SEPTEMBER 2 :2024 CHENNAI

Customer Churn By

Prediction
Radhin Krishna R

22376003

BSC Data Science

Telecom customer Churn Prediction Using ANN and deploy it V th Semester

in Azure and display result using Tableau Dashboard.

CONTENTS

Introduction
Provide a brief description of the
project’s aim, prospects, and scope

Abstract
Predicting customer churn to enhance
retention strategies

Project Overview

Explaination about each stages about

the project execution

literature Review
Reviewing existing models and methods
for churn prediction

Introduction This project aims to develop a robust churn prediction system

capable of providing real-time insights into customer and employee
attrition. By integrating various components, including data
extraction, transformation, loading (ETL), KPI calculation, data
cleaning and preprocessing, feature selection, machine learning
model development, and visualization, this system will enable
organizations to make informed decisions and implement proactive
retention strategies. Our goal is not to become experts in a single tool
or method, but to excel as solution experts. That's why we've
integrated both Machine Learning (ML) and Artificial Neural Networks
(ANN) into this project. Although we select the one with superior
accuracy, time, and space complexity, we acknowledge the
potential of employing multiple ML algorithms through an ensemble
model. Presently, each model and dashboard is constructed using a
static dataset.
 Methodology:-
We aim to create a project that is reusable and flexible for all types
of churn analysis, whether it's customer, segment, or employee
churn. Each dataset varies by attributes and individual
characteristics, so a one-size-fits-all architecture isn't feasible.
However, we can develop a model that can analyze the data,
retrain itself when accuracy dips below a certain threshold, and
operate in a cloud environment for efficient resource
management. This approach will also allow the use of real-time
data, making it a versatile tool for various churn predictions within
an organization.

Customer Churn Prediction and Real-Time Analysis

Abstract This project aims to develop a comprehensive solution for

predicting customer churn in the telecommunications industry and
providing real-time insights for effective resource management
and policy evaluation. By leveraging machine learning techniques,
the project seeks to identify customers at risk of churning and offer
actionable recommendations to mitigate losses.

Key components of the project include:

• Data Extraction and Preprocessing: Extraction of relevant

data from various sources followed by cleaning,
normalization, and feature engineering to ensure data
quality and consistency.
• Machine Learning Model Development: Training and
evaluation of machine learning models, such as ANNs, to
accurately predict customer churn based on historical
data. The models will be carefully selected and optimized
to achieve the highest possible predictive performance.
• Real-Time Dashboard: Creation of an interactive
dashboard to visualize churn metrics, predictions, and
trends segmented by region, department, gender, and
other relevant factors. The dashboard will provide a clear
and intuitive overview of customer churn patterns and
enable timely decision-making.
• Cloud-Based Architecture: Design and implementation of a
scalable cloud architecture to enable real-time data
ingestion, model deployment, and prediction generation.
The architecture will be designed to handle large volumes
of data and ensure high availability and performance.
• Continuous Learning: Incorporation of mechanisms for
model retraining and updates based on new data to
ensure ongoing accuracy and relevance. The model will be
continuously evaluated and refined to adapt to changing
customer behavior and market dynamics.
The project has made significant progress in the ETL process, KPI
calculation, data cleaning, and feature selection. The next steps
involve integrating the components into a cloud environment,
selecting the optimal ML model, and developing a user-friendly
web application for accessing the predictions.

By successfully implementing this project, telecommunications

companies can gain valuable insights into customer behavior,
optimize resource allocation, and improve customer retention
 strategies. The project will provide a powerful tool for making data-
driven decisions and enhancing overall business performance

Project Overview
MODULE 1 : ETL

1.1 Data extraction

The provided Python code effectively extracts data from a remote API and saves it in a structured CSV
format, making it suitable for subsequent analysis and machine learning tasks. It starts by importing
essential libraries: requests for handling HTTP requests, csv for working with CSV files, and json for parsing
JSON data

The script then configures the API endpoint and API key, which are crucial for authentication and
identification. It constructs a dictionary of request parameters, including the API key and any other
required parameters. A GET request is sent to the API using the requests.get() function, and the response is
evaluated. If the request is successful (status code 200), the JSON data is parsed using the response.json()
method, converting it into a Python dictionary for easier manipulation.

The extracted data is then written to a CSV file using the csv module. A CSV writer object is created, and
the header row containing column names is written to the file. The loop iterates through each item (data
point) in the parsed JSON data, extracting the desired fields and writing them as a row to the CSV file. This
ensures that the data is organized in a tabular format for efficient analysis.

To handle APIs that return paginated results, the code can be extended to include a loop that iterates
through multiple pages until there are no more results. This is achieved by updating the page number in
the request parameters and checking for a "next_page" indicator in the API response.

The code incorporates error handling to gracefully deal with unsuccessful requests. If the API request
returns an error (non-200 status code), an informative error message is printed, and the program execution
is terminated. This prevents further processing of potentially corrupted or incomplete data.

1.2 Data Transformation

The initial phase of the project involved transforming the extracted telecom churn data within MS SQL
Server. Four tables were created to effectively organize and manage the data, ensuring efficient
analysis. A comprehensive data transformation process was conducted to ensure data quality and
consistency. This included cleaning the data by removing duplicates, handling null values, and
correcting errors, as well as converting data types to match their intended use. Additionally, data
standardization was implemented to ensure consistent formats and units across different columns and
tables. To enhance the predictive power of the model, feature engineering techniques were applied,
creating new features or transforming existing ones.

Once the data was transformed and cleaned, ten key performance indicators (KPIs) were defined to
gain valuable insights into the telecom churn dataset. These KPIs were designed to measure various
aspects of customer behaviour, service usage, and churn patterns. Examples include churn rate,
average revenue per user (ARPU), customer lifetime value (CLTV), contract length, usage patterns,
service bundle popularity, and customer satisfaction. SQL queries were leveraged to calculate these
KPIs within MS SQL Server, providing a comprehensive understanding of customer churn trends and
identifying areas for improvement.

This Data is saved for creating the first ANN model and analyse which all components are related and
formalize the ANN architecture and to be used as an initial phase of visualization in tableau.

A live connection from this Sql Server is established to Tableau and Azure Server.
Module 2: Data Pre-processing- Analysis

2.1 Feature engineering

Data Cleaning
• Removal of Irrelevant Column: The "Customer ID" column is dropped as it's unlikely to have a
direct impact on churn prediction.
• Handling Missing Values: Rows with missing values in the "Total Charges" column are identified
and removed to ensure data consistency and avoid potential errors in subsequent analysis.
Data Type Conversion
• Conversion of "Total Charges": To ensure compatibility with numerical operations, the "Total
Charges" column, initially a string, is converted to a numeric data type. This enables
calculations and statistical analysis on this feature.
Data Exploration and Visualization
• Churn Distribution: Histograms are created to visualize the distribution of tenure and monthly
charges for both churning and non-churning customers. This helps identify potential patterns or
differences between these groups.
Feature Encoding
• Handling Categorical Features:
o Replacement of Categorical Values: The values "No internet service" and "No phone
service" are replaced with "No" for consistency and to simplify feature encoding.
o Encoding of Binary Features: Columns representing binary features (e.g., "Married",
"Phone Service") are converted to numeric values (1 for "Yes", 0 for "No"). This allows the
model to directly process these features.
o One-Hot Encoding: For columns with more than two categories (e.g., "Offer"), one-hot
encoding is applied. This creates separate binary columns for each unique category,
preventing the model from assuming an ordinal relationship between categories.
Data Normalization
• Scaling Numerical Features: Features with a wide range of values, such as "Tenure in Months",
"Monthly Charge", and "Total Revenue", are scaled using MinMaxScaler. This ensures that
features with larger scales don't dominate the model's learning process, leading to more
balanced feature contributions.
Impact of Feature Engineering
• Improved Data Quality: Cleaning and handling missing values ensure that the model trains on
reliable and consistent data, reducing the risk of errors and biases.
• Enhanced Model Performance: Feature encoding and normalization contribute to improved
model performance by allowing the model to effectively process categorical data and avoid
biases due to varying feature scales.
• Better Model Interpretability: One-hot encoding makes it easier to understand the influence of
different categories on churn prediction, providing insights into the factors driving customer
churn.

2.2 Data Modeling & Analysis

The provided code performs a series of data visualizations, analyses, and feature engineering steps to
prepare a telecom customer churn dataset for machine learning model training. Here's a detailed
breakdown of these processes and their impact on understanding and predicting customer churn

Tenure vs. Churn: This visualization employs histograms to compare the distribution of "Tenure in Months"
for customers who churned (typically shown in red) and those who stayed (usually in green).
Analyzing this plot helps identify potential correlations between customer tenure and churn. For
instance, a significantly higher churn rate at a specific tenure might indicate areas needing
improvement in customer retention strategies. By visually comparing the distribution shapes, you
can assess if churn is more likely for newer or more established customers.

Impact: These visualizations provide an initial exploration of the data, revealing potential patterns
related to churn. They are instrumental in understanding general trends but may not pinpoint
specific factors influencing churn decisions. However, they can guide further analysis and feature
selection for the machine learning model.

Feature Analysis:

Unique Values: The print_unique_col_values function identifies the unique values present in each
categorical column. This helps understand the composition of the data and identify potential
inconsistencies in how certain categories are represented (e.g., "No internet service" vs. a more
standardized "No").

Impact: Feature analysis is crucial for determining how to handle categorical features during
model training. By understanding the unique values within each category, you can decide on
appropriate encoding techniques (e.g., one-hot encoding or label encoding) to transform these
features into a format suitable for the machine learning model.

Unique Values: The print_unique_col_values function identifies the unique values present in each
categorical column. This helps understand the composition of the data and identify potential
inconsistencies in how certain categories are represented (e.g., "No internet service" vs. a more
standardized "No").

Impact: Feature analysis is crucial for determining how to handle categorical features during
model training. By understanding the unique values within each category, you can decide on
appropriate encoding techniques (e.g., one-hot encoding or label encoding) to transform these
features into a format suitable for the machine learning model.

Replacing Missing Values: The code replaces specific values like "No internet service" and "No
phone service" with a simpler "No" to ensure consistency in the data. This step helps address
potential issues that might arise during model training due to inconsistencies in how missing or non-
standard values are represented within a category.

Converting Yes/No to Numeric: Columns representing binary features like "Married" are converted
to numeric (typically 1 for Yes and 0 for No). This transformation allows the model to understand
these features quantitatively. By converting these features from strings to numerical values, the
model can learn the relationships between these binary features and the target variable (customer
churn) more effectively.
more effectively.

One-Hot Encoding: Columns with more than two categories (e.g., "Offer") undergo one-hot
encoding. This technique creates separate binary columns for each unique category within the
original column. For instance, if "Offer" has three categories ("A", "B", and "C"), one-hot encoding
would result in three new binary columns, where a value of 1 indicates belonging to that specific
category and 0 indicates not belonging.

Impact: Feature encoding is a critical step in preparing categorical data for machine learning
models. By converting features and handling missing values consistently, the model can learn from
them more effectively. One-hot encoding, in particular, improves model interpretability as it allows
you to understand the influence of each individual category within a multi-category feature on the
model's predictions.

Module 3: Artificial Nueral network

The provided code defines a multi-layer perceptron (MLP) neural network for binary classification
using the Keras library in TensorFlow. MLPs are powerful tools for various tasks, excelling in finding
complex relationships between input features and a binary output (0 or 1).

Network Layers and Activation Functions:

• Input Layer:
o The first layer, keras.layers.Dense(26, input_shape=(70,), activation='relu'), acts as the
entry point for your data. It can handle input samples with 70 features (None, 70
shape). The None in the first dimension signifies that the batch size can be flexible
during training. This layer has 26 neurons, which are essentially processing units that
learn patterns from the data. The activation='relu' part introduces non-linearity
through the rectified linear unit (ReLU) activation function. ReLU allows the network to
learn more complex relationships between features compared to linear activation.
• Hidden Layer:
o The subsequent layer, keras.layers.Dense(15, activation='relu'), receives the processed
output from the input layer. It has 15 neurons, further refining the learned features. The
ReLU activation is applied again to maintain non-linearity, crucial for modeling real-
world problems that are rarely linear.
• Output Layer:
o The final layer, keras.layers.Dense(1, activation='sigmoid'), takes the output from the
hidden layer. This layer has a single neuron, responsible for generating the binary
classification output. The activation='sigmoid' function is well-suited for this task
because it maps any real number to a value between 0 and 1, representing the
probability of an instance belonging to the positive class (typically class 1).
Training and Evaluation:
• Compilation:

o The model.compile(...) section sets up the training process. The Adam optimizer is
employed with a learning rate of 0.01 to efficiently adjust the network's weights during
training. The loss='binary_crossentropy' function measures the difference between the
predicted probabilities and the true labels (0 or 1). It's suitable for binary classification
problems. Additionally, the metrics=['accuracy'] argument tracks the model's
accuracy during training, providing a basic performance indicator.

• Training:

o The model.fit(X_train, y_train, epochs=100) line trains the model on the provided
training data (X_train for features and y_train for labels) for 100 epochs (iterations).
During each epoch, the model iterates through the training data, calculates the loss,
and updates its internal weights to minimize the loss gradually. This process helps the
network learn the patterns in the data to make accurate predictions on unseen
 and updates its internal weights to minimize the loss gradually. This process helps the
network learn the patterns in the data to make accurate predictions on unseen examples.

Prediction and Thresholding:

• yp = model.predict(X_test) instructs the trained model to make predictions on the unseen

test data (X_test). The predictions are stored in the yp list. These predictions are initially in the
form of probabilities between 0 and 1 (due to the sigmoid activation).

• The subsequent code iterates through yp and creates a new y_pred list. It applies a
threshold of 0.5. Any value in yp greater than 0.5 is classified as 1 (positive class), and those
below 0.5 are classified as 0 (negative class). This is a common approach to convert the
model's probabilistic outputs into definitive class labels for binary classification tasks.

Evaluation Metrics:

• print(classification_report(y_test,y_pred)) and model.evaluate(X_test, y_test) assess the

model's performance on the test data. The classification_report provides a detailed
breakdown of metrics like precision, recall, F1-score, and support for each class, offering
valuable insights into the model's strengths and weaknesses. The model.evaluate function
calculates the built-in evaluation metrics (often accuracy and loss) on the test data, giving
a more concise performance summary.

F1-Score, Recall, Accuracy Calculation:

• The code extracts F1-score, recall, and accuracy from the classification report for further
analysis. These metrics provide complementary perspectives on the model's performance.
F1-score considers both precision (correctly predicted positives) and recall (correctly
identified positive cases), offering a balanced view. Recall focuses on how well the model
identifies all positive cases, while accuracy measures the overall proportion of correct
predictions.

Binary Crossentropy Calculation (using NumPy):

The code calculates binary crossentropy loss using NumPy:

• The predicted values (y_pred) are converted to a NumPy array for compatibility with
TensorFlow's loss functions.
• The tf.keras.losses.BinaryCrossentropy(from_logits=True) function is used to calculate the
binary crossentropy loss between the true labels (y_test) and the predicted probabilities
(y_pred_array). The from_logits=True argument indicates that the input values are logits (pre-
activation values) rather than probabilities.
• The numpy().mean() method calculates the average binary crossentropy loss across all
samples.
•
Creating a evaluation DataFrame:

• The code creates a Pandas DataFrame to store the calculated metrics for easy analysis
and visualization.
• The DataFrame includes columns for the predicted values, F1-score, recall, accuracy, and
binary crossentropy.
• The values are populated with the corresponding calculated metrics.

This DataFrame provides a structured way to organize and interpret the model's performance,
making it easier to evaluate the effectiveness of the neural network in the given binary
classification task.
Regularization Techniques
To prevent overfitting, apply regularization techniques such as L1, L2 regularization, or dropout.
Dropout randomly disables a fraction of neurons during training, which helps the model
generalize better by preventing it from relying too heavily on any single neuron.

Hyperparameter Tuning
Hyperparameters like learning rate, batch size, and the number of epochs can significantly affect
model performance. Use techniques like grid search or random search to find the optimal
hyperparameters. Tools like Keras Tuner can automate this process and help you find the best
configuration.

Optimization Algorithms
The choice of optimization algorithm can also impact the model’s accuracy. While stochastic
gradient descent (SGD) is a common choice, other algorithms like Adam, RMSprop, or Adagrad
might converge faster and lead to better performance. Experiment with different optimizers to
see which one works best for your model.

Data Augmentation
If you’re working with image data, data augmentation can help improve model accuracy by
artificially increasing the size of your training dataset. Techniques like rotation, scaling, and flipping
can create new training examples, helping the model generalize better
Module 4: Dash Board

The objective of this dashboard is to provide a comprehensive overview of customer churn, focusing
on various aspects like customer information, service usage, churn predictions, and account details.
The dashboard will be built using Tableau and will leverage data from a SQL server hosted on Azure
for real-time updates.
The dashboard will consist of five primary pages:
Overview Page:
• Display key performance indicators (KPIs) such as total churn rate, customer retention rate,
and average revenue per user (ARPU).
• Visualize the overall churn trend over time using a line chart.
• Analyze churn rates across different customer segments (e.g., demographic, geographic, or
service plan) using a bar chart
• Explore the correlation between churn and various factors (e.g., service usage, customer
satisfaction) using scatter plots or correlation matrices.
Customer Page:
• Show customer information like age, gender, and location using bar charts or pie charts.
• Visualize Customer Lifetime Value (CLTV) distribution using a histogram or density plot.
• Display customer segmentation based on clustering or other techniques using scatter plots or
color-coded maps.
• Track customer interactions and behaviors over time to identify potential churn indicators.
Service Page:
• Analyze service usage patterns (e.g., minutes used, data consumed) using line charts or
area charts.
• Track service outages and their impact on customer churn using bar charts or heatmaps.
• Display customer satisfaction ratings for different services using star ratings or bar charts.
• Examine the effectiveness of service bundles or promotions in reducing churn.
Prediction Page:
• Integrate the results from the neural network model into the dashboard.
• Visualize the predicted churn probability for each customer using a histogram or scatter plot.
• Rank customers based on their predicted churn risk using a bar chart or table.
• Evaluate the accuracy of the prediction model using metrics like precision, recall, and F1-
score.
Account Page:
• Analyze account-level activities (e.g., logins, payments) using line charts or time series
visualizations.
• Track contract expiration dates and potential churn risks using a calendar or timeline.
• Visualize customer support interactions (e.g., tickets, calls) using bar charts or heatmaps.
• Examine billing and payment patterns to identify potential issues that may lead to churn.
To ensure the dashboard displays the most current information, data will be extracted from the SQL
server on Azure in real-time. This can be achieved using Tableau's live data connection feature or by
scheduling data refreshes.
• Customer Segmentation: Experiment with different customer segmentation methods to
identify target groups for retention efforts.
• Customer Lifetime Value (CLTV) Analysis: Calculate and analyze CLTV to prioritize customer
retention efforts.
• Customer Journey Mapping: Visualize the customer journey to identify pain points and areas
for improvement.
• Geo-Spatial Analysis: If applicable, visualize data based on geographic location to identify
regional trends or patterns.
• Integration with Other Tools: Consider integrating the dashboard with other tools, such as
CRM systems or marketing automation platforms, for a more holistic view of customer
interactions.
• Automated Alerts: Set up automated alerts to notify relevant stakeholders of significant
changes in churn rates or other key metrics.
Modules 5: Azure Pipeline

stream Data to Event Hub

Azure Event Hubs is a big data streaming platform and event ingestion service. It can receive and
process millions of events per second.
• Create an Event Hub: In the Azure portal, create an Event Hub namespace and an Event Hub
within it.
• Send Data to Event Hub: Use an SDK or Azure CLI to send your streaming data to the Event
Hub.
Process Data with Stream Analytics
Azure Stream Analytics is a real-time analytics and complex event-processing engine.
• Create a Stream Analytics Job: In the Azure portal, create a Stream Analytics job.
• Input Configuration: Set the Event Hub as the input source for the Stream Analytics job.
• Query Definition: Write a Stream Analytics query to process the incoming data. For example,
you can filter, aggregate, or join data streams.
• Output Configuration: Set the output to Azure Synapse Analytics.
Store Processed Data in Synapse Analytics
Azure Synapse Analytics is a limitless analytics service that brings together big data and data
warehousing.
• Create a Synapse Workspace: In the Azure portal, create a Synapse workspace.
• Create a Dedicated SQL Pool: This will be used to store the processed data.
• Connect Stream Analytics to Synapse: Configure the Stream Analytics job to output data to
the Synapse SQL pool.
Archive Data in Blob Storage
Azure Blob Storage is Microsoft’s object storage solution for the cloud.
• Create a Storage Account: In the Azure portal, create a storage account and a blob
container within it.
• Configure Synapse to Archive Data: Set up a pipeline in Synapse to periodically move data to
Blob Storage for archival purposes.
Use Azure ML for Predictions
Azure Machine Learning is a cloud-based environment you can use to train, deploy, automate, and
manage machine learning models.
• Register Your Model: Register your trained ANN model in the Azure ML workspace.
• Create an Inference Pipeline: Set up an inference pipeline to use the model for predictions on
the incoming data.

• Deploy the Model: Deploy the model as a web service or a real-time endpoint.
Visualize Data with Power BI
Power BI is a business analytics service that provides interactive visualizations and business intelligence
capabilities.
• Connect Power BI to Synapse: Use the Power BI desktop to connect to your Synapse SQL pool.
• Create Reports and Dashboards: Build interactive reports and dashboards to visualize the
customer churn predictions and other insights.
Module 6: web App

First, ensure you have Python and Streamlit

installed on your machine. Streamlit is a powerful
tool for creating web apps with minimal effort,
making it perfect for this task.

Preparing Your Model

You already have a trained ANN model for

predicting customer churn. Save this model using
a library like joblib or pickle. This allows you to
load the model later in your Streamlit app for
making predictions.

Creating the Streamlit App

1. Import Libraries: Start by importing the

necessary libraries, including Streamlit,
pandas for handling data, and joblib for
loading your model.

2. App Layout: Design the layout of your

app. Begin with a title and a section to
display some sample data. This helps users
understand the kind of data your model
works with. Load a sample dataset and
display the first few rows.

3. User Input: Create input fields for users to

enter customer details. For example, you
might ask for age, income, and tenure.
Streamlit provides easy-to-use widgets for
this purpose, such as sliders and number
input fields.

4. Making Predictions: Add a button that,

when clicked, takes the user input,
processes it, and uses your ANN model to
make a prediction. The model will output
the probability of customer churn.

5. Displaying Results: Show the prediction

result to the user. This could be the
probability of churn expressed as a
percentage, giving users a clear
understanding of the risk.

Running the App

Once your app is set up, you can run it using a

simple command in your terminal. This will start a
local server, and you can interact with your app
through a web browser.

Summary

In summary, you’ll create a Streamlit web app

Project Significance

Project Significance

Current Methodologies in Churn Prediction

Churn prediction is a critical aspect of customer relationship management, aiming to identify

customers who are likely to discontinue using a service or product. Traditional methodologies
predominantly utilize machine learning models such as logistic regression, decision trees, support
vector machines (SVM), and ensemble learning techniques like random forests and gradient boosting.
These models analyze a plethora of historical customer data, including usage patterns, transaction
history, and demographic information, to pinpoint key indicators of churn. For example, logistic
regression might be employed to estimate the probability of a customer churning based on their past
behavior, while decision trees can elucidate the decision paths leading to churn. Despite their
effectiveness, these models often encounter significant challenges. Handling large volumes of data,
integrating data from multiple disparate sources, and providing real-time predictions are some of the
common hurdles. Additionally, these traditional models may struggle with capturing complex, non-
linear relationships within the data, potentially leading to less accurate predictions.

Leveraging Cloud-Based ANN and Dashboards

Your innovative cloud-based Artificial Neural Network (ANN) and dashboard system effectively
addresses several limitations inherent in traditional churn prediction methods. By harnessing the power
of cloud computing, your system can effortlessly manage vast amounts of data from various sources,
ensuring that the data remains up-to-date and is accessible from any location. The ANN, with its
advanced capability to learn intricate patterns and relationships within the data, offers more precise
and nuanced predictions compared to traditional models. This is particularly beneficial in capturing
complex, non-linear interactions that are often missed by conventional methods. Furthermore, the
integrated dashboard provides real-time insights and visualizations, enabling stakeholders to monitor
churn indicators continuously and take proactive measures swiftly. This combination not only enhances
the accuracy of predictions but also significantly improves decision-making efficiency. By offering a
comprehensive, real-time view of customer behavior, your system empowers businesses to implement
timely interventions, thereby reducing churn rates and enhancing customer retention.
Literature Survey

Customer Churn Prediction: A Comparative Analysis of Machine Learning Algorithms

Customer churn analysis and prediction in the telecom sector is an important area of research due
to the high cost of acquiring new customers compared to retaining existing ones (e.g.,
Senthilnayaki B, Swetha M, Nivedha D). Machine learning techniques such as Logistic Regression,
Random Forest, and K-Nearest Neighbors are commonly used to build churn prediction models

Authors: Senthilnayaki B, Senthilnayaki B, Senthilnayaki B.

This study aims to develop and evaluate the performance of various machine learning models for
predicting customer churn in the telecom industry. By comparing the accuracy, precision, recall,
and F1-score of Logistic Regression, Random Forest, and K-Nearest Neighbors algorithms, we aim
to identify the most effective model for predicting customer churn and providing valuable insights
for targeted retention strategies.

Enhancing Customer Loyalty in the Telecom Industry: A Machine Learning Approach

This paper, presented at the Highlights in Science, Engineering and Technology CMLAI
conference (Volume 39, 2023), investigates the application of machine learning algorithms for
predicting customer churn in the telecom industry.

Author:Pulin Yang, from Guangxi University of Science and Technology, China, emphasizes the
growing importance of customer retention strategies due to rising competition and increasing
churn rates. The study focuses on developing a churn prediction model that not only identifies
customers at risk of leaving but also prioritizes high-value customers for targeted retention efforts.

The research utilizes a telecom customer dataset from Kaggle to analyze usage patterns and
identify factors contributing to churn. By employing data visualization techniques, the study
explores trends and relationships within the data. Subsequently, the paper compares the
performance of three machine learning models: Random Forest, Support Vector Machine (SVM),
and Gradient Boosting Decision Tree (GBDT). The results demonstrate that the Random Forest
model achieves superior classification accuracy, particularly in predicting churn among high-
value customers. This finding suggests that Random Forest offers a valuable tool for telecom
operators to optimize their customer retention strategies and minimize churn rates.
Customer Churn Prediction: A Comparative Analysis of Ordinary Artificial Neural Networks and
Convolutional Neural Networks

Authors: Omer Faruk Seymen, Abdulkadir Hiziroglu İzmir ,Orhan Er

This paper, published in the Gazi University Journal of Science, explores the application of both
ordinary artificial neural networks (ANNs) and convolutional neural networks (CNNs) for predicting
customer churn in the retail sector. The study aims to compare the performance of these two
deep learning approaches and evaluate their effectiveness in handling the challenges
associated with imbalanced datasets and complex customer behavior patterns.

By transforming customer transaction and demographic data into an image format, the authors
leverage the unique capabilities of CNNs for feature extraction and classification. The
experimental results demonstrate that the CNN-based model outperforms traditional churn
prediction methods, highlighting the potential of deep learning techniques in enhancing
customer retention strategies.
Table 1 summarizes the related work in the field of customer churn prediction, highlighting the
performance metrics used and the best-performing models. As noted, most deep learning-based
studies in this area have focused on the telecommunications sector, where contractual
relationships exist between customers and companies.
Deep Learning Applications:
• Feedforward Neural Networks (FNNs): [7] compared FNNs (both small and large versions)
with ANNs and CNNs for telecom churn prediction, using two different datasets. Accuracy
was the primary metric used for comparison.
• Convolutional Neural Networks (CNNs) and Long Short-Term Memory (LSTM): [13]
developed a daily churn prediction model using CNN and LSTM, outperforming RFM-
based models. The models were evaluated using AUC, F1 Score, Lift, Log Loss, and EPMC.
• CNN for Image-Based Prediction: [14] transformed call details, SMS, Recharge, and Data
usage records into images and used CNN for churn prediction. Precision, recall, and F1-
scores were compared against SVM, Random Forest, and Gradient Boosting Classifier.
• CNN with Natural Language Processing (NLP): [12] combined CNN with NLP to analyze
customer call transcripts and predict churn. Precision, recall, and F1-score were used for
evaluation.
Other Deep Learning Studies:
• CNN and Deep CNN: [15, 17] used CNN and Deep CNN for telecom customer churn
prediction. They evaluated performance using Confusion Matrix outputs, accuracy, AUC, loss
margin, maximum dice coefficient, and Jaccard coefficient.
• Vector Embedding Model: [18] used a Vector Embedding Model for loss estimation in a
telecom dataset. Accuracy and F1-score were reported.
• CNN in Financial Services: [19, 20] applied CNN models to predict churn in the financial
service domain. [19] used textual and structural data, while [20] compared RNN and LSTM
models. Evaluation metrics included Precision, AUC, and TDL.
• Deep Neural Network (DNN): [21] compared DNN with MPL for banking customer churn
prediction. Accuracy and loss margin were used, along with different activation functions
and training algorithms.
Other Machine Learning Approaches:
• SVM with Bayesian Optimization: [22] optimized SVM kernels using Bayesian Optimization.
Linear, Polynomial, Radial, and Sigmoid kernels were compared using accuracy, precision,
recall, and F1-score.
• Extended Convolutional Decision Tree (ECDT): [23] applied ECDT for retail employee churn
prediction. Grid Search Optimization was used to improve accuracy, which was compared
against ECDT, KNN, CNN, SVM, NB, and DT.
• Machine Learning Models for Online Gaming: [24-26] used various machine learning models
to predict churn in the online gaming industry, based on player behavior data. Accuracy,
AUC, and F1-score were used for evaluation.
CNN Enhancements:
• Shallow-Based CNN: [27] combined deep and shallow learning models for churn prediction.
• LSTM Integration: [28] added LSTM output as an input to a CNN model.
• Static Data Integration: [8] integrated static data (demographics) with dynamic features
(RFM and relationship length) in a CNN model.
Retail Domain Studies:
• LSTM for Smartphone Buyer Churn: [29] used LSTM to predict churn based on questionnaire
responses from smartphone buyers. AUC, Precision, and Recall were used for evaluation.
• CNN and Restricted Boltzmann Machine: [30] compared CNN and Restricted Boltzmann
Machine for customer churn prediction in the retail domain. Various performance metrics
were used for evaluation.
• While there are studies on data mining and machine learning in the retail domain, deep
learning applications are still relatively limited.
Conclusion

This project successfully developed a sophisticated real-time customer churn prediction system,
demonstrating the power of Artificial Neural Networks (ANNs) and the scalability of Azure cloud
infrastructure. The system, equipped with a comprehensive dashboard and a user-friendly web
application, accurately identifies at-risk customers, empowering businesses to proactively address
churn and enhance customer retention.

By leveraging the capabilities of ANNs, the model effectively predicts customer churn based on a
diverse set of factors, including historical usage patterns, demographic data, and recent customer
interactions. The Azure platform provides a scalable and resilient environment for real-time data
processing and model deployment, ensuring the system's ability to handle increasing workloads and
adapt to evolving customer behaviors.

The integrated dashboard offers a clear visualization of key performance indicators, enabling
stakeholders to monitor the system's effectiveness and identify emerging trends. The web
application provides a user-friendly interface for accessing churn predictions, empowering
businesses to take timely actions to retain their valuable customers.

Beyond its immediate benefits, this project serves as a proof of concept for the application of
advanced machine learning techniques in customer retention. The system's success highlights the
potential of AI-driven solutions to transform business operations and drive growth. Future
enhancements could include incorporating additional features, such as customer sentiment analysis
and predictive marketing campaigns, to further refine the churn prediction capabilities and provide
even more valuable insights to businesses.

In conclusion, this project demonstrates the efficacy of ANNs in predicting customer churn and the
advantages of deploying such models on cloud platforms like Azure. The system has the potential to
significantly enhance customer retention rates and drive business growth. By harnessing the power
of AI and leveraging the scalability of cloud technology, businesses can gain a competitive edge
and deliver exceptional customer experiences.