app.py

from data_utils.data_generation import UpliftSimulation
from data_utils.exploratory_data_analysis import ExploratoryAnalysis
from data_utils.feature_importance import FeatureImportance
from models_utils.ml_models import ModelTraining
from eval_utils.evaluation import ModelEvaluator

import matplotlib.pyplot as plt
import plotly.express as px
import plotly.graph_objects as go

import numpy as np
import pandas as pd
import streamlit as st

X_names = [
            'AgeIndex', 'IncomeIndex', 'PurchaseFrequencyIndex',
            'AccountLifetimeIndex', 'AverageTransactionValueIndex', 'PreferredPaymentMethodIndex', 'RegionIndex',
            'EmailDiscountCTRIndex', 'WebDiscountCTRIndex', 'SocialMediaEngagementIndex',
            'DirectMailDiscountResponseIndex', 'InAppDiscountEngagementIndex', 'FlashSaleParticipationIndex',
            'SeasonalPromoInterestIndex', 'LoyaltyProgramEngagementIndex', 'ReferralBonusUsageIndex',
            'DiscountCodeRedemptionIndex', 'VIPSaleAccessIndex', 'EarlyAccessOptInIndex',
            'ProductReviewAfterDiscountIndex', 'UpsellConversionIndex', 'CrossSellInterestIndex',
            'BundlePurchaseIndex', 'SubscriptionUpgradeIndex', 'CustomerFeedbackIndex',
            'BrowserTypeIndex', 'DeviceCategoryIndex', 'OperatingSystemIndex',
            'SessionStartTimeIndex', 'LanguagePreferenceIndex', 'NewsletterSubscriptionIndex',
            'AccountVerificationStatusIndex', 'AdBlockerPresenceIndex'
        ]

# Title
st.title("Uplift Modeling in Retail Demo")

tabs = st.sidebar.radio("Navigation", [
    "Overview",
    "Data generation",
    "Exploratory analysis",
    "Model training",
    "Economic effects"]
)

if tabs == "Overview":
    st.header("Overview")
    st.write("""
    This app demonstrates the use of uplift modeling to understand the effect of different actions (like promotions) on customer behavior. We generate a simulated dataset and use it to train a model that predicts the uplift effect of different treatments on customer behavior. We then evaluate the model using the Qini curve, which measures the uplift effect of a model compared to a random model.
    """)

    VIDEO_URL = "https://www.youtube.com/watch?v=AIL0QZWSWBY"
    st.video(VIDEO_URL)

    st.write("To get started, select the 'Data generation' tab from the sidebar.")
    st.write("For more details, reach out to [email protected]")

if tabs == "Data generation":

    st.header("Data Generation")

    # Description
    st.write("""
    This app creates a simulated dataset for a special kind of analysis called uplift modeling, which helps understand the effect of different actions (like promotions) on customer behavior. We use some default settings to make things easy:
    - We're looking at whether customers make a purchase or not.
    - We compare different types of promotions (like no discount, 5% off, etc.).
    - The dataset includes 15 different pieces of information (features) about each customer.
    """)

    # Interactive number of samples selection
    n = st.number_input('Number of Samples (n)', min_value=1000, value=10000, step=1000,
                        help="Total number of samples to generate in the dataset.")

    # Default values for other variables
    y_name = 'conversion'
    treatment_group_keys = ['control', 'discount_05', 'discount_10', 'discount_15']
    n_classification_features = 15
    n_classification_informative = 7
    n_classification_repeated = 0
    n_uplift_increase_dict = {'discount_05': 4, 'discount_10': 3, 'discount_15': 3}
    n_uplift_decrease_dict = {'discount_05': 0, 'discount_10': 0, 'discount_15': 0}
    positive_class_proportion = 0.05
    random_seed = 8097

    # Button to generate dataset
    if st.button('Generate Dataset'):
        uplift_sim = UpliftSimulation(n=n, y_name=y_name, treatment_group_keys=treatment_group_keys,
                                    n_classification_features=n_classification_features,
                                    n_classification_informative=n_classification_informative,
                                    n_classification_repeated=n_classification_repeated,
                                    n_uplift_increase_dict=n_uplift_increase_dict,
                                    n_uplift_decrease_dict=n_uplift_decrease_dict,
                                    positive_class_proportion=positive_class_proportion,
                                    random_seed=random_seed)
        uplift_sim.simulate_dataset()
        uplift_sim.apply_discounts_and_clean()
        uplift_sim.postprocess_tables()
        uplift_sim.add_monetary_effect()
        st.session_state.uplift_sim = uplift_sim # Store in session state

        st.write("Dataset Generated Successfully!")

        st.subheader("User profiles")
        st.write('Features that represent a customer such as age, income, purchase frequency, etc')
        st.dataframe(uplift_sim.dataframes[0].head(3))

        st.subheader("Treatments data")
        st.write('Information about the different treatments (discounts) that were applied to the customers as discounts in different channels (web, email, mobile), early access, etc')
        st.dataframe(uplift_sim.dataframes[1].head(3))

        st.subheader("Other data")
        st.write('Other data that can be used in the analysis')
        st.dataframe(uplift_sim.dataframes[2].head(3))

if tabs == "Exploratory analysis":

    st.header("Exploratory Analysis")

    if 'uplift_sim' in st.session_state:

        st.subheader('Summary statistics')
        uplift_sim = st.session_state.uplift_sim
        eda = ExploratoryAnalysis(uplift_sim.df)

        st.write('We begin by computing the total sum of conversions, sales (discounted price) and platform benefit. We can see that the total conversions and the total sales grows as the discount value is bigger. However, the platform benefit decreases.')
        
        sum_conversions, mean_conversions = eda.compute_summaries()
        st.write(sum_conversions)
        st.write(mean_conversions)

        st.write('We can also visualize the tradeoff between conversions and platform benefit by plotting the mean benefit per user on the y-axis and the mean conversion rate on the x-axis, for each treatment group.')
        mean_benefit_vs_conversion = eda.compute_mean_benefit_vs_conversion()

        # fig, ax = plt.subplots()
        # mean_benefit_vs_conversion.plot.scatter(x='conversion', y='benefit', c='DarkBlue', s=50, ax=ax)
        # st.pyplot(fig)

        fig = px.scatter(mean_benefit_vs_conversion, x='conversion', y='benefit', color_discrete_sequence=['LightBlue'], size_max=50)
        st.plotly_chart(fig)

        st.write('''
                 We further compute the Average Treatment Effect (ATE) for both the mean conversion rate and the mean benefit per user:
                 - Conversion ATE = Mean Conversion rate in the discounted group minus Mean Conversion rate in the control group
                 - Benefit ATE = Mean Benefit per user in the discounted group minus Mean Benefit per user in the control group
                 This helps illustrate how the discount value affects Conversion ATE and Benefit ATE.
                 ''')
        mean_conversions_ate = eda.compute_ate()

        # fig, ax = plt.subplots()
        # mean_conversions_ate.plot.scatter(x='conversion', y='benefit', c='DarkBlue', s=50, ax=ax)
        # st.pyplot(fig)

        fig = px.scatter(mean_conversions_ate, x='conversion', y='benefit', color_discrete_sequence=['LightBlue'], size_max=50)
        st.plotly_chart(fig)

        st.subheader('Feature importance')

        # Allow users to select a treatment group
        treatment_group = st.selectbox(
            'Select a treatment group',
            options=['discount_05', 'discount_10', 'discount_15'],
            index=0  # default to 'discount_05'
        )

        feature_importance = FeatureImportance(uplift_sim.df, X_names, y_name = 'conversion', treatment_group = treatment_group)
        fi = feature_importance.compute_feature_importance()
        # fig, ax = plt.subplots()
        di_df_sorted = fi.sort_values(by='score', ascending=False)
        # di_df_sorted[['feature', 'score']].plot.barh(x='feature', y='score', ax=ax)
        # st.pyplot(fig)
        fig = px.bar(di_df_sorted, y='feature', x='score', orientation='h')
        st.plotly_chart(fig)


        st.write("""
                    - AccountLifetimeIndex: Longer-standing accounts are key predictors of customer response to promotions \n
                    - CustomerFeedbackIndex: Customer feedback significantly influences the success of marketing strategies \n
                    - UpsellConversionIndex: The success rate of upselling is an important factor \n
                    - PurchaseFrequencyIndex: More frequent purchases indicate higher engagement and response to marketing efforts \n
                    - ReferralBonusUsedIndex and LoyaltyProgramEngagementIndex: Engagement with these programs is highly indicative of responsiveness to promotions
                 """)

    else:
        st.error("Please generate the dataset first.")

if tabs == "Model training":

    st.header("Model Training")
    st.write("""
        In this section, we train a model to predict the uplift effect of different treatments on customer behavior.
        We use the XGBoost algorithm to train the model. The model can be used to predict the conversion rate or the benefit per user for each treatment group.
        We can also analyze the economic effects of the treatments by comparing the uplift in conversion rate and benefit per user.
    """)

    if 'uplift_sim' in st.session_state:

        uplift_sim = st.session_state.uplift_sim

        model_trainer = ModelTraining(uplift_sim.df, 'conversion', X_names)

        model_type = st.radio("Choose the model type", ('Conversion Model', 'Benefit Model'))

        params = {
            'n_estimators': st.slider('Number of Estimators', 10, 100, 50),
            'max_depth': st.slider('Max Depth', 1, 10, 4),
            'colsample_bytree': st.slider('Colsample by Tree', 0.1, 1.0, 0.2),
            'subsample': st.slider('Subsample', 0.1, 1.0, 0.2),
        }
        control_name = 'control' # st.text_input('Control Group Name', 'control')
        test_size = st.slider('Test Size', 0.1, 0.9, 0.5)
        random_state = 20143 # st.slider('Random State', 0, 10000, 20143)

        if st.button('Train Model'):

            model_trainer.split_data(test_size=test_size, random_state=random_state)

            if model_type == 'Conversion Model':
                y_name = 'conversion' # st.selectbox('Select target variable for conversion', options=uplift_sim.target_options)
                model_trainer.y_name = y_name
                tau = model_trainer.fit_predict_classifier(params, control_name)
            elif model_type == 'Benefit Model':
                y_name = 'benefit' # st.selectbox('Select target variable for benefit', options=uplift_sim.benefit_options)
                model_trainer.y_name = y_name
                tau = model_trainer.fit_predict_regressor(params, control_name)

            st.session_state.model_trainer = model_trainer

            feature_importances = model_trainer.compute_feature_importance()

            st.subheader('Feature Importances')
            fig, ax = plt.subplots()

            # for k, v in feature_importances.items():
            #     st.write(f"Feature importance for {k}")
            #     v.plot(kind='barh', ax=ax)
            #     ax.set_xlabel("Importance")
            #     ax.set_ylabel("Feature")
            #     ax.set_title(f"Feature Importance for {model_type}")
            #     st.pyplot(fig)

            for k, v in feature_importances.items():
                # Reset index if 'v' is a Series or its index contains the feature names
                if isinstance(v, pd.Series) or 'feature' not in v.columns:
                    v = v.reset_index()
                    v.columns = ['feature', 'score']  # Adjust column names accordingly

                # Assuming 'v' now has columns ['feature', 'score']
                fig = px.bar(v, y='feature', x='score', orientation='h', 
                            title=f"Feature Importance for {model_type} ({k})", 
                            labels={'score': 'Importance', 'feature': 'Feature'})
                fig.update_layout(yaxis={'categoryorder':'total ascending'})  # Optional: This sorts the bars
                st.plotly_chart(fig)

    else:
        st.error("Please generate and preprocess the dataset first.")

if tabs == "Economic effects":

    st.header("Economic Effects Analysis")
    st.write("""
             We can evaluate our models by looking at the Qini curves. We can use the CATE conversion model to evaluate the performance on both the Conversion and the Benefit as a function of the fraction of users targeted.
            The Qini curve is a measure of the uplift effect of a model. It shows the difference between the uplift model and a random model.
        """)

    if 'uplift_sim' in st.session_state and 'model_trainer' in st.session_state:
        df_test = st.session_state.model_trainer.df_test
        model_type = st.radio("Choose the model type for analysis", ('Conversion Model', 'Benefit Model'))

        # Determine which model to use based on user selection
        if model_type == 'Conversion Model':
            model = st.session_state.model_trainer.conversion_learner_t
        elif model_type == 'Benefit Model':
            model = st.session_state.model_trainer.benefit_learner_t
        else:
            st.error("Invalid model type selected.")
            st.stop()

        if model == None:
            st.error("Please train the model first.")
            st.stop()

        evaluator = ModelEvaluator(model,
                                   df_test,
                                   X_names # df_test.columns.drop(['conversion', 'benefit', 'treatment_group_key'])
                                   )
        discounts = ['discount_05', 'discount_10', 'discount_15']
        qini_conversions = {}
        qini_benefits = {}

        for discount in discounts:
            qini_conv, qini_ben = evaluator.eval_performance(discount)
            qini_conversions[discount] = qini_conv
            qini_benefits[discount] = qini_ben
        
        # Plotting CATE Conversion
        st.subheader("CATE Conversion vs Targeted Population")
        
        # fig, ax_conversion = plt.subplots()
        # for discount, color in zip(discounts, ['b', 'g', 'y']):
        #     qini_conversions[discount].plot(ax=ax_conversion, x='index', y='S', color=color)
        #     qini_conversions[discount].plot(ax=ax_conversion, x='index', y='Random', color='r', ls='--')
        
        # ax_conversion.legend([f'{d} model' for d in discounts] + [f'{d} random' for d in discounts], prop={'size': 10})
        # ax_conversion.set_xlabel('Fraction of Targeted Users')
        # ax_conversion.set_ylabel('CATE Conversion')
        # ax_conversion.set_title('CATE Conversion vs Targeted Population')
        # st.pyplot(fig)

        # Initialize a figure object
        fig = go.Figure()

        # Define colors for each discount level and the random baseline
        colors = ['blue', 'green', 'yellow']
        random_line_dash = 'dash'

        # Iterate over each discount to add its line to the plot
        for i, discount in enumerate(discounts):
            # Add the model line
            fig.add_trace(go.Scatter(x=qini_conversions[discount]['index'], 
                                    y=qini_conversions[discount]['S'], 
                                    mode='lines', 
                                    name=f'{discount} model',
                                    line=dict(color=colors[i])))
            
            # Add the random baseline line
            fig.add_trace(go.Scatter(x=qini_conversions[discount]['index'], 
                                    y=qini_conversions[discount]['Random'], 
                                    mode='lines', 
                                    name=f'{discount} random',
                                    line=dict(color='red', dash=random_line_dash)))

        # Update the layout of the figure
        fig.update_layout(
            title='CATE Conversion vs Targeted Population',
            xaxis_title='Fraction of Targeted Users',
            yaxis_title='CATE Conversion',
            legend_title='Legend',
            legend=dict(
                x=0,
                y=1,
                traceorder='normal',
                font=dict(
                    size=10,
                )
            )
        )

        # Display the figure in Streamlit
        st.plotly_chart(fig)

        # Plotting CATE Benefit
        st.subheader("CATE Benefit vs Targeted Population")

        # fig, ax_benefit = plt.subplots()
        # for discount, color in zip(discounts, ['b', 'g', 'y']):
        #     qini_benefits[discount].plot(ax=ax_benefit, x='index', y='S', color=color)
        #     qini_benefits[discount].plot(ax=ax_benefit, x='index', y='Random', color='r', ls='--')

        # ax_benefit.legend([f'{d} model' for d in discounts] + [f'{d} random' for d in discounts], prop={'size': 10})
        # ax_benefit.set_xlabel('Fraction of Targeted Users')
        # ax_benefit.set_ylabel('CATE Benefit')
        # ax_benefit.set_title('CATE Benefit vs Targeted Population')
        # st.pyplot(fig)


        # Initialize a figure object
        fig = go.Figure()

        # Define colors for each discount level and the random baseline
        colors = ['blue', 'green', 'yellow']
        random_line_dash = 'dash'

        # Iterate over each discount to add its line to the plot
        for i, discount in enumerate(discounts):
            # Add the model line
            fig.add_trace(go.Scatter(x=qini_benefits[discount]['index'], 
                                    y=qini_benefits[discount]['S'], 
                                    mode='lines', 
                                    name=f'{discount} model',
                                    line=dict(color=colors[i])))
            
            # Add the random baseline line
            fig.add_trace(go.Scatter(x=qini_benefits[discount]['index'], 
                                    y=qini_benefits[discount]['Random'], 
                                    mode='lines', 
                                    name=f'{discount} random',
                                    line=dict(color='red', dash=random_line_dash)))

        # Update the layout of the figure
        fig.update_layout(
            title='CATE Benefit vs Targeted Population',
            xaxis_title='Fraction of Targeted Users',
            yaxis_title='CATE Benefit',
            legend_title='Legend',
            legend=dict(
                x=0,
                y=1,
                traceorder='normal',
                font=dict(
                    size=10,
                )
            )
        )

        # Display the figure in Streamlit
        st.plotly_chart(fig)


        # Plotting CATE Benefit vs CATE Conversion
        st.subheader("CATE Benefit vs CATE Conversion")

        # fig, ax_comp = plt.subplots()
        # colors = ['b', 'g', 'y']
        # for i, discount in enumerate(discounts):
        #     qini_conc_test = pd.concat([qini_conversions[discount][['S']], qini_benefits[discount][['S']]], axis=1)
        #     qini_conc_test.columns = ['cate_conversion', 'cate_benefit']
        #     qini_conc_test.plot(ax=ax_comp, x='cate_conversion', y='cate_benefit', color=colors[i], label=f'{discount} model')
        
        # st.write('To simplify the comparison, we can plot the CATE Benefit as a function of the CATE conversion.')
        # st.write('In the last plot for example we can see that there is a region where offering 15% discount to a targeted group of users is more efficient than giving 10% to everyone. We can obtain the same impact in overall conversion uplift while reducing our benefit loss considerably.')

        # ax_comp.legend(prop={'size': 10})
        # ax_comp.set_xlabel('CATE Conversion')
        # ax_comp.set_ylabel('CATE Benefit')
        # ax_comp.set_title('CATE Benefit vs CATE Conversion')
        # st.pyplot(fig)

        # Initialize a figure object
        fig = go.Figure()

        # Define colors for each discount level
        colors = ['blue', 'green', 'yellow']

        # Iterate over each discount to add its scatter plot to the figure
        for i, discount in enumerate(discounts):
            qini_conc_test = pd.concat([qini_conversions[discount]['S'], qini_benefits[discount]['S']], axis=1)
            qini_conc_test.columns = ['cate_conversion', 'cate_benefit']

            # Add the scatter plot for each discount level
            # Adjust marker size with `size` and line width with `line=dict(width=2)`
            fig.add_trace(go.Scatter(x=qini_conc_test['cate_conversion'], 
                                    y=qini_conc_test['cate_benefit'], 
                                    mode='markers+lines', 
                                    name=f'{discount} model',
                                    marker=dict(color=colors[i], size=6),  # Adjust marker size here
                                    line=dict(width=2)))  # Adjust line width here

        # Update the layout of the figure to adjust aspect ratio and margins if needed
        fig.update_layout(
            title='CATE Benefit vs CATE Conversion',
            xaxis_title='CATE Conversion',
            yaxis_title='CATE Benefit',
            legend_title='Legend',
            legend=dict(
                x=0,
                y=1,
                traceorder='normal',
                font=dict(
                    size=10,
                )
            ),
            # Optionally adjust plot and margin size for a "thinner" appearance
            margin=dict(l=20, r=20, t=50, b=20),  # Adjust margins to change plot boundary
            height=400,  # Adjust height for overall "thinness"
            width=600   # Adjust width as needed
        )

        # Display the figure in Streamlit
        st.plotly_chart(fig)


        st.write('To simplify the comparison, we can plot the CATE Benefit as a function of the CATE conversion.')
        st.write('In the last plot for example, we can see that there is a region where offering a 15% discount to a targeted group of users is more efficient than giving 10% to everyone. We can obtain the same impact on overall conversion uplift while reducing our benefit loss considerably.')


    else:
        st.error("Please ensure the model is trained and the dataset is prepared.")