Customer churn prediction

# Standard import pandas as pd import numpy as np # Pycaret from pycaret.classification import * # Plots from plotly.offline import iplot import plotly.graph_objects as go import plotly.figure_factory as ff from plotly.subplots import make_subplots import seaborn as sns # Sklearn tools from sklearn.preprocessing import LabelEncoder from sklearn.metrics import * # Extras from datetime import date import warnings warnings.filterwarnings("ignore") # Datapath and Setup data_path = "/datasets/telco-customer-churn/" random_seed = 142

# Helper functions for structured data ## Get info about the dataset def dataset_info(dataset, dataset_name: str): print(f"Dataset Name: {dataset_name} | Number of Samples: {dataset.shape[0]} | Number of Columns: {dataset.shape[1]}") print(30*"=") print("Column Data Type") print(dataset.dtypes) print(30*"=") missing_data = dataset.isnull().sum() if sum(missing_data) > 0: print(missing_data[missing_data.values > 0]) else: print("No Missing Data on this Dataset!") print(30*"=") print(f"Memory Usage: {np.round(dataset.memory_usage(index=True).sum() / 10e5, 3)} MB") ## Dataset Sampling def data_sampling(dataset, frac: float, random_seed: int): data_sampled_a = dataset.sample(frac=frac, random_state=random_seed) data_sampled_b = dataset.drop(data_sampled_a.index).reset_index(drop=True) data_sampled_a.reset_index(drop=True, inplace=True) return data_sampled_a, data_sampled_b ## Bar Plot def bar_plot(data, plot_title: str, x_axis: str, y_axis: str): colors = ["#0080ff",] * len(data) colors[0] = "#ff8000" trace = go.Bar(y=data.values, x=data.index, text=data.values, marker_color=colors) layout = go.Layout(autosize=False, height=600, title={"text" : plot_title, "y" : 0.9, "x" : 0.5, "xanchor" : "center", "yanchor" : "top"}, xaxis={"title" : x_axis}, yaxis={"title" : y_axis},) fig = go.Figure(data=trace, layout=layout) fig.update_layout(template="simple_white") fig.update_traces(textposition="outside", textfont_size=14, marker=dict(line=dict(color="#000000", width=2))) fig.update_yaxes(automargin=True) iplot(fig) ## Plot Pie Chart def pie_plot(data, plot_title: str): trace = go.Pie(labels=data.index, values=data.values) layout = go.Layout(autosize=False, title={"text" : plot_title, "y" : 0.9, "x" : 0.5, "xanchor" : "center", "yanchor" : "top"}) fig = go.Figure(data=trace, layout=layout) fig.update_traces(textfont_size=14, marker=dict(line=dict(color="#000000", width=2))) fig.update_yaxes(automargin=True) iplot(fig) ## Histogram def histogram_plot(data, plot_title: str, y_axis: str): trace = go.Histogram(x=data) layout = go.Layout(autosize=False, title={"text" : plot_title, "y" : 0.9, "x" : 0.5, "xanchor" : "center", "yanchor" : "top"}, yaxis={"title" : y_axis}) fig = go.Figure(data=trace, layout=layout) fig.update_traces(marker=dict(line=dict(color="#000000", width=2))) fig.update_layout(template="simple_white") fig.update_yaxes(automargin=True) iplot(fig) # Particular case: Histogram subplot (1, 2) def histogram_subplot(dataset_a, dataset_b, feature_a: str, feature_b: str, title: str, title_a: str, title_b: str): fig = make_subplots(rows=1, cols=2, subplot_titles=( title_a, title_b ) ) fig.add_trace(go.Histogram(x=dataset_a[feature_a], showlegend=False), row=1, col=1) fig.add_trace(go.Histogram(x=dataset_b[feature_b], showlegend=False), row=1, col=2) fig.update_layout(template="simple_white") fig.update_layout(autosize=False, title={"text" : title, "y" : 0.9, "x" : 0.5, "xanchor" : "center", "yanchor" : "top"}, yaxis={"title" : "<i>Frequency</i>"}) fig.update_traces(marker=dict(line=dict(color="#000000", width=2))) fig.update_yaxes(automargin=True) iplot(fig) # Calculate scores with Test/Unseen labeled data def test_score_report(data_unseen, predict_unseen): le = LabelEncoder() data_unseen["Label"] = le.fit_transform(data_unseen.Churn.values) data_unseen["Label"] = data_unseen["Label"].astype(int) accuracy = accuracy_score(data_unseen["Label"], predict_unseen["Label"]) roc_auc = roc_auc_score(data_unseen["Label"], predict_unseen["Label"]) precision = precision_score(data_unseen["Label"], predict_unseen["Label"]) recall = recall_score(data_unseen["Label"], predict_unseen["Label"]) f1 = f1_score(data_unseen["Label"], predict_unseen["Label"]) df_unseen = pd.DataFrame({ "Accuracy" : [accuracy], "AUC" : [roc_auc], "Recall" : [recall], "Precision" : [precision], "F1 Score" : [f1] }) return df_unseen # Confusion Matrix def conf_mat(data_unseen, predict_unseen): unique_label = data_unseen["Label"].unique() cmtx = pd.DataFrame( confusion_matrix(data_unseen["Label"], predict_unseen["Label"], labels=unique_label), index=['{:}'.format(x) for x in unique_label], columns=['{:}'.format(x) for x in unique_label] ) ax = sns.heatmap(cmtx, annot=True, fmt="d", cmap="YlGnBu") ax.set_ylabel('Predicted') ax.set_xlabel('Target'); ax.set_title("Predict Unseen Confusion Matrix", size=14);

dataset = pd.read_csv(data_path+"customers.csv") dataset.head(3)

dataset[dataset.duplicated()]

dataset_info(dataset, "customers")

dataset["TotalCharges"] = pd.to_numeric(dataset["TotalCharges"], errors="coerce") print(f"The Feature TotalCharges is type {dataset.TotalCharges.dtype} now!")

pie_plot(dataset["Churn"].value_counts(), plot_title="<b>Client Churn Distribution<b>")

df_aux = dataset.query('Churn == "No"') df_aux = df_aux["Contract"].value_counts() bar_plot(df_aux, "<b>Contract Types of not Churned Clients</b>", "<i>Contract</i>", "<i>Count</i>")

df_aux = dataset.query('Churn == "Yes"') df_aux = df_aux["Contract"].value_counts() bar_plot(df_aux, "<b>Contract Types of Churned Clients</b>", "<i>Contract</i>", "<i>Count</i>")

df_aux = dataset.query('(Contract == "Month-to-month") and (Churn == "No")') histogram_subplot(df_aux, df_aux, "MonthlyCharges", "TotalCharges", "<b>Charges Distribution for Month-to-month contracts for not Churned Clients</b>", "(a) Monthtly Charges Distribution", "(b) Total Charges Distribution")

df_aux = dataset.query('(Contract == "Month-to-month") and (Churn == "Yes")') histogram_subplot(df_aux, df_aux, "MonthlyCharges", "TotalCharges", "<b>Charges Distribution for Month-to-month contracts for Churned Clients</b>", "(a) Monthtly Charges Distribution", "(b) Total Charges Distribution")

df_aux = dataset.query(('Contract == Month-to-month') and ('Churn == "Yes"')) df_aux = df_aux["PaymentMethod"].value_counts() bar_plot(df_aux, "<b>Payment Method of Month-to-month contract Churned Clients</b>", "<i>Payment Method</i>", "<i>Count</i>")

df_aux = dataset.query(('Contract == Month-to-month') and ('Churn == "Yes"')) df_aux = df_aux["tenure"].value_counts().head(5) bar_plot(df_aux, "<b>Tenure of Month-to-month contract for Churned Clients</b>", "<i>Tenure</i>", "<i>Count</i>")

data, data_unseen = data_sampling(dataset, 0.9, random_seed) print(f"There are {data_unseen.shape[0]} samples for Unseen Data.")

exp01 = setup(data=data, target="Churn", session_id=random_seed, ignore_features=["customerID"], numeric_features=["SeniorCitizen"], normalize=True, feature_selection=True, remove_outliers=True, remove_multicollinearity=True, transformation=True, ignore_low_variance=True, pca=True, bin_numeric_features=["MonthlyCharges", "TotalCharges"])

compare_models(fold=10, sort="F1")

base_alg = "qda" base_model = create_model(base_alg)

plot_model(base_model, plot="parameter")

tuned_model = tune_model(base_alg, fold=10, optimize="F1")

plot_model(tuned_model, plot="parameter")

bagged_model = ensemble_model(tuned_model)

blended_model = blend_models(estimator_list=[tuned_model, bagged_model], method="soft")

stacked_model = stack_models([tuned_model, bagged_model], method="soft")

best_model = blended_model plot_model(best_model, plot="auc")

predict_model(best_model);

final_model = finalize_model(best_model)

predict_unseen = predict_model(final_model, data=data_unseen); score_unseen = test_score_report(data_unseen, predict_unseen) print(score_unseen.to_string(index=False)) conf_mat(data_unseen, predict_unseen);

save_experiment("expQDA_"+date.today().strftime("%m-%d-%Y")) save_model(tuned_model, "modelQDA_"+date.today().strftime("%m-%d-%Y"))