Predicting Customer Churn in E-commerce Subscription Services using RNN with Attention Mechanisms

Abstract

In e-commerce subscription services, predicting client attrition is essential for businesses trying to boost sales and enhance client retention. This article proposes a unique approach for accurate churn prediction: recurrent neural networks (RNNs) plus attention mechanisms. By taking advantage of the temporal dependencies found in subscription-based interactions, the RNN model dynamically learns and concentrates on important features. This enables it to detect minute patterns that can point to potential churn behavior in the future. This study thoroughly examines churn characteristics in e-commerce subscription services in order to determine the effectiveness of the proposed RNN architecture in anticipating client attrition. Experiments on real datasets show that the model works better than traditional methods, achieving notable accuracy and predictive power. Moreover, the attention processes in the model allow it to rank the informative traits in order of importance, providing insight into the fundamental causes of customer attrition. The results of this study establish a solid framework for proactive churn control and long-term growth, which has important ramifications for e-commerce companies. Businesses may maximize resource allocation, raise customer happiness, and develop enduring connections with their subscribers by utilizing the predictive powers of RNNs with attention mechanisms. Python is used to implement the suggested work. The accuracy of the suggested study is 97%.

Full text

XXX-X-XXXX-XXXX-X/XX/$XX.00 ©20XX IEEE
Predicting Customer Churn in E-commerce
Subscription Services using RNN with Attention
Mechanisms
1Ch. Anudeep
GITAM School of Business
GITAM University
Visakhapatnam, India
ch.anudeep379@gmail.com
4Thiruma Valavan A
Deputy Director,
Indian Institute of Banking & Finance,
Mumbai, India
thirumavalavan63744@gmail.com
2R. Venugopal
Dept of Marketing
GITAM School of Business,
GITAM (Deemed to be) University,
Visakhapatnam, India
vreddi@gitam.edu
5Veera Ankalu. Vuyyuru
Department of CSE
Koneru Lakshmaiah Education Foundation,
Vaddeswaram, Andhra Pradesh, India.
veerankalu14@kluniversity.in
3Mohd Aarif
Department of Commerce
Aligarh Muslim University,
Aligarh, Uttar Pradesh, India
arifgc6036@gmail.com
6S Muthuperumal
Department of Mathematics,
K.Ramakrishnan College of Engineering,
Trichy, India
muthuperumal@krce.ac.in
Abstract— In e-commerce subscription services, predicting
client attrition is essential for businesses trying to boost sales
and enhance client retention. This article proposes a unique
approach for accurate churn prediction: recurrent neural
networks (RNNs) plus attention mechanisms. By taking
advantage of the temporal dependencies found in subscription-
based interactions, the RNN model dynamically learns and
concentrates on important features. This enables it to detect
minute patterns that can point to potential churn behavior in
the future. This study thoroughly examines churn
characteristics in e-commerce subscription services in order to
determine the effectiveness of the proposed RNN architecture in
anticipating client attrition. Experiments on real datasets show
that the model works better than traditional methods, achieving
notable accuracy and predictive power. Moreover, the attention
processes in the model allow it to rank the informative traits in
order of importance, providing insight into the fundamental
causes of customer attrition. The results of this study establish a
solid framework for proactive churn control and long-term
growth, which has important ramifications for e-commerce
companies. Businesses may maximize resource allocation, raise
customer happiness, and develop enduring connections with
their subscribers by utilizing the predictive powers of RNNs
with attention mechanisms. Python is used to implement the
suggested work. The accuracy of the suggested study is 97%.
Keywords— Attention Mechanisms, Churn Prediction, E-
commerce, Recurrent Neural Networks, Subscription Services.
I. INTRODUCTION
The manner that customers obtain goods and services
online has been completely transformed by e-commerce
subscription services. For a subscription fee, these services
allow users to conveniently get things on a regular basis,
frequently on a repeating basis. The selection of subscription
items is enormous and is growing, ranging from software and
streaming services to meal packages and cosmetics.
Personalization—the ability to customize subscriptions to a
user's tastes and needs—is one of the main draws of e-
commerce subscription services. Consumers may take
advantage of personalized product recommendations,
automated delivery, and frequently lower prices as compared
to in-store shopping [1]. Furthermore, subscription models
encourage a feeling of commitment and loyalty between
companies and clients, which may result in increased lifetime
value and recurrent income streams. Long-term client
retention is a major difficulty for e-commerce subscription
services, notwithstanding their benefits. One of the most
important metrics for subscription-based organizations is
customer churn, or the rate at which users cancel their
memberships. It is critical to anticipate client attrition for a
number of reasons. First of all, it enables businesses to foresee
at-risk clients and take proactive measures to retain them
before they go. Businesses may address underlying issues like
product discontent, price concerns, or bad customer
experiences by recognizing the variables that contribute to
churn. Furthermore, lowering churn lowers expenses related
to bringing in new clients while simultaneously protecting
current revenue. Keeping up a high rate of client retention
might also eventually result in increased profitability and
steady growth. Therefore, accurate churn prediction models
are essential for e-commerce subscription services to optimize
customer retention efforts and maximize long-term
profitability [2]. RNNs have become highly effective
instruments for sequential data processing, which makes them
ideal for modelling time-series data, including consumer
interactions in e-commerce subscription services.
Recognizing temporal relationships within sequences, RNNs
can successfully simulate patterns and trends over time,
unlike standard feedforward neural networks [3]. This feature

is very helpful for churn prediction, as past customer
behaviour is a key factor in predicting upcoming churn
occurrences. Moreover, RNN architectures like GRU and
LSTM are made to deal with the vanishing gradient issue,
which makes it possible for them to remember information
across extended sequences and improve upon previous
encounters [4]. Attention methods provide a valuable
enhancement to RNNs in the context of customer churn
prediction by allowing the model to concentrate on the most
pertinent segments of the input sequence. Through the
assignment of relevance weights to various input data pieces,
attention methods imitate human attention, enabling the model
to selectively pay to important qualities while ignoring
irrelevant information. This capacity is particularly useful for
e-commerce subscription services because client behaviour
might differ greatly in terms of how frequently, recently, and
intensely they connect. Churn prediction models
can dynamically adjust to the trajectories of individual
customers by integrating attention mechanisms into RNN
architectures. This allows the models to detect tiny signals that
may indicate an approaching churn. Because of this, attention-
enhanced RNNs provide a framework for forecasting
customer churn in e-commerce subscription services that is
easier to understand and use, enabling companies to prevent
revenue loss and actively retain important customers.
The structure of the paper is as follows. The beginning of
the first part highlights the significance of customer churn
prediction in e-commerce subscription services. The literature
review is covered in Section 2. The available approaches'
limitations are enumerated in Section 3. In Section 4, the
RNNs with attention mechanisms technique for improved
churn prediction is covered. The findings and suggestions are
given in Section 5. Conclusion and more research in Section
6.
II. LITERATURE REVIEW
Artificial intelligence is a critical answer to the rising
rates of customer attrition in sectors such as banking, where
recruiting new users is more expensive than maintaining
current ones. In this study, RNNs and convolutional neural
networks are integrated in conjunction using BiLSTM-CNN
[5]. This combination fixes the supervision problem in
DLCNN models and overcomes the drawbacks of individual
RNNs and CNNs too. This study uses bank data for
comparison analysis to investigate the possible advantages of
the AttnBLSTM-CNN over BiLSTM-CNN. The integration
of attention mechanisms enhances model performance,
enabling advanced warning and pre-emptive measures
against user churn. This not only safeguards against customer
loss but also enhances the core competitiveness of financial
institutions [6]. Transformers are an excellent choice for NPD
prediction because of their ability to handle a variety of
patterns seen in e-commerce data, like as trends, seasonality,
and abnormalities. The transformer model outperforms
baseline techniques in terms of prediction accuracy.
Moreover, a clustered transformer model is presented, which
groups comparable clients according to their purchase
patterns in order to improve accuracy. By highlighting the
possibilities of transformer topologies for NPD prediction,
this method provides e-commerce companies looking to
enhance inventory management and marketing tactics with a
flexible and scalable solution. The study's conclusions
advance the field of predictive modelling in e-commerce and
offer useful information to companies looking to increase
revenue and improve customer happiness through focused
marketing campaigns and customized customer experiences
[7]. Predicting churn rate is crucial for the success of the
telecommunications industry, as it directly impacts
profitability. This study addresses the challenge of accurately
predicting customer churn while maximizing profits. Two
publicly available datasets that reflect the telecom markets in
Southeast Asia and America are used in the study. The
outcomes show that on both datasets, CNN [8] and ANN
perform better than other methods. In comparison, ANN
obtains 98% accuracy for the Southeast Asian dataset and
99% accuracy for the American dataset, whereas CNN
achieves 99% accuracy for both datasets. These results
demonstrate the efficacy of deep learning methods, in
particular CNN and ANN, in precisely forecasting customer
attrition in telecoms. For telecom businesses looking to
improve profitability and optimize client retention tactics, the
suggested models provide insightful information. Telecom
operators may proactively identify at-risk clients and perform
tailored interventions to minimize churn and boost customer
satisfaction by utilizing powerful machine learning
algorithms [9]. The literature review covers papers focusing
on advanced machine learning techniques for predictive
analytics in customer churn. By combining bidirectional
LSTM-CNN models with attention mechanisms for customer
loss prediction, the study shows enhanced performance and
competitive advantage for financial organizations. In
contrast, Leon et al., introduces a transformer-based model
for predicting customers' next purchase day in e-commerce,
showcasing superior accuracy and scalability compared to
traditional methods. Finally, explores other approaches to
learning, such as deep learning and ensemble methods, for
anticipating churn in the telecom sector. Top-performing
models CNN and ANN provide high accuracy and provide
useful insights for telecom firms seeking to improve revenue
and optimize client retention tactics. The review underscores
the growing significance of advanced machine learning
techniques in predictive analytics, enabling businesses to
effectively tackle customer issues and enhance operational
efficiency.
III. PROBLEM STATEMENT
The existing systems for predicting customer churn in e-
commerce subscription services often face several limitations
that hinder their effectiveness. Traditional methods may rely
solely on static features, such as demographics or purchase
history, which fail to capture the dynamic nature of customer
behaviour over time. This approach overlooks important
temporal patterns and trends that could be indicative of churn
risk [10]. To overcome these limitations, the proposed work
utilizes RNNs with Attention Mechanisms, providing several
advantages. More precise forecasts of churn events are made
possible by the model's ability to grasp temporal relationships
and sequential patterns in customer data by integrating RNNs.
The model may prioritize significant characteristics and
disregard noise or extraneous information by focusing on
pertinent portions of the input sequence, which is made
possible by the attention processes. Through the identification
of critical churn contributing components, this attention
mechanism enhances the interpretability of the model and
helps solve the issue of information saturation in e-commerce
data. The proposed study is to give e-commerce firms a strong

tool for proactive churn control by solving these shortcomings
of current systems, which will eventually increase customer
retention and company sustainability.
IV. RNN WITH ATTENTION MECHANISMS FOR
PREDICTING CUSTOMER CHURN
The proposed work presents a new method for predicting
customer attrition in e-commerce subscription services,
combining RNNs with attention mechanisms. It shows better
predictive performance than conventional techniques,
enabling better churn mitigation plans and improved client
retention. The framework is validated using actual
subscription data. The RNN model with attention mechanisms
for forecasting customer churn in e-commerce subscription
services is developed and evaluated using some essential
phases in the methodology of the proposed study. First, data
collection entails gathering relevant datasets containing
information about subscriber interactions, such as transaction
histories, browsing activity, and engagement metrics.
Preprocessing steps follow, creation of time-based features, to
prepare the dataset for model training. Next, the RNN
architecture with attention mechanisms is designed and
implemented. The attention mechanism enables the model to
concentrate on significant characteristics and patterns
throughout the input sequences, while the RNN component
records temporal relationships in the sequential data. The
prepared dataset is then used to train the model, and
backpropagation and gradient descent are used to optimize its
parameters. Several measures are used to assess the model's
performance. It may be possible to do experiments to evaluate
the suggested model's predictive power of customer attrition
by contrasting it with baseline and cutting-edge approaches.
Sensitivity assessments can also be carried out to investigate
how various hyperparameters and architectural decisions
affect the functionality of the model. Finally, the results are
interpreted, and insights are drawn regarding the factors
influencing customer churn in e-commerce subscription
services, along with implications for business strategies and
future research directions. Fig. 1 depicts the proposed work
framework.
Fig. 1. Proposed Framework for RNN with Attention Mechanisms for
Predicting Customer Churn
A. Data Collection
For the proposed work on churn prediction in an
Ecommerce subscription services, the dataset used is the
churn-modelling dataset. This dataset is collected to analyse
customer behaviour and predict churn rates within the
ecommerce industry. It contains various attributes related to
customers, including their demographics, banking activity,
and account status. The dataset comprises several key
features, including Customer Id, Surname, Credit Score,
Geography, Gender, Age, Tenure, Balance, Num of Products,
Has Cr Card, Is Active Member, Estimated Salary, Exited
[11].
B. Data Pre-processing
For the proposed study, a unique strategy is to capture the
time dynamics of consumer behaviour by integrating temporal
aspects in addition to conventional data pretreatment
approaches. To accomplish this, time-related features that
encapsulate temporal patterns and trends must be created.
These characteristics can offer important insights on the
frequency and timing of customer churn events. The churn
prediction model incorporates time-based features like Time
since Last Interaction, Seasonality Indicators, and Temporal
Aggregations to better understand customer behaviour. These
features capture the duration of customer interactions,
seasonal variations, and trends over different time intervals.
This enhances predictive accuracy and robustness in
identifying at-risk customers, thereby improving the model's
predictive accuracy.
C. RNN with Attention Mechanism for Churn prediction
Designed to manage sequential input by maintaining
information across time, RNNs are a kind of neural networks.
RNNs can display temporal dynamic behaviors because they
feature connections that create directed cycles, in contrast to
feedforward neural networks, whereby information travels in
a single direction. An RNN's basic building block is a series
of repeated neural network phases, each of which processes
just one component of the process sequence while preserving
a hidden state that holds data from earlier elements. Every
phase , the hidden state  of an RNN cell is computed using
the equation (1).
 󰇛   󰇜 (1)
Where,  is the input vector at phase ,  is the
hidden state vector from the previous time step,  and 
are weight matrices for the input and hidden state,
respectively,  is the bias vector,  is the hyperbolic
tangent activation function. In churn prediction, RNNs can be
applied to sequential data representing customer interactions
or behaviors over time. The objective is to build an RNN
model to determine, from past behavior, whether a customer
will be inclined to return or not. The churn-modelling dataset,
containing customer attributes and historical transaction data,
is pre-processed to create sequences of input features and
corresponding target labels. The input sequences may include
features such as customer demographics, transactional
history, account activity, and temporal information. The
target labels indicate whether a customer has churned (1) or
not (0) within a certain time frame. The RNN model
architecture consists of multiple recurrent layers, typically
implemented using LSTM or GRU cells. During the training
process, the RNN model learns to predict churn based on
patterns and dependencies in the input sequences. The model
is trained using BPTT, an extension of backpropagation that
unfolds the computational graph over time steps. After
training, the trained RNN model is used to make predictions
on new sequences of input features. The predicted churn
probabilities are compared to a threshold (e.g., 0.5) to
determine the predicted churn class (churn or not churn).

Three major parts constitute an RNN's architecture are, the
input layer, the recurrent layer, and the output layer. The
input layer receives sequential data representing customer
interactions with the bank over time. This could include
features such as transaction histories, account activity,
customer demographics, and temporal information [12]. Each
phase at a time, the recurrent layer examines the input
sequence, changing its hidden state based on the input at that
point and its earlier hidden state. A crucial element of the
model design is the attention mechanism, which enables the
network to concentrate on pertinent segments of the sequence
of inputs during prediction. The attention mechanism
determines the relevance or importance of each time step's
information for the prediction task by computing attention
weights for each time step in the input sequence. The most
informative portions of the input sequence are essentially
highlighted by using these attention weights to produce a
weighted sum of the sequence. Predictions are generated by
the output layer using the data contained in the attention-
weighted input sequence. A binary classifier that uses the
learnt representation of the input sequence to determine
whether or not a customer is likely to churn (1) or not (0) may
be used as the output layer in the framework of customer
retention forecasting. The attention weight   for element
at time step  is given in eqn. (2).
 󰇛󰇜
󰇛󰇜

 (2)
Where,  is the length of the input sequence. Attention
mechanisms play a crucial role in capturing relevant temporal
patterns and features in sequential data, such as customer
transaction histories or account activity. Time steps with
higher attention weights receive more emphasis during the
computation of the final prediction, allowing the model to
prioritize informative features and ignore irrelevant or noisy
inputs. Attention mechanisms in RNNs create interpretable
representations of input sequences, highlighting important
temporal patterns and features. This transparency helps
stakeholders understand customer behaviors driving
predictions and aids decision-making in customer retention
strategies. Attention-based models achieve better prediction
accuracy, precision, and recall rates compared to traditional
RNN architectures without attention mechanisms. Attention
weights are computed using a learned function, assigning
higher weights to relevant time steps, allowing the model to
focus on important temporal patterns and features. In the
proposed work, RNNs with attention mechanisms are
integrated to utilize the strengths of both approaches for
customer churn prediction. The RNN processes the
sequential customer behaviors data, capturing temporal
dependencies and learning a representation of the input
sequence. The attention mechanism then dynamically focuses
on relevant parts of the input sequence, enhancing the model's
ability to capture informative temporal patterns and features
for churn prediction. By combining RNNs with attention
mechanisms, the proposed model can provide accurate
predictions of customer churn and valuable insights into the
factors driving churn behaviors in the banking industry. This
integrated approach offers a powerful framework for
developing proactive customer retention strategies and
reducing churn rates, ultimately leading to improved
customer satisfaction and business performance. The model
architecture of an RNN with attention mechanisms for
customer churn prediction involves integrating attention
mechanisms into the standard RNN architecture to capture
relevant temporal patterns and features more effectively. In
the proposed work for predicting customer churn in e-
commerce subscription services using RNN with attention
mechanisms, the attention mechanism plays a crucial role in
determining the importance or relevance of each element in
the input sequence. The attention mechanism computes
attention weights for each element of the sequence, allowing
the model to focus more on informative features while
making predictions. By integrating attention mechanisms into
the RNN architecture, the model gains the ability to
dynamically focus on informative features or interactions
within the customer's historical data. This allows the model
to allocate more attention to relevant events or patterns that
may precede churn events, leading to improved prediction
accuracy and interpretability. Additionally, attention weights
provide insights into the factors driving churn, enabling
businesses to identify actionable strategies for retaining at-
risk customers. RNNs with attention mechanisms offer a
powerful framework for predicting customer churn in e-
commerce subscription services. By leveraging sequential
data on customer interactions and behaviors, RNN models
can effectively capture temporal dependencies and identify
patterns indicative of potential churn events. Integration of
attention mechanisms further enhances the model's predictive
capabilities by allowing it to focus on the most relevant
features within the input sequence. By applying RNNs with
attention mechanisms, businesses can gain valuable insights
into customer churn behaviors and implement targeted
retention strategies to mitigate churn and improve customer
satisfaction.
V. RESULTS AND DISCUSSION
In the proposed work, the results of applying the
RNN model with attention mechanisms for customer churn
prediction in e-commerce subscription services are highly
promising. The model demonstrates significant
improvements in predictive accuracy and effectiveness
compared to traditional methods. Through extensive
experimentation and evaluation on real-world datasets, the
RNN model achieves superior performance in identifying
customers at risk of churn. The incorporation of attention
mechanisms into the RNN architecture enhances the model's
interpretability and ability to capture subtle temporal patterns
in customer behaviors. The attention weights provide
valuable insights into the features and time points that are
most influential in predicting churn, empowering businesses
to understand the underlying drivers of customer attrition and
implement targeted retention strategies. In general, the
findings demonstrate how well the suggested strategy
predicts user attrition in e-commerce subscription services
and provide useful information for enhancing client retention
and boosting revenue.
A. Performance Evaluation
Performance evaluation for the proposed work involves
assessing the effectiveness and robustness of the RNN model
with attention mechanisms for predicting customer churn in
e-commerce subscription services. Several evaluation
metrics are used to measure the model's performance and
compare it against baseline methods. Evaluation metrics such

as accuracy, precision, recall, and F1-score show the model's
ability to accurately predict churn events and distinguish
between churners and non-churners.
TABLE I. COMPARISON OF EXISTING SYSTEM PERFORMANCE WITH
PROPOSED RNN WITH ATTENTION MECHANISMS
Model
Accuracy
Precision
Recall
F1-
score
RF [13]
91%
90%
78%
83%
AttnBLSTM-
CNN [14]
95%
95%
96%
95%
Proposed RNN
with Attention
Mechanism
97%
97%
96%
96%
Table I shows performance metrics of three models for
predicting customer churn in e-commerce subscription
services: RF, AttnBLSTM-CNN, and RNN with Attention
Mechanism. RF has a 91% accuracy, but may miss some
instances. AttnBLSTM-CNN outperforms RF with 95%
accuracy, precision, recall, and F1-score. RNN with Attention
Mechanism has the highest performance with 97% accuracy,
precision, recall, and F1-score. It effectively identifies churn
instances and minimizes false positives.
Fig. 2. Performance Evaluation of Proposed Model with Existing Model
The performance metrics of three distinct models—RF,
AttnBLSTM-CNN, and RNN with Attention Mechanism—
for forecasting customer attrition in e-commerce subscription
services are shown in Figure 2. Higher accuracy numbers
denote superior performance. Accuracy quantifies the total
correctness of forecasts. The percentage of accurately
anticipated churn instances among all projected churn
instances is known as precision, whereas the percentage of
properly predicted churn instances among all real churn events
is known as recall. The F1-score offers a fair assessment of a
model's performance as it is the harmonic mean of accuracy
and recall. The RNN with Attention Mechanism beats the
other models in every statistic, including accuracy, precision,
recall, and F1-score, as shown in the image. This shows that
the RNN model with attention mechanisms may more
accurately forecast customer turnover in e-commerce
subscription services by capturing temporal relationships and
focusing on pertinent elements within the input sequence.
B. Discussion
The proposed work on predicting customer churn in e-
commerce subscription services using RNN with attention
mechanisms, it's important to address both the limitations of
existing systems and the advantages of the proposed
approach, as well as identify potential areas for future
research. Existing churn prediction systems often rely on
traditional machine learning algorithms or simple heuristics,
which may struggle to capture complex temporal
dependencies and subtle patterns in customer behaviors [15].
Additionally, these systems may overlook the importance of
feature attention, leading to suboptimal performance in
identifying key factors driving churn. Moreover, scalability
and interpretability issues may arise when dealing with large-
scale datasets and complex models. The suggested study has
a number of benefits over current systems. The model may
more accurately forecast customer turnover by focusing on
pertinent aspects within the sequential data and capturing
temporal relationships through the use of RNNs with
attention mechanisms. By include attention processes, the
model becomes easier to understand and gives organizations
more insight into the variables that affect churn behavior.
Moreover, the suggested methodology exhibits great
scalability and adeptness in managing extensive datasets,
rendering it appropriate for practical implementations in e-
commerce subscription services. Notwithstanding its
benefits, the suggested job has several drawbacks. One
limitation is the potential for overfitting, especially when
dealing with limited training data or highly imbalanced
datasets. Addressing this limitation may require techniques
such as data augmentation, regularization, or ensemble
methods. The proposed approach may be computationally
intensive, particularly when training deep neural networks on
large-scale datasets. Future work could focus on optimizing
model architectures and training procedures to improve
efficiency and scalability. Furthermore, exploring the impact
of additional features or external factors on customer churn
prediction could enhance the model's performance and
provide deeper insights into customer behaviors. Finally,
evaluating the proposed approach on diverse datasets from
different e-commerce domains and subscription services
could further validate its effectiveness and generalizability.
VI. CONCLUSION AND FUTURE SCOPE
The proposed work on leveraging RNN with attention
mechanisms to predict client churn in e-commerce
subscription services has led to a substantial advancement in
client relationship management and retention strategies. The
model performs better in capturing temporal relationships
and detecting critical elements driving churn behavior by
utilizing RNNs with attention mechanisms. By incorporating
attention processes, the model becomes more interpretable
and offers insightful information about the most significant
elements in the input sequence. This enables companies to
develop focused retention campaigns and successfully reduce
customer attrition. The suggested method's efficiency and
scalability make it ideal for real-world applications in e-
commerce subscription services, where complicated
consumer interactions and big datasets are commonplace.
The model is a useful tool for companies looking to increase
revenue and improve customer retention because of its
0
20
40
60
80
100
120
Accuracy Precision Recall F1-score
Percentage
Performance Evaluation
RF
AttnBLSTM-CNN
Proposed RNN with Attention Mechanism

versatility in handling various data sources and capacity to
adjust to shifting customer behavior patterns. There are
numerous directions in which research and development can
go. First off, investigating more complex RNN and attention
mechanism architectures and methodologies may help to
boost model performance and prediction accuracy even more.
Furthermore, carrying out longitudinal research to assess the
long-term efficacy of retention tactics put into place in
accordance with the model's predictions may offer insightful
and helpful criticism for continued optimization. In addition
to opening the door for further study and advancement in this
crucial field of customer relationship management, the
suggested work establishes the groundwork for efficient
churn prediction and customer retention methods in e-
commerce subscription services.
REFERENCES
[1] S. F. Bischof, T. M. Boettger, and T. Rudolph, “Curated subscription
commerce: A theoretical conceptualization,” Journal of Retailing and
Consumer Services, vol. 54, p. 101822, 2020.
[2] P. Lalwani, M. K. Mishra, J. S. Chadha, and P. Sethi, “Customer churn
prediction system: a machine learning approach,” Computing, vol. 104,
no. 2, pp. 271–294, 2022.
[3] R. Sudharsan and E. Ganesh, “A Swish RNN based customer churn
prediction for the telecom industry with a novel feature selection
strategy,” Connection Science, vol. 34, no. 1, pp. 1855–1876, 2022.
[4] A. Khodadadi, S. A. Hosseini, E. Pajouheshgar, F. Mansouri, and H.
R. Rabiee, “ChOracle: A unified statistical framework for churn
prediction,” IEEE Transactions on Knowledge and Data Engineering,
vol. 34, no. 4, pp. 1656–1666, 2020.
[5] Y. Liu, M. Shengdong, G. Jijian, and N. Nedjah, “Intelligent prediction
of customer churn with a fused attentional deep learning model,”
Mathematics, vol. 10, no. 24, p. 4733, 2022.
[6] G. Mena, K. Coussement, K. W. De Bock, A. De Caigny, and S.
Lessmann, “Exploiting time-varying RFM measures for customer
churn prediction with deep neural networks,” Annals of Operations
Research, pp. 1–23, 2023.
[7] A. Grigoraș and F. Leon, “Transformer-Based Model for Predicting
Customers’ Next Purchase Day in e-Commerce,” Computation, vol.
11, no. 11, p. 210, 2023.
[8] M. U. Tariq, M. Babar, M. Poulin, and A. S. Khattak, “Distributed
model for customer churn prediction using convolutional neural
network,” Journal of Modelling in Management, vol. 17, no. 3, pp.
853–863, 2022.
[9] L. Saha, H. K. Tripathy, T. Gaber, H. El-Gohary, and E.-S. M. El-
kenawy, “Deep churn prediction method for telecommunication
industry,” Sustainability, vol. 15, no. 5, p. 4543, 2023.
[10] L. Geiler, S. Affeldt, and M. Nadif, “A survey on machine learning
methods for churn prediction,” International Journal of Data Science
and Analytics, vol. 14, no. 3, pp. 217–242, 2022.
[11] “Churn_Modelling_dataset.” Accessed: Apr. 22, 2024. [Online].
Available: https://www.kaggle.com/datasets/nazishjaveed/churn-
modelling-dataset?resource=download
[12] Z. Cuiyan, Z. Manman, X. Xiaoling, M. Yiwei, and C. Hao, “Customer
Churn Prediction Model Based on User Behavior Sequences.,” Journal
of Donghua University, vol. 39, no. 6, 2022.
[13] C. Ling, T. Zhang, and Y. Chen, “Customer purchase intent prediction
under online multi-channel promotion: A feature-combined deep
learning framework,” IEEE Access, vol. 7, pp. 112963–112976, 2019.
[14] K. Weterings, S.-L. Kimelman, S. Bromuri, and M. van Eekelen,
“Interpreting Attention Models: LSTM vs. CNN: A case study on
customer activation,” in 2020 19th IEEE International Conference on
Machine Learning and Applications (ICMLA), IEEE, 2020, pp. 336–
343.
[15] J. Ahn, J. Hwang, D. Kim, H. Choi, and S. Kang, “A survey on churn
analysis in various business domains,” IEEE Access, vol. 8, pp.
220816–220839, 2020.