold else "✅ Employee is likely to stay."
        explanation = explain_prediction(input_df)
        return f"{result} (Probability: {prediction_prob:.2%})\n\nExplanation:\n{explanation}"
    except Exception as e:
        return f"❌ Error: {str(e)}"

# SHAP Explainability (Directly Integrated)
def explain_prediction(input_df):
    try:
        explainer = shap.TreeExplainer(model)
        shap_values = explainer.shap_values(input_df)
        
        # Generate and save SHAP explanation as an image
        shap.initjs()
        plt.figure()
        shap.waterfall_plot(shap.Explanation(values=shap_values[0], 
                                             base_values=explainer.expected_value,
                                             data=input_df.iloc[0].values,
                                             feature_names=input_df.columns))
        plt.savefig("shap_explanation.png")
        return "SHAP explanation generated for this prediction."
    except Exception as e:
        return f"❌ Error in SHAP: {str(e)}"

# Gradio interface with dynamic threshold and SHAP
def gradio_interface():
    interface = gr.Interface(
        fn=predict_employee_status,
        inputs=[
            gr.Number(label="Satisfaction Level (0.0 - 1.0)"),
            gr.Number(label="Last Evaluation (0.0 - 1.0)"),
            gr.Number(label="Number of Projects (1 - 10)"),
            gr.Number(label="Average Monthly Hours (80 - 320)"),
            gr.Number(label="Time Spent at Company (Years)"),
            gr.Radio([0, 1], label="Work Accident (0 = No, 1 = Yes)"),
            gr.Radio([0, 1], label="Promotion in Last 5 Years (0 = No, 1 = Yes)"),
            gr.Radio([0, 1, 2], label="Salary (0 = Low, 1 = Medium, 2 = High)"),
            gr.Dropdown(
                ['RandD', 'accounting', 'hr', 'management', 'marketing',
                 'product_mng', 'sales', 'support', 'technical'], 
                label="Department"
            ),
            gr.Slider(0.1, 0.9, value=0.5, step=0.05, label="Prediction Threshold")
        ],
        outputs="text",
        title="Employee Retention Prediction System (With SHAP & ROC Threshold)",
        description="Predict whether an employee is likely to stay or quit based on their profile. Adjust the threshold for accurate predictions.",
        theme="dark"
    )
    interface.launch()

gradio_interface()