See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/378067630

CusCP: AI-Driven System for Predictive Modeling of Customer Churn in E-

commerce Using Machine Learning

Conference Paper · December 2023

DOI: 10.1109/ICIIP61524.2023.10537663

CITATIONS READS

0 125

2 authors, including:

Shantanu Neema
Georgia Institute of Technology
15 PUBLICATIONS 16 CITATIONS

SEE PROFILE

All content following this page was uploaded by Shantanu Neema on 08 February 2024.

The user has requested enhancement of the downloaded file.

 CusCP: AI-Driven System for Predictive Modeling
of Customer Churn in E-commerce Using Machine
Learning
Jagdeep Sharma Shantanu Neema
School of Management and Commerce, Senior Data Scientist,
RIMT University, Syntelli Solutions Inc.,
Mandi Gobindgarh, India Charlotte, United States
sharmajagdeep001@gmail.com shantanuneema@gmail.com

Abstract— Client churn has emerged as a key worry for Businesses may identify at-risk clients and anticipate possible
firms looking to maintain and grow their client base in the ever churn with the use of statistical analysis and customer
changing world of e-commerce. In order to anticipate customer segmentation. Businesses may modify their strategy to retain
turnover, this study takes a comprehensive approach and important clients by examining purchase history, behaviour,
evaluates the effectiveness of four different predictive models:
decision trees, neural networks, logistic regression, and NN-
and feedback. The main reasons of churn may be found by
Keras. To build a solid dataset, the research starts with actively engaging with consumers and soliciting their input.
meticulous data collecting and preparation. We thoroughly It also gives the chance to talk about client problems and
examine these models to determine their individual advantages suggest solutions. Reactivation campaigns are used by e-
and disadvantages. The distribution of true positives and true commerce companies to entice former clients back by
negatives offered by logistic regression is balanced, but it providing rewards, savings, or a reminder of the advantages
struggles with a large proportion of false negatives. Decision of their goods or services. In order to evaluate the
trees are excellent at detecting genuine positives but have a effectiveness of your e-commerce firm, it is crucial to track
greater tendency to detect false positives. While NN-Keras churn rates through time and compare them to industry
favors genuine negatives, neural networks establish a healthy
equilibrium between real positives and true negatives. The
standards [2]. For e-commerce enterprises to be successful in
research gives E-commerce companies insightful information on the long run, client attrition must be decreased. Businesses
how to choose a model while taking into account the trade-offs may strengthen customer loyalty, boost their bottom line, and
between accuracy, recall, and interpretability. This research succeed in a cutthroat online market by comprehending why
paper emphasizes the necessity of continual model improvement consumers leave and putting into practice efficient retention
and the investigation of ensemble tactics to strengthen client measures.
retention strategies and improve prediction skills as the e- Depending on the sector, business strategy, and factors
commerce industry continues to change. such as the causes for the churn, customer turnover may have
a major impact on a company. Here are a few of the most
Keywords— Customer Churn, E-Commerce, Prediction,
Logistic Regression, Neural Network, Decision Tree important consequences of client turnover in a company. The
revenue of a corporation is directly impacted by customer
attrition. Customers who no longer purchase goods or
I. INTRODUCTION services see a decline in sales, which impacts revenue. Losing
The phenomenon where consumers who used to make long-term clients may be especially expensive since they tend
transactions on an online shopping site or website stop doing to earn more income the longer they do business with a
so is known as customer retention in online shopping, also company. It is sometimes more costly to acquire new clients
known as client attrition or customer turnover. It's a crucial than to keep your current ones. Businesses may have to spend
indicator for e-commerce companies since it's often less more money on finding new clients when customer turnover
expensive to keep current consumers than to get new ones. rates are high, which may be expensive. The profitability of
To maintain and expand an e-commerce firm, it is crucial to a corporation may suffer as a result of turnover. Profit
comprehend and manage client turnover. Crucial things to margins may narrow when revenue declines and client
think about when it comes to client turnover in e-commerce. acquisition expenses increase, making it harder for a
The proportion of consumers that no longer use your online company to sustain strong financial performance. The brand
shopping site or make transactions during a certain period of and goodwill image of a firm may be harmed by high
time, such as one month or a year, is generally used to customer turnover [3]. Customers who are not happy could
calculate customer churn [1]. Consumers churn in e- complain to others, which can deter prospective new clients
commerce for a variety of reasons, such as unhappiness with from doing business with the company. A company's market
the good or service in question, better offers from rivals, share and competitive edge may be eroded through customer
modifications in their lives, or even lack of interest. High attrition. When clients go to rivals, the company loses sales
churn rates may be bad for an online store since they can lead and provides competitors a chance to expand their share of
to lost sales, higher user acquisition expenses, and worse the market. The typical customer's lifetime value may be
overall profitability. Many different tactics are often used by dramatically decreased by high churn rates. The total income
e-commerce companies to lower client turnover. These might a company may anticipate from a client over the course of the
include user experience improvements, exceptional customer relationship is measured by the CLV statistic. When turnover
service, personalized marketing, and loyalty programs. is significant, CLV declines, which has an impact on the

XXX-X-XXXX-XXXX-X/XX/$XX.00 ©20XX IEEE

company's long-term financial stability. The amount of churn provide personalized offers, suggestions, and
might provide insightful feedback. A company might find communications to at-risk clients to entice them to remain by
opportunities for improvement in the quality of the product, analysing customer data. Machine learning may assist in the
customer service, price, or other facets of the consumer creation and analysis of A/B testing to identify the tactics that
experience by understanding the reasons why consumers are most successful at lowering churn. It may reveal which
leave. Customer turnover may result in fierce rivalry over a initiatives are most effective for certain client categories. On
single pool of prospective consumers in businesses that are the basis of recent customer interactions, deep learning
highly competitive. Customers often have a wide range of models, particularly LSTMs, may forecast churn in real-time,
options when switching providers, thus it is vital for allowing for prompt intervention. To acquire insights into the
companies to stand out. As organizations always need to causes of churn, it might be helpful to analyse customer
modify their strategy, marketing initiatives, and customer feedback, evaluations, and comments using sentiment
service in order to acquire and maintain new consumers, assessment and natural language processing (NLP)
frequent churn may result in operational interruptions. High approaches. Machine learning models may be updated and
churn rates might make investors lose faith in a company [4]. improved often, including fresh information and feedback to
Consistently high customer turnover may result in diminished raise the accuracy of churn prediction over time. Based on
value to shareholders and investor support since investors model projections, online retailers may adopt certain
often consider customer retention as an indication of a retention methods for consumers who are at risk, such as
successful and sustainable firm. Businesses must control discounts, top-notch customer support, or the implementation
client turnover if they want to remain profitable and expand. of loyalty programmes. Machine learning may assist
Customer service enhancements, product or service organisations in making the most use of their resources by
improvements, the implementation of loyalty programs, and focusing their attention on the clients most likely to leave,
the use of data analytics to better understand customer lowering expenses, and enhancing the results of retention
behaviour and preferences are all possible strategies for initiatives. E-commerce companies may proactively detect
minimizing churn. Businesses may lessen these negative and handle client churn by using the effectiveness of machine
consequences and build a more reliable and prosperous learning as well as deep learning, which will enhance
company by addressing the root reasons for churn and customer retention and income. By assisting companies in
concentrating on client retention. better-comprehending consumer behaviour, these tactics
facilitate the use of data in decision-making and the creation
II. DECODING E-COMMERCE CUSTOMER CHURN: AI'S of churn avoidance initiatives.
SECRET SAUCE
In e-commerce, anticipating and controlling client III. LITERATURE REVIEW
attrition is significantly aided by artificial intelligence and In this work, the authors used k-means segmentation of
deep learning. By using these strategies, organisations may clients and support vector machine (SVM) prediction to
analyse huge amounts of data and get insightful knowledge forecast customer turnover in B2C e-commerce, successfully
that can be used to spot consumers who are at danger of resolving the difficulties presented by longitudinal periods and
leaving. They assist in foreseeing and resolving client many data factors. Additionally, they assessed how well SVM
turnover in the following ways. Large volumes of data from prediction and logistic regression worked for developing
many sources, including as transaction history, consumer clients and discovered that SVM prediction performed more
behaviour, demographics, and more, may be processed and accurately than logistic regression. Their study fits in with the
larger field trend, since machine learning techniques have
analysed using machine learning models [5]. In order to build
shown potential in predicting customer turnover across
useful input characteristics for predictive models, feature
several sectors, including support vector machines, decision
engineering entails choosing and manipulating relevant data trees, and artificial neural networks. They also looked at the
elements, including purchase frequency, average value of use of combinatorial classifiers, such as XGBoost and
orders, and client engagement indicators. Forecasting models AdaBoost, which combine a number of weak classifiers into a
for customer attrition may be created using machine learning powerful predicting tool that is especially useful for data with
techniques like logistic regression, random forests, decision historical characteristics. Through the use of sophisticated
trees, and support vector machines. Deep learning predictive modelling tools, the research offers useful insights
approaches, like as neural networks, are highly suited for into improving client retention tactics in the e-commerce
predicting customer churn because they can capture intricate industry [7]. The authors used data mining methods to analyse
patterns in time-series information, particularly recurrent a set of data from a Tehran, Iran, food ordering online service
artificial neural networks (RNNs) as well as short-term for their study. Multiple experiments were conducted as part
mental networks (LSTMs). Based on behaviour, tastes, and of their research to assess different data mining techniques,
other traits, machine learning may segment the client base. with a focus on forecasting customer turnover by using web
Using segmentation enables more focused churn prevention characteristics and user behaviour. Gradient Boosted Trees,
measures. It is possible to train machine learning algorithms which achieved an excellent accuracy rate of 86.90%, stood
to spot unusual client behaviour. A dramatic decrease in order out as the most accurate approach among the other ones
frequency or other unusual patterns may indicate impending examined.
churn [6]. Individual consumers may be given churn risk In addition to shedding light on the dynamics of customer
ratings by models. Concentrating on high-risk consumers churn within the context of an online food ordering service,
who are more inclined to churn, enables organisations to this research also highlights the efficacy of The gradient
prioritise their efforts. The marketing strategy may be tailored Boosted Trees as a predictive tool for enhancing retention
thanks to machine learning. E-commerce platforms may tactics in this sector, providing insightful information for
similar businesses looking to lower churn and raise customer effectiveness in improving model performance. The research
satisfaction. The primary approach used in this study to also emphasises the critical necessity for clear standards on
forecast customer turnover in the banking sector is Artificial the assessment of model performance indicators.
Neural Networks (ANN). Using layers of linked nodes or Predominantly quantitative methodologies are used in all of
neurons, ANNs are skilled at digesting incoming data, the research that have been analysed, emphasising the value
allowing operations like extraction of features, recognition of of numerical information in predictive modelling. The lack of
patterns, regression, and classification. The Rectified Linear studies concentrating on the extraction of information-rich
Unit (ReLU) and Mean Squared Error (MSE) are the material from customer-company interactions, including
activation and loss functions, respectively, that the authors channels like emails, phone conversations, or customer
apply in this work [8]. They use the Adam Optimizer to service chat logs, however, stands out as a significant research
enhance the model's prediction skills. The authors develop gap. This discrepancy suggests a promising direction for
three unique models to investigate the relationship between future study and the possibility of using underutilised data
characteristic data and labels, making it easier for them to be sources to improve machine learning-driven customer-
applied to fresh data via mathematical calculations. The focused approaches and analytics across a wide range of
research uses supervised training, whereby models are trained sectors [13]. In the suggested technique, customer turnover is
with data that will be anticipated as input and output. These analysed and predicted over the course of various stages using
models have a default rate of learning of 0.001 and a variety a systematic manner. The procedure begins with preliminary
of layers and nodes, enabling alterations to be made depending data processing and feature analysis, followed by feature
on what is needed prediction accuracy. This study digs into selection using the gravitational search method in the third
the banking industry's use of ANNs as a powerful tool for stage. After that, the dataset is split into test and training sets
churn prediction, providing insights into model design and in an 80/20 split. Several prediction models are used to the
methods to strengthen client retention tactics. The report training set, including logistical regression, naive bayes
undertook a thorough examination of the literature, looking at models, support vector machines, random forests, and
40 relevant publications that were gathered from Google decision trees. Boosting and ensemble modelling approaches
Scholar rankings and had been released between 2010 and are also used to improve model accuracy. For hyperparameter
June 2020 [9]. Each publication was subjected to a thorough adjustment and to reduce the danger of overfitting, K-fold
review across six important aspects, including references, cross-validation is used. AUC curve and a confusion matrix
study fields, primary goals, dataset sources, applied are used to analyse the outcomes and provide a thorough
approaches, and research findings. The thorough review of evaluation of model performance. Surprisingly, the Adaboost
this corpus of information revealed important conclusions, and the XGBoost classifier come out on top, with the greatest
emphasising the widespread use of particular data mining accuracy ratings of 81.71% and 80.8%, as well as the best
methods. Notably, the most often used approaches in the AUC score of 84%. With an emphasis on using modern
research under evaluation were decision tree models (DT), machine learning algorithms and methodologies, this study
support vector machines (SVM), and logistical regression provides useful insights into efficient tactics for anticipating
(LR). In addition to giving an in-depth account of the state of and controlling customer turnover [14]. In this study, the
the field's research, this review emphasises the importance of authors develop an efficient early-churn-detection method
these methods for data mining in relation to addressing a specifically for the telecoms sector using a stacked data
variety of research objectives and goals across a variety of mining strategy.
datasets, providing insightful information for upcoming
studies as well as applications in based on data domains. In They specifically integrate support vector machines
this work, customer attrition prediction is investigated using (SVMs) alongside the chi-square automatic contact detection
information from the telecoms sector [10]. (CHAID) tree of choices, demonstrating that this combination
is the best option for churn prediction [15]. The four main
The study uses artificial neural network (ANN) and phases of the data mining procedure in this work are model
random forest (RF) approaches to identify the variables learning, model assessment, and decision-making
affecting customer attrition [11]. An innovative hybrid ANN- improvement. To assure data quality, data preparation entails
based technique for Customer Churn Predict (CCP) is removing raw data from a variety of sources and fixing any
presented in the research, which is noteworthy. Three problems and inaccuracies [16]. Then, model learning makes
models—RF, ANN with four layers that are concealed, and use of a variety of techniques to reveal hidden patterns and
ANN that has two hidden layers—have their classification insights in the dataset. The effectiveness of the models is
accuracy compared in a comparison experiment. The results evaluated using metrics like accuracy, with the improvement
show that the suggested approach of using four hidden layers of how decisions are made as the main goal. This study
in an ANN achieves the greatest accuracy rate, achieving an highlights the possibility for combined data mining methods
outstanding 90.34%, exceeding the performance of using two and their influence on decision-making for retention tactics,
hidden layers in an ANN (88.14%) and RF [12]. This study offering a viable method for early churn identification in the
not only demonstrates the value of ANN and RF in predicting telecom industry.
customer turnover in the telecom industry, but it also
emphasises the effectiveness of the innovative ANN design, IV. PROPOSED MODEL OF CUSTOMER CHURN PREDICTION IN
offering useful insights for companies looking to improve E-COMMERCE
customer retention tactics in this field. In order to shed light
on important developments in the field, this study offers a The study's method of e-commerce research Numerous
thorough overview of the popular machine-learning structured processes must be followed in order to predict
approaches used by academics in recent years. In particular, it client attrition. Data collection is the laboriously assembling
highlights the importance of hybrids and ensemble approaches of customer information from many sources, including
in a variety of applications and highlights their considerable transactional and demographic information. The second stage
of data preparation involves meticulous dataset curation. This A. Training, Testing and Validation Accuracy
process includes the detection and removal of outliers, the Insights on the performance of the four predictive models—
management of values that are missing, and the modification Logistic Regression, Decision Tree, Neural Network, and
of variables to ensure the consistency and quality of the data. NN-Keras—in the context of predicting customer turnover in
The prediction models being examined are listed below: e-commerce are provided by the assessment of the models.
Neural networks come in many forms, including decision The accuracy measures for logistic regression are stable and
trees, regular neural networks, keras-based neural networks, consistent, with balanced performance on the training, test,
and logistic regression. The fundamental algorithms, and validation datasets. Because of its superior
structures, and predictive analytic techniques of each model interpretability, it is a practical option where model
are fully described. The experimental design is rigorous and transparency is crucial. The Decision Tree model, in
entails ensuring class balance and minimizing potential bias comparison, has a remarkable training accuracy,
while properly separating the information being studied into demonstrating its ability to recognize complex patterns in the
learning and evaluation groups. The model's training phase training data. On the test and validation sets, this model does,
comes next, during which hyperparameters are carefully however, show a little decline in accuracy, indicating some
adjusted and convergence criteria are constantly monitored. overfitting. Although it is skilled at capturing intricate
We check the models using a three-pronged approach. First, decision boundaries, it has to be pruned carefully to achieve
the Train Accuracy serves as an internal benchmark and improved generalization. The Neural Network model
gauges the model's ability to reproduce the training data. It performs well across all datasets, and its training accuracy
helps establish how well the model recognizes patterns in the reflects its capacity for understanding intricate correlations in
dataset and ensures that the system can learn from the training the data. This model has significant generalization skills,
data. Second, the Test Accuracy is a critical metric for especially in cases requiring detailed pattern recognition,
assessing the model's capacity for generalization. By putting despite a modest drop in test and validation accuracy. Similar
the model to the test on a distinct test dataset that wasn't to Logistic Regression, the NN-Keras model displays
utilized during training, we can measure the model's ability accuracy numbers, highlighting its balanced performance
to make accurate predictions on unseen data, which is across datasets. This model is an appealing option for a
essential for real-world applications. Additionally, Validation variety of machine learning applications since it combines
For the models to be optimized and to prevent overfitting, simplicity and efficacy. In conclusion, all four models show
accuracy is crucial. By monitoring the model's performance competitive accuracy in forecasting customer attrition in e-
on a validation dataset while it is being trained, we may adjust commerce. While Decision Trees and Neural Networks
its parameters and improve its hyperparameters. exhibit better training accuracies, indicating their capacity to
Additionally, Confusion Matrices provide a thorough view of recognize complex patterns, they also exhibit some
model performance by showing True Positives (TP), True overfitting. On the other hand, NN-Keras and logistic
Negatives (TN), False Positives (FP), and False Negatives regression retain accuracy consistency and provide
(FN). These matrices are essential for assessing the recall and interpretability. The best appropriate model must take into
accuracy of the models, allowing for a more full account interpretability, computing resources, and model
understanding of their predictive capabilities. We can ensure complexity in addition to the unique needs of the E-
that our prediction models are thoroughly assessed, prevent commerce industry. Making an educated choice may be aided
them from fitting to the best of our abilities, and provide a by a more thorough examination of accuracy, memory, F1-
thorough review of their performance by employing these score, and alignment with particular application goals.
validation criteria.
V. RESULTS AND DISCUSSIONS
It was determined how successfully several predictive
models, including Logistic Regression, Decision Trees,
Neural Networks, and NN-Keras, predicted E-commerce
Customer Churn. We identified certain advantages and
disadvantages after assessing these models according to a
broad variety of criteria. The promise of logistic regression for
precise churn prediction may be seen in its astounding
accuracy and F1-score performance. While Decision Trees
had good recall values, making them effective at spotting
genuine churn events, Neural Networks displayed outstanding
ROC-AUC values, indicating their capability for class
separation. A thorough understanding of the generalization
and convergence capabilities of the models is also provided by
the metrics for Train Accuracy, Test Accuracy, and Validation
Accuracy. Confusion matrices research offered insight into
the trade-offs between accuracy and recall in the models and
offered a full view on false positives and false negatives.
Fig. 1. Performance Evaluation of Training, Testing and Validation
Accuracy for E-Commerce Customer Churn Prediction
 TABLE I: TRAINING, TESTING AND VALIDATION ACCURACY The model successfully classifies non-churn instances, as seen
Accuracy by its 660 true negatives, which is a plus. It also records 50
Algorithm
Train Test Validation
false positives, indicating times when the model mistakenly
Logistic
thought certain customers were quitting when they weren't.
0.83 0.832 0.822 The Decision Tree approach, in conclusion, creates a balance
Regression
between real positives and true negatives. However, the
Decision Tree 0.87 0.85 0.84
significant number of false negatives highlights the need for
Neural Network 0.859 0.854 0.849 additional model improvement to boost its predictive accuracy
NN-Keras 0.83 0.832 0.828 and better retention tactics in the area of predicting customer
attrition in e-commerce.
B. Confusion Matrix – Logistic Regression
The Logistic Regression model's confusion matrix offers a
thorough analysis of how well it predicts E-commerce client
attrition. Notably, this model got 40 true positives,
demonstrating its accuracy in identifying clients who really
did leave. A larger number of 158 false negatives, however,
were noted, suggesting that the model missed a significant
part of true churn occurrences. This statistic emphasizes the
need of raising awareness of customer turnover, since
ignoring these situations might result in lost chances for
attempts to increase retention. Positively, the model showed
650 true negatives, demonstrating its accuracy in
categorizing non-churn situations. Additionally, it showed 40
false positives, indicating that there may have been some Fig. 3. Confusion Matrix of Decision Tree for E-Commerce Customer
cases when the model mistook clients for churners when they Churn Prediction
really weren't. Conclusion: Although Logistic Regression
achieves a balance between true positives and true negatives, D. Confusion Matrix – Neural Network and Neural Network
the significant number of false negatives calls for further with Keras
model refinement to improve its ability to accurately identify The Neural Network model records 130 false negatives in
customers at risk of leaving and, as a result, improve retention addition to 65 true positives for customer turnover, showing
strategies in the e-commerce domain. a high number of lost chances for churn prediction. It
successfully categorizes non-churn instances by maintaining
a balance between 42 false positives and 668 genuine
negatives. The Neural Network using Keras, in comparison,
functions similarly, with 51 true positives and 149 erroneous
negatives. With 757 true negatives and 43 false positives, it
does very well at properly identifying non-churn situations,
but it also misses a significant portion of genuine churn cases.
In the context of e-commerce, both models have the potential
to be improved to increase their sensitivity to customer
turnover and maximize prediction accuracy for more
successful client retention efforts.

Fig. 2. Confusion Matrix of Logistic Regression for E-Commerce

Customer Churn Prediction

C. Confusion Matrix – Decision Tree

The Decision Tree model's confusion matrix offers a
thorough analysis of how well it performs in foretelling E-
commerce client attrition. An outstanding feature of this
model is its ability to correctly detect 71 instances of client
turnover. This shows that it is skilled at identifying situations
when consumers have in fact left. The model does, however,
also record a sizeable 127 false negatives, which is a
significant number of instances when it was wrongly projected
that consumers did not churn when they really did. This
indicator emphasizes how important it is to increase
sensitivity to customer turnover since failing to recognize Fig. 4. Confusion Matrix of Neural Network for E-Commerce Customer
these situations may result in lost chances for client retention. Churn Prediction
 [1] X. Xiahou and Y. Harada, “Customer churn prediction using
AdaBoost classifier and BP neural network techniques in the E-
commerce industry,” American Journal of Industrial and Business
Management, vol. 12, no. 3, pp. 277–293, 2022.
[2] S. Kim and H. Lee, “Customer churn prediction in influencer
commerce: An application of decision trees,” Procedia Comput
Sci, vol. 199, pp. 1332–1339, 2022.
[3] X. Xiahou and Y. Harada, “B2C E-commerce customer churn
prediction based on K-means and SVM,” Journal of Theoretical
and Applied Electronic Commerce Research, vol. 17, no. 2, pp.
458–475, 2022.
[4] P. Gopal and N. Bin MohdNawi, “A Survey on Customer Churn
Prediction using Machine Learning and data mining Techniques in
E-commerce,” in 2021 IEEE Asia-Pacific Conference on
Computer Science and Data Engineering (CSDE), 2021, pp. 1–8.
[5] Seema and G. Gupta, “Development of fading channel patch based
convolutional neural network models for customer churn
prediction,” International Journal of System Assurance
Engineering and Management, pp. 1–21, 2022.
[6] P. Nagaraj, V. Muneeswaran, A. Dharanidharan, M. Aakash, K.
Balananthanan, and C. Rajkumar, “E-Commerce Customer Churn
Fig. 5. Confusion Matrix of Neural Network with Keras for E-Commerce Prediction Scheme Based on Customer Behaviour Using Machine
Customer Churn Prediction Learning,” in 2023 International Conference on Computer
Communication and Informatics (ICCCI), 2023, pp. 1–6.
[7] X. Xiahou and Y. Harada, “Customer churn prediction using
VI. CONCLUSION AdaBoost classifier and BP neural network techniques in the E-
commerce industry,” American Journal of Industrial and Business
In this research, we set out on a thorough investigation of Management, vol. 12, no. 3, pp. 277–293, 2022.
[8] Z. Wu, L. Jing, B. Wu, and L. Jin, “A PCA-AdaBoost model for
predictive models for predicting client attrition in e- E-commerce customer churn prediction,” Ann Oper Res, pp. 1–18,
commerce. The evaluation of Logistic Regression, Decision 2022.
Trees, Neural Networks, and NN-Keras revealed a range of [9] K. Matuszelański and K. Kopczewska, “Customer Churn in Retail
performance traits and subtleties, illuminating their individual E-Commerce Business: Spatial and Machine Learning Approach,”
Journal of Theoretical and Applied Electronic Commerce
strengths and weaknesses. The results of the logistic Research, vol. 17, no. 1, pp. 165–198, 2022.
regression showed a balanced distribution of true positives [10] P. Nagaraj, V. Muneeswaran, A. Dharanidharan, M. Aakash, K.
and true negatives, but there were also a significant number of Balananthanan, and C. Rajkumar, “E-Commerce Customer Churn
false negatives, suggesting room for sensitivity enhancement. Prediction Scheme Based on Customer Behaviour Using Machine
Learning,” in 2023 International Conference on Computer
Although they produced more false positives than true Communication and Informatics (ICCCI), 2023, pp. 1–6.
positives, decision trees showed a strong capability for [11] X. Xiahou and Y. Harada, “B2C E-commerce customer churn
genuine positives. Neural networks performed well, prediction based on K-means and SVM,” Journal of Theoretical
establishing a nice balance between real positives and true and Applied Electronic Commerce Research, vol. 17, no. 2, pp.
458–475, 2022.
negatives. NN-Keras performed exceptionally in true [12] N. Babayev and K. Israfilzade, “Creating complexity matrix for
negatives but had a little lower proportion of genuine classifying artificial intelligence applications in e-commerce: New
positives. Based on trade-offs between accuracy, recall, and perspectives on value creation,” JOURNAL OF LIFE
model interpretability, these results provide the basis for the ECONOMICS, vol. 10, no. 3, pp. 141–156, 2023.
[13] A. Shi, C. Y. Lim, and S. L. Ang, “Customer Churn Analysis for
selection of models adapted to certain E-commerce situations. Live Stream E-Commerce Platforms by Using Decision Tree
In order to improve predictive accuracy, optimize customer Method,” in International Conference on Advanced
churn prediction, and ultimately boost customer retention Communication and Intelligent Systems, 2023, pp. 168–178.
tactics in the dynamic environment of e-commerce, the study [14] C. Singh, M. K. Dash, R. Sahu, and A. Kumar, “Artificial
intelligence in customer retention: a bibliometric analysis and
emphasizes the significance of continual model refining. future research framework,” Kybernetes, 2023.
Future research may concentrate on optimizing these models [15] L. Li, J. Lin, W. Luo, and X. R. Luo, “INVESTIGATING THE
and investigating ensemble ways to further enhance predictive EFFECT OF ARTIFICIAL INTELLIGENCE ON CUSTOMER
skills, assuring E-commerce enterprises can respond RELATIONSHIP MANAGEMENT PERFORMANCE IN E-
COMMERCE ENTERPRISES,” Journal of Electronic Commerce
proactively to customer churn concerns and strengthen their Research, vol. 24, no. 1, pp. 68–83, 2023.
competitive advantage. [16] J. Shobana, C. Gangadhar, R. K. Arora, P. N. Renjith, J. Bamini,
and Y. devidas Chincholkar, “E-commerce customer churn
REFERENCES prevention using machine learning-based business intelligence
strategy,” Measurement: Sensors, vol. 27, p. 100728, 2023.

View publication stats