app.py

import gradio as gr
import pickle
import pandas as pd
import shap
from shap.plots._force_matplotlib import draw_additive_plot
import gradio as gr
import numpy as np
from PIL import Image
import matplotlib.pyplot as plt

# load the model from disk
loaded_model = pickle.load(open("model-reduced.pkl", 'rb'))

# Setup SHAP
explainer = shap.Explainer(loaded_model) # PLEASE DO NOT CHANGE THIS.

# Create a dictionary for feature name mapping to make SHAP interpretable 
feature_names_mapping = {
    'Tenure': 'Tenure',
    'LearningDevelopment': 'Learning & Development',
    'WorkEnvironment': 'Work Environment',
    'WorkLifeBalance': 'Work-Life Balance',
    'WellBeing': 'Well-Being',
    'Engagement': 'Employee Engagement',
    'GM3': 'General Manager Leadership'
}

# Create the main function for server
def main_func(Tenure, LearningDevelopment, WorkEnvironment, WorkLifeBalance, WellBeing, Engagement, GM3):
    
    tenure_key = {
        '< 6 Months' : 2,
        '6 Months - 2 Years' : 4,
        '2 to 5 Years' : 6,
        '5+ Years' : 7,
    }

    default_tenure = 5
    val_tenure = tenure_key.get(Tenure, default_tenure)
    
    new_row = pd.DataFrame.from_dict({
        'Tenure':val_tenure,
        'LearningDevelopment':LearningDevelopment,
        'WorkEnvironment':WorkEnvironment,
        'WorkLifeBalance': WorkLifeBalance,
        'WellBeing':WellBeing,
        'Engagement':Engagement,
        'GM3':GM3}, orient = 'index').transpose()
    
    prob = loaded_model.predict_proba(new_row)
    shap_values = explainer(new_row)

    fig, ax = plt.subplots(figsize=(6, 4))
    values = shap_values.values[0] 
    features = new_row.columns 
    colors = ['#FF0000' if v < 0 else '#1E4380' for v in values]
    sorted_indices = np.argsort(np.abs(values))[-7:]  
    sorted_values = values[sorted_indices]
    sorted_features = [feature_names_mapping[features[i]] for i in sorted_indices]
    sorted_colors = [colors[i] for i in sorted_indices]
    ax.barh(sorted_features, sorted_values, color=sorted_colors)
    ax.set_xlabel("SHAP Value Impact")
    ax.set_title("Feature Importance (SHAP)")

    plt.tight_layout()
    local_plot = plt.gcf()
    plt.close()

    return {"Leave ❌": float(prob[0][0]), "Stay ✅": 1-float(prob[0][0])}, local_plot

# Create the UI
title = "**Employee Turnover Predictor & Interpreter** 🏨🛎️"
description1 = """
This app predicts whether an employee is likely to stay with Hilton or leave based on their satisfaction across seven key workplace factors. The app produces two outputs: 1) the predicted probability of staying or leaving, and 2) a Shapley force plot that visualizes the impact of each factor on the stay/leave prediction."""

description2 = """
To use the app, click on one of the examples below, or adjust the values of the employee satisfaction factors, and click on Analyze. ✨ 
""" 

# Function to change logo color to #1E4380
def change_logo_color(logo_path, hex_color="#FFFFFF"):
    r = int(hex_color[1:3], 16)
    g = int(hex_color[3:5], 16)
    b = int(hex_color[5:7], 16)
    img = Image.open(logo_path).convert("RGBA")
    data = np.array(img)
    alpha = data[:, :, 3]
    mask = alpha > 0
    colored_data = data.copy()
    colored_data[mask, 0] = r
    colored_data[mask, 1] = g
    colored_data[mask, 2] = b
    colored_img = Image.fromarray(colored_data)
    colored_logo_path = "colored_logo.png"
    colored_img.save(colored_logo_path)
    return colored_logo_path

# Create the colored logo
colored_logo_path = change_logo_color("logo.png")


def change_specific_color(logo_path, target_hex="#232d4b", new_hex="#FFFFFF", tolerance=30):
    # Convert hex colors to RGB
    target_r = int(target_hex[1:3], 16)
    target_g = int(target_hex[3:5], 16)
    target_b = int(target_hex[5:7], 16)
    
    new_r = int(new_hex[1:3], 16)
    new_g = int(new_hex[3:5], 16)
    new_b = int(new_hex[5:7], 16)
    
    # Open the image
    img = Image.open(logo_path).convert("RGBA")
    data = np.array(img)
    
    # Create a color distance mask using Euclidean distance in RGB space
    r_diff = (data[:,:,0].astype(int) - target_r)**2
    g_diff = (data[:,:,1].astype(int) - target_g)**2
    b_diff = (data[:,:,2].astype(int) - target_b)**2
    
    # Calculate color distance
    color_distance = np.sqrt(r_diff + g_diff + b_diff)
    
    # Create mask for pixels within tolerance of target color and having alpha > 0
    color_mask = (color_distance <= tolerance) & (data[:,:,3] > 0)
    
    # Apply new color to matched pixels
    modified_data = data.copy()
    modified_data[color_mask, 0] = new_r
    modified_data[color_mask, 1] = new_g
    modified_data[color_mask, 2] = new_b
    
    # Create a new image with the modified data
    modified_img = Image.fromarray(modified_data)
    
    # Save the modified image
    output_path = "modified_logo.png"
    modified_img.save(output_path)
    
    return output_path

uva_logo_path = change_specific_color("UVA-Logo.png")


css = """
.container img {
    max-width: 50%;
    max-height: auto;
    display: inline-block;
    margin: 0;
}

# body {
#     background-color: white;  # looks bad, keep here for now though
# }
"""

# Create Gradio interface with the colored logo
with gr.Blocks(title=title, css=css) as demo:
    with gr.Row(elem_classes=["container"]):
        gr.Image(colored_logo_path, show_label=False, container=False)
        gr.Image(uva_logo_path, show_label=False, container=False)
    gr.Markdown(f"## {title}")
    gr.Markdown(description1)
    gr.Markdown("""---""")
    gr.Markdown(description2)
    gr.Markdown("""---""")
    with gr.Row():        
        with gr.Column():
            Tenure = gr.Dropdown(["< 6 Months", "6 Months - 2 Years", "2 to 5 Years", "5+ Years"], label="Employee Tenure", value="< 6 Months")
            LearningDevelopment = gr.Slider(label="Learning & Development", minimum=1, maximum=5, value=4, step=.1)
            WorkEnvironment = gr.Slider(label="Work Environment", minimum=1, maximum=5, value=4, step=.1)
            WorkLifeBalance = gr.Slider(label="Work-Life Balance", minimum=1, maximum=5, value=4, step=.1)
            WellBeing = gr.Slider(label="Well Being", minimum=1, maximum=5, value=4, step=.1)
            Engagement = gr.Slider(label="Employee Engagement", minimum=1, maximum=5, value=4, step=.1)
            GM3 = gr.Slider(label="General Manager Leadership (GM3)", minimum=1, maximum=5, value=4, step=.1)
            submit_btn = gr.Button("Analyze")
        with gr.Column(visible=True,scale=1, min_width=600) as output_col:
            label = gr.Label(label = "Predicted Label")
            local_plot = gr.Plot(label = 'Shap:')
        
            submit_btn.click(
                main_func,
                [Tenure, LearningDevelopment, WorkEnvironment, WorkLifeBalance, WellBeing, Engagement, GM3],
                [label,local_plot], api_name="Employee_Turnover"
            )
    
    gr.Markdown("### Click on any of the examples below to see how it works:")
    gr.Examples([
                ["6 Months - 2 Years", 2.9,3.4,3.1,3.3,3.4,3.2],
                ["6 Months - 2 Years", 4,4.1,3.9,4,4.2,4], 
                ["6 Months - 2 Years", 4.8,4.9,4.7,4.8,4.9,4.8]], 
                [Tenure, LearningDevelopment, WorkEnvironment, WorkLifeBalance, WellBeing, Engagement, GM3], 
                [label,local_plot], main_func, cache_examples=True)

demo.launch(share=True)