Available via license: CC BY-NC-ND 4.0
Content may be subject to copyright.
VOLUME XX, 2017 1
Customer Behaviour Analysis and
Predictive Modelling in Supermarket
Retail: A Comprehensive Data Mining
Approach
Dr. Kavitha Dhanushkodi 1, Akila Bala2, Nithin Kodipyaka3 and Shreyas V4
1Assistant Professor Sr., School of Computer Science and Engineering, Vellore Institute of Technology, Chennai, India
2School of Computer Science and Engineering, Vellore Institute of Technology, Chennai, India
3School of Computer Science and Engineering, Vellore Institute of Technology, Chennai, India
4School of Computer Science and Engineering, Vellore Institute of Technology, Chennai, India
Corresponding author: Dr. D. Kavitha (e-mail: kavitha.d@vit.ac.in).
This work was supported by Vellore Institute of Technology, Chennai, India.
ABSTRACT In the dynamic landscape of supermarket retail, understanding customer behavior is paramount
for optimizing business strategies and enhancing profitability. This paper presents a comprehensive data
mining approach to analyze customer behavior and build predictive models within the supermarket retail
domain. Leveraging advanced data analytics techniques, our methodology encompasses data preprocessing,
exploratory data analysis, feature engineering, model selection, and evaluation. This paper presents a
comprehensive approach to customer behavior analysis and predictive modelling within the context of
supermarket retail. We delve into the intricacies of data mining methodologies, exploring how retailers can
leverage diverse datasets to uncover valuable insights and build predictive models that drive business growth
and customer satisfaction. From data preprocessing to model evaluation, each step in the process is
meticulously examined, highlighting best practices and key considerations for effective implementation.
INDEX TERMS Customer behavior analysis, Data mining, Predictive modelling, Retail, Sequential pattern
mining.
I. INTRODUCTION
In the contemporary landscape of retail, particularly within
the dynamic realm of supermarket operations, the quest to
understand and predict customer behavior stands as a
cornerstone of strategic decision-making. Amidst the
proliferation of data sources and the advent of sophisticated
analytical methodologies, retailers are compelled to adopt
comprehensive data mining approaches to derive actionable
insights that drive business success. This paper embarks on a
journey through the intricate domain of customer behavior
analysis and predictive modelling within the context of
supermarket retail, elucidating the pivotal role of data-driven
strategies in optimizing operational efficiencies and
enhancing customer satisfaction.
The exponential growth of data availability in recent years
has paved the way for transformative advancements in retail
analytics. With transaction records, customer demographics,
loyalty program data, and product information constituting a
rich tapestry of insights, retailers are presented with an
unprecedented opportunity to unravel the intricacies of
consumer behavior. By harnessing advanced data mining
techniques, retailers can distil these vast datasets into
actionable intelligence, empowering them to make informed
decisions across a spectrum of operational domains,
including marketing, merchandising, and inventory
management.
Against this backdrop, this paper endeavors to elucidate a
systematic approach to customer behavior analysis and
predictive modelling in the context of supermarket retail.
Through a structured methodology encompassing data
preprocessing, exploratory data analysis, feature engineering,
model selection, and evaluation, retailers can unlock the
latent potential of their data assets and derive meaningful
insights into customer preferences, purchasing patterns, and
brand affinities.
Furthermore, the integration of predictive modelling
techniques not only presents an opportunity to forecast
forthcoming consumer behaviors but also empowers retailers
to proactively adjust their strategies. This proactive approach
allows for the customization of product offerings to align
with evolving consumer preferences, the tailoring of
marketing campaigns to resonate with specific customer
segments, and the optimization of resource allocation
strategies for improved operational efficiency. Through a
meticulous examination of each phase within the data mining
process, coupled with insights derived from practical
implementations, this paper aims to furnish retailers with the
necessary expertise and tools to embrace data-driven
decision-making effectively. By doing so, retailers can not
only achieve sustained growth but also gain a competitive
edge in the dynamic landscape of supermarket retail.
This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2024.3407151
This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see https://creativecommons.org/licenses/by-nc-nd/4.0/
VOLUME XX, 2017 1
Embracing data-driven strategies as a cornerstone of their
operations, retailers can navigate through the complexities of
the market, capitalize on emerging opportunities, and foster
long-term success in an ever-evolving retail environment.
II. RELATED WORK
Customer segmentation in retail, particularly in online retail,
plays a crucial role in tailoring marketing strategies. The
RFM model, focusing on Recency, Frequency, and Monetary
value, is a significant framework for this purpose [1]. Data
mining offers advantages in understanding customer
behavior and preferences, enhancing data quality, and
enabling streamlined operations in retail. However, privacy
concerns must be addressed to maintain consumer trust [2].
The integration of data mining and machine learning in retail,
akin to the banking industry, leads to improved customer
satisfaction and retention rates through personalized services
[3]. Data warehousing aids in customer segmentation and
inventory management, facilitating personalized marketing
and operational efficiency in retail [4]. E-commerce data
mining enables retailers to understand market demands,
optimize product assortment, refine pricing strategies, and
ultimately drive revenue growth [5]. Data mining has become
a pivotal tool in the retail sector, revolutionizing operations
with its capacity to derive actionable insights from vast
datasets. Its application offers numerous advantages,
foremost among them being the acquisition of valuable
customer insights. By analyzing purchasing patterns,
preferences, and behaviors, retailers can tailor their
marketing strategies and enhance customer experiences.
Customer segmentation is a critical strategy in online retail,
allowing businesses to tailor marketing efforts effectively.
The RFM model, which analyzes Recency, Frequency, and
Monetary value, provides a comprehensive framework for
this purpose [1]. Data mining offers significant advantages in
retail, including understanding customer behavior, enhancing
data quality, and improving operational efficiency. However,
addressing privacy concerns is essential to maintaining
consumer trust and compliance with regulations [2]. The
integration of data mining and machine learning techniques
in retail, drawing parallels from the banking sector, has
shown promising results in enhancing customer satisfaction
and retention rates through personalized services [3]. Data
warehousing plays a vital role in retail operations, facilitating
customer segmentation and inventory management. This
enables retailers to optimize marketing strategies and
streamline inventory levels for improved efficiency [4]. E-
commerce data mining provides valuable insights into market
trends, enabling retailers to make informed decisions
regarding product assortment, pricing strategies, and
promotional activities. This ultimately drives revenue growth
and ensures competitiveness in the online retail landscape [5].
Data mining has emerged as a pivotal tool for enhancing
competitive advantage in both banking and retail sectors. In
banking, data mining empowers institutions to sift through
vast troves of customer data encompassing transactional
histories, demographics, and behavioural patterns. This
enables the segmentation of customers according to their
distinct needs, preferences, and risk profiles, thereby
facilitating tailored services and targeted marketing
strategies. Moreover, predictive analytics and anomaly
detection techniques bolster risk management efforts by
enabling early identification of potential threats and
fraudulent activities, thus mitigating financial losses.
Similarly, in the retail industry, data mining plays a crucial
role in customer segmentation based on purchase behaviors,
browsing habits, and demographic insights. This
segmentation aids retailers in crafting personalized shopping
experiences and devising targeted promotional campaigns.
Furthermore, data mining serves as a potent tool in fraud
detection, helping retailers combat revenue losses stemming
from theft, shoplifting, or the circulation of counterfeit goods.
Overall, the integration of data mining techniques equips both
banking and retail sectors with valuable insights, enabling
them to optimize operations, enhance customer satisfaction,
and fortify against financial risks and fraudulent activities [6].
Data mining techniques have become integral to the retail
industry, offering diverse applications that enhance
operational efficiency and strategic decision-making. One
such crucial application is customer segmentation, where
clustering and classification algorithms categorize customers
based on demographics, purchase behavior, and preferences.
This segmentation enables retailers to tailor marketing
strategies effectively, thereby optimizing customer
engagement and satisfaction [7]. Market basket analysis, a
subtype of association rule mining, further aids retailers in
uncovering patterns among products frequently purchased
together. This insight facilitates optimized product
placement, promotions, and cross-selling opportunities,
ultimately boosting sales and profitability [7]. Moreover, data
mining techniques play a vital role in demand forecasting,
employing methods such as time series analysis and
regression modelling to predict future demand accurately. By
leveraging historical sales data, seasonal patterns, trends, and
external factors, retailers can optimize inventory
management and ensure adequate stock levels to meet
customer demand [7]. In the realm of e-commerce,
personalized online sales leveraging web usage data mining
have gained significant attention for enhancing user
experience and increasing sales. This process begins with
comprehensive data collection, encompassing user
interactions such as clicks, views, searches, and purchases.
Subsequent data preprocessing ensures data quality by
removing noise, handling missing values, and transforming
data into an analyzable format. User profiling, based on
extracted web usage data, enables tailored recommendations
and marketing strategies, fostering customer engagement and
loyalty [8]. The study titled "Performance Study of
Classification Algorithms for Consumer Online Shopping
Attitudes and Behavior Using Data Mining" investigates
various classification algorithms' effectiveness in analyzing
consumer online shopping attitudes and behavior. Following
a structured approach involving data collection,
preprocessing, and feature engineering, the study conducts
unbiased model training and evaluation through data
splitting. Rigorous optimization of classification algorithms
through model training and hyperparameter tuning is
performed, followed by thorough model evaluation using
techniques such as cross-validation to assess generalization
capabilities. The study's findings provide valuable insights
into the strengths and weaknesses of each algorithm in
predicting consumer behaviour, offering valuable guidance
This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2024.3407151
This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see https://creativecommons.org/licenses/by-nc-nd/4.0/
VOLUME XX, 2017 1
for researchers and practitioners in online consumer
behaviour analysis [9]. Srikant and Agrawal's seminal paper
titled "Mining Sequential Patterns" has significantly
influenced the data mining landscape, particularly in
sequential data analysis. Their work offers crucial insights
and methodologies for extracting meaningful patterns from
sequences of data, introducing concepts like the Apriori-
based algorithm for sequence pattern mining. Over the years,
researchers have built upon this foundation, exploring
various generalizations and performance enhancements to
advance sequential pattern mining techniques. This paper
remains a cornerstone in the literature, serving as a
benchmark for subsequent research endeavors aimed at
harnessing insights within sequential data structures [10].
Magnini, Honeycutt Jr, and Hodge explore data mining
applications in the hotel industry, highlighting its role in
extracting actionable insights from vast data reservoirs. They
emphasize benefits like enhanced customer segmentation and
demand forecasting, alongside challenges such as data
quality and privacy concerns [11]. Ritbumroong investigates
online analytical mining (OLAM) methodologies, combining
data mining and online analytical processing (OLAP) to
extract actionable insights. OLAM's integration allows
businesses to understand customer preferences, anticipate
trends, and tailor strategies effectively [12]. Hemalatha's
study focuses on market basket analysis in Indian retailing,
aiming to understand consumer purchasing behaviors and
associations between products. Leveraging data mining
methodologies, this research aids in targeted marketing and
product placement decisions [13]. Huang, Chang, and
Narayanan delve into customer behaviour in dynamic
markets, using data mining techniques to comprehend and
anticipate shifts in preferences. Their study highlights the
importance of adapting to changing market dynamics for
effective forecasting and response [14]. Punpukdee et al.
conducted a comprehensive investigation into determinants
of consumer behaviour in Thailand's offline marketing
sphere, synthesizing insights from existing literature and
employing data mining techniques to uncover patterns and
connections among factors shaping consumer decisions [15].
III. DATASET
A. COLUMNS
• BasketID: Unique identifier for each basket or
transaction.
• BasketDate: Date and time of the basket or
transaction.
• Sale: Sale amount for each product in the basket.
• CustomerID: Unique identifier for each customer.
• CustomerCountry: Country of the customer.
• ProdID: Unique identifier for each product.
• ProdDescr: Description of the product.
• Qta: Quantity of each product in the basket.
B. DATA ENTRIES
Each row represents a product purchased within a specific
basket or transaction.
The dataset contains information about various transactions,
including the date, customer details, product details, and
quantities purchased.
C. DATA TYPES
• BasketID: numeric identifier.
• BasketDate: Date and time format.
• Sale: Numeric
• CustomerID: Numeric
• CustomerCountry: Categorical, representing the
country name.
• ProdID: alphanumeric identifier.
• ProdDescr: Text description.
• Qta: Numeric, representing quantity.
D. DATA QUALITY
To assess the quality of the data, the process began by
eliminating duplicate entries, amounting to 5232 instances,
which represented approximately 1.11% of the entire dataset.
This left the analysis with 466,678 rows for further
examination. Upon visual inspection of the numerical
attributes, it became apparent from the plots in Figure 1a that
both attributes exhibited notably high outliers, both positive
and negative. However, the presence of negative values was
inconsistent with the semantics of the attributes, as they
should inherently be positive. From Figure 1 and Figure 2,
further investigation revealed that negative values in the
"Qta" attribute likely represented refunds, supported by the
symmetric behaviour of the attribute. Additionally, nearly all
records with negative "Qta" values were associated with
BasketIDs starting with "C," indicative of cancellations.
Notably, some records with negative "Qta" values lacked a
corresponding "C" BasketID prefix, yet analysis of their
respective "ProdDescr" suggested they pertained to errors or
damaged items. Regarding negative "Sale" values, only two
records exhibited this property, which upon examination of
the "ProdDescr" ("ADJUST BAD DEBT") were attributed to
errors. Importantly, all rows identified as errors were
associated with null CustomerIDs. Subsequently, the analysis
proceeded by removing entries corresponding to the 65,073
null CustomerID values, constituting approximately 13.94%
of the dataset. This action was deemed necessary as the
primary objective was to analyze customer behaviour,
rendering entries with null CustomerIDs irrelevant. This
removal process also eliminated the previously identified
errors. Additionally, ProdIDs that did not conform to the
defined format, consisting solely of letters such as 'POST',
'D', 'C2', 'M', 'BANK CHARGES', etc., accompanied by
respective ProdDescrs such as 'POSTAGE', 'Discount',
'CARRIAGE', 'Manual', 'Bank Charges', etc., were
eliminated from the dataset. Consequently, 1273 entries were
dropped from the dataset.
This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2024.3407151
This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see https://creativecommons.org/licenses/by-nc-nd/4.0/
VOLUME XX, 2017 1
FIGURE 1. Boxplot before Outlier Removal.
FIGURE 2. Boxplot after Outlier Removal.
E. VARIABLE DISTRIBUTION
FIGURE 3. Yearly Variable Distribution.
FIGURE 4. Monthly Variable Distribution.
FIGURE 5. Weekly Variable Distribution.
FIGURE 6. Daily Variable Distribution.
From Figure 3, we can see that the attribute is highly
unbalanced; in fact, we have that almost all the
records are related to transactions of 2011, while the objects
from 2010 are very few. Indeed, the rows of
2011 represents about 93% of the whole dataset.
This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2024.3407151
This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see https://creativecommons.org/licenses/by-nc-nd/4.0/
VOLUME XX, 2017 1
Furthermore, Figure 4, represents an estimation of the
probability density function of BasketDate
divided by year, we can appreciate the two different
distributions.
In fact, for 2010, we have a very uneven plot, which indicates
that the records are not uniformly distributed
concerning the days in a month as shown in Figure 5. This is
because, for the majority of the months in 2010, there were
registered only transactions from a single day from Figure 6,
the 12th; this is the value for which the plot shows the peak.
On the other
hand, the distribution for 2011 is much more homogeneous,
meaning that the transactions were registered
for most days in the months of that year.
F. SUMMARY
The dataset is transactional sales data for a retail business, l
for an online store.
Each transaction consists of multiple products.
It includes details such as the date of purchase, customer
information, product details, and quantities sold.
The dataset covers sales across different countries, as
indicated by the "CustomerCountry" column.
IV. EXPLORATORY DATA ANALYSIS
A. DATASET CLEANING
1) REMOVAL OF NULL CUSTOMERID ENTRIES
Identify rows in the dataset where the CustomerID variable is
null.
Remove these rows from the dataset as they lack essential
customer identification information.
2) IDENTIFICATION OF PECULIAR TRANSACTIONS
Examine the ProdID variable to identify transactions
characterized by codes consisting solely of letters.
Investigate these transactions to understand their nature, such
as port charges, bank charges, or discounts.
3: IDENTIFICATION OF CANCELED TRANSACTIONS
Check for entries in the BasketID variable prefixed with "C,"
indicating cancelled transactions.
Ensure that for each cancelled transaction, there exists a
corresponding entry with a negative quantity (Qta)
representing the cancellation.
B. FEATURE EXTRACTION
In the project, RFM (Recency, Frequency, Monetary)
analysis is employed as a fundamental technique for customer
segmentation and understanding consumer behaviour in the
context of online shopping [4]. At its core, RFM analysis
operates on the premise of the Pareto Principle, often referred
to as the "80-20 rule," which suggests that approximately
80% of a company's revenue is generated by around 20% of
its customers. By segmenting customers according to RFM
criteria, businesses can identify this vital subset of customers
who contribute significantly to their bottom line. The first
component of RFM analysis is recency, which measures how
recently a customer has made a purchase or interacted with
the business as shown in Table I. Customers who have
engaged with the business more recently are often considered
more valuable, as they may be more likely to make additional
purchases in the near future. Frequency refers to the number
of transactions or interactions a customer has had with the
business within a specific period. Customers who make
frequent purchases or engage with the business regularly are
typically seen as more loyal and valuable.
Monetary value assesses the total amount of money a
customer has spent on purchases or transactions with the
business. Customers who have spent more money over time
are generally considered high-value customers, as they
contribute more to the company's revenue. RFM analysis
enables the categorization of customers based on their
transactional history, focusing on three key metrics:
• Recency Feature: This metric measures the time
elapsed since a customer's last purchase. Customers
who have made recent purchases are likely to be
more engaged with the business and may exhibit
different behaviour compared to those who have not
made purchases in a while.
Calculate the number of days between the present
date and the date of the customer's last purchase.
This feature quantifies how recently a customer
made a purchase, indicating their engagement level
with the business.
Example: recency = present_date -
last_purchase_date
• Frequency Feature: Frequency refers to the
number of transactions made by each customer
within a given period. Customers who make
frequent purchases may represent loyal or high-
value segments, while those with lower frequencies
may require targeted marketing efforts to increase
engagement.
Count the number of orders or transactions for each
customer.
This feature represents how frequently a customer
makes purchases, indicating their loyalty or
engagement with the business.
Example: frequency = total_number_of_orders
• Monetary Feature: The monetary metric
represents the total amount spent by each customer
over a specified timeframe. This metric helps
identify high-spending customers who contribute
significantly to revenue generation and can inform
strategies for personalized marketing and
promotions. Calculate the total purchase amount or
spending for each customer.
This feature represents the monetary value of the
customer's transactions, indicating their profitability
or contribution to the business.
Example: monetary = sum_of_purchase_amounts
This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2024.3407151
This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see https://creativecommons.org/licenses/by-nc-nd/4.0/
VOLUME XX, 2017 1
TABLE I. RFM Quartile Table.
FIGURE 7. RFM Segmentation Table.
Customer segmentation is a crucial strategy for businesses to
effectively target their marketing efforts and maximize
profitability. One popular method for segmenting customers
is RFM analysis, which evaluates customers based on
Recency, Frequency, and Monetary value. By dividing
customers into segments based on their RFM scores,
businesses can identify their most valuable customers and
implement targeted marketing campaigns. In this dataset, we
have segmented customers into six categories: Best
Customers, Loyal Customers, Big Spenders, Almost Lost,
Lost Customers, and Lost Cheap Customers. In Figure 7 each
segment represents a different level of engagement and
potential value to the business, allowing for tailored
marketing strategies to be implemented.
• Best Customers: These are customers who have the
best scores in all three RFM categories (Recency,
Frequency, and Monetary). Specifically, they have
the best Recency score (1), the best Frequency score
(1), and the best Monetary score (1). In your dataset,
there are 440 customers in this segment.
• Loyal Customers: Loyal customers may not have the
absolute best scores in all three categories, but they
still demonstrate strong loyalty to the business. In
your dataset, there are 1012 customers classified as
Loyal Customers.
• Big Spenders: These customers may not purchase as
frequently as the Best Customers or Loyal
Customers, but they spend a significant amount
when they do make a purchase. They have the best
Monetary score (1) but may not necessarily have the
best scores in the other categories. There are 1051
customers in this segment in your dataset.
• Almost Lost: These customers may have shown
signs of decreased activity or engagement with the
business. They may have previously been more
active but have become less so recently. In your
dataset, there are 105 customers classified as Almost
Lost.
• Lost Customers: These customers have not made
purchases in a long time and are considered lost to
the business. They have the worst scores in the
Recency category. In your dataset, there are 11
customers classified as Lost Customers.
• Lost Cheap Customers: These customers have not
made purchases in a long time, like Lost Customers,
but when they did make purchases, they typically
spent less than other segments. They have low
scores in both the Recency and Monetary categories.
In your dataset, there are 431 customers classified as
Lost Cheap Customers.
Feature correlation analysis is essential to understand the
relationships between different features in a dataset. In the
context of RFM analysis for customer segmentation, we can
perform correlation analysis to examine how the Recency,
Frequency, and Monetary features are related to each other.
By assessing the correlation in Figure 8 between these three
key dimensions of customer behaviour, we can gain insights
into patterns such as whether customers who make purchases
more frequently tend to spend more money per transaction or
if customers who have made recent purchases are more likely
to make repeat purchases. Moreover, understanding feature
correlations helps in identifying redundant or highly
correlated features, which can be beneficial for feature
selection and model simplification.
FIGURE 8. Feature Correlation Heatmap.
V. SYSTEM ARCHITECTURE
Implementation Description for Implementation Diagram
Mapping according to Figure 10:
A. DATA UNDERSTANDING AND PREPARATION
• Data Semantics Analysis: Initially, explore the
dataset using pandas to understand its structure,
features, and data types.
• Distribution and Statistics Analysis: Utilize
descriptive statistics to understand the distribution
of variables such as sales, customer IDs, product
IDs, etc.
R
F
M
0.25
22.0
1.0
298.11
0.50
58.0
2.0
646.30
0.75
151.0
4.0
1567.20
This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2024.3407151
This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see https://creativecommons.org/licenses/by-nc-nd/4.0/
VOLUME XX, 2017 1
• Data Quality Assessment: Use techniques such as
checking for missing values, outliers, and
inconsistencies in the dataset to ensure data quality.
• Variables Transformation and Generation:
Perform necessary transformations on variables,
such as converting data types, encoding categorical
variables, and generating new features as required
by the task.
• Pairwise Correlations Analysis: Calculate
pairwise correlations between variables to identify
relationships and potential redundancies. Eliminate
redundant variables if necessary.
B. CLUSTERING ANALYSIS
• K-means Clustering: Identify the optimal value of
k using techniques such as the elbow method or
silhouette score as depicted in Figure 9.
• Density-based Clustering: Study clustering
parameters such as minimum samples and epsilon
for DBSCAN.
Characterize and interpret clusters obtained from
DBSCAN.
• Hierarchical Clustering: Compare different
hierarchical clustering results using different
linkage methods (e.g., single, complete, average).
Visualize and analyze dendrograms to understand
cluster hierarchy and structure.
• Alternative Clustering Techniques: Explore
additional clustering techniques provided by the
clustering library, such as agglomerative clustering
or G-means clustering.
C. PREDICTIVE ANALYSIS
• Customer Profile Definition: Define a customer
profile based on indicators computed during data
preparation, considering attributes like the total
number of items purchased, distinct items bought,
maximum items purchased per session, and
Shannon entropy of purchasing behaviour.
• Label Generation: Compute labels for customers
indicating high-spending, medium-spending, or
low-spending categories based on defined
thresholds.
• Model Selection and Evaluation: Evaluate the
performance of different predictive models (e.g.,
logistic regression, decision trees, random forests)
on training and test sets. Discuss preprocessing
techniques applied to manage potential challenges in
prediction.
D. SEQUENTIAL PATTERN MINING
• Modelling Customer Sequences: Model customers
as sequences of baskets, where each basket contains
purchased items.
• Sequential Pattern Mining Algorithm: Implement
the Generalized Sequential Pattern (GSP) algorithm
to mine frequent sequential patterns in customer
basket sequences as shown in Figure 11.
• Pattern Analysis and Interpretation: Discuss
resulting sequential patterns, considering their
support, length, and interpretation in the context of
customer purchasing behaviour.
These implementation steps outline the process flow for
analyzing customer behaviour, performing clustering
analysis, predictive modelling, and sequential pattern
mining based on the provided project description and
sequential pattern mining code. Each step contributes to
understanding customer preferences, segmenting
customers, and making predictions to enhance
supermarket retail strategies [7].
FIGURE 9. Silhouette Method for Optimal k.
FIGURE 10. Architectural Diagram of the Proposed Process.
This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2024.3407151
This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see https://creativecommons.org/licenses/by-nc-nd/4.0/
VOLUME XX, 2017 1
FIGURE 11. Sequential Diagram of the Proposed Process.
VI. MODULAR DECOMPOSITION
Given below in Figure 12 is the modular decomposition for
the proposed technology
A. OBJECTIVES
• Analyze customer behaviour in supermarket retail.
• Develop predictive models for customer behaviour.
• Utilize data mining techniques for comprehensive
analysis.
B. COMPONENTS
• Data Collection
• Data Preprocessing
• Customer Segmentation
• Predictive Modeling
• Algorithm Selection
• Model Training
• Model Evaluation
• Feature Importance Analysis
• Result Interpretation
• Conclusion
C. TASKS
• Gather supermarket retail data including sales,
transactions, and customer demographics.
• Clean, preprocess, and integrate the collected data
for analysis.
• Segment customers based on their shopping
behaviour, preferences, and demographics.
• Develop predictive models to forecast customer
behaviour, such as purchase propensity or churn.
• Select appropriate data mining algorithms for
predictive modelling.
• Train the predictive models using the prepared data.
• Evaluate the performance of the trained models
using relevant metrics.
• Analyze the importance of features in predicting
customer behaviour.
• Interpret the results of the analysis and conclude.
• Provide recommendations based on the findings to
improve supermarket retail strategies.
This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2024.3407151
This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see https://creativecommons.org/licenses/by-nc-nd/4.0/
VOLUME XX, 2017 1
FIGURE 12. Flowchart Diagram of the Proposed Architecture.
Each component serves a specific purpose, allowing for a
structured and organized workflow. For instance, data
collection involves gathering diverse datasets encompassing
sales, transactions, and customer demographics, while
preprocessing focuses on cleaning and integrating these
datasets for analysis. Customer segmentation aims to
categorize customers based on various parameters,
facilitating targeted marketing strategies. Predictive
modelling utilizes advanced algorithms to forecast customer
behaviour, such as purchase propensity or churn. Algorithm
selection ensures the use of appropriate techniques tailored to
the specific objectives.
Model training and evaluation assess the performance of
predictive models using relevant metrics, while feature
importance analysis highlights key factors influencing
customer behaviour predictions. Finally, the result
interpretation and conclusion provide insights derived from
the analysis, offering recommendations to enhance
supermarket retail strategies.
The structured workflow for analyzing supermarket retail
strategies begins with data collection, where diverse datasets
encompassing sales transactions, customer demographics,
and product information are gathered. This comprehensive
dataset ensures a holistic view of the retail environment,
facilitating more accurate analysis and insights.
Subsequently, data preprocessing focuses on cleaning and
integrating the collected datasets to ensure data quality and
consistency. Tasks such as handling missing values,
removing duplicates, and standardizing formats prepare the
data for analysis and modelling. Customer segmentation
follows, aiming to categorize customers based on parameters
such as purchase history, demographics, and behaviour
patterns. By segmenting customers into distinct groups,
retailers can tailor marketing strategies, promotions, and
product offerings to meet the specific needs and preferences
of each segment. Predictive modelling utilizes advanced
algorithms to forecast customer behaviour, such as purchase
propensity, product preferences, or churn likelihood. Model
training and evaluation assess the performance of predictive
models using relevant metrics, ensuring the reliability and
effectiveness of the algorithms in making accurate
predictions and capturing underlying patterns in the data.
Feature importance analysis examines the significance of
different factors or variables in influencing customer
behaviour predictions, allowing retailers to prioritize
resources and efforts on the most impactful strategies.
Finally, the result interpretation and conclusion synthesize
the findings of the analysis, providing actionable insights for
improving supermarket retail strategies. By following this
structured workflow, retailers can leverage data-driven
insights to optimize operations, enhance customer
experiences, and drive business growth in the competitive
retail landscape.
The structured workflow for analyzing supermarket retail
strategies is a systematic approach designed to extract
actionable insights from data to enhance business decision-
making. It begins with meticulous data collection,
encompassing a wide range of datasets including sales
transactions, customer demographics, product information,
and other relevant sources. This comprehensive dataset
serves as the foundation for subsequent analysis, providing a
holistic view of the retail landscape.
This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2024.3407151
This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see https://creativecommons.org/licenses/by-nc-nd/4.0/
VOLUME XX, 2017 9
VII. PREDICTIVE ANALYSIS
For this, we will be using Neural Networks and SVC
• Data Preparation: Prepare the dataset with RFM
features (Recency, Frequency, Monetary) as input
variables and customer segments as the target
variable.
Encode categorical target variables (customer
segments) if necessary.
• Splitting Data: Split the dataset into training and
testing sets for model training and evaluation,
typically using a 70-30 or 80-20 split [9][10].
• Model Training: Train an SVC classifier using the
training data.
Train a Neural Network classifier using the training
data.
Experiment with different hyperparameters and
architectures to optimize model performance.
To train an SVC classifier and a Neural Network
classifier using the training data, we'll use Python's
scikit-learn library for SVC and TensorFlow/Keras
for the Neural Network. We will also experiment
with different hyperparameters and architectures to
optimize model performance.
This approach allows us to experiment with
different hyperparameters and architectures to
optimize the performance of both classifiers.
Adjustments can be made to the hyperparameter
and architecture search spaces based on the specific
requirements of the problem and the characteristics
of the dataset.
• Model Evaluation:
For SVC:
FIGURE 13. SVC Learning Curve.
The learning curve in Figure 13 depicts the relationship
between the model's performance (on the y-axis) and the size
of the training set (on the x-axis). In this specific case, the y-
axis represents the average score, which could be accuracy,
precision, recall, or F1 score.
Here are some observations based on the graph:
Overall Performance: The average score appears to be
relatively high across the entire training set size range,
indicating that the SVC model is performing well.
Training vs. Cross-Validation: The two curves in the graph
represent the training score (solid line) and the cross-
validation score (dashed line). The training score shows a
slight upward trend as the size of the training set increases,
which is expected as the model learns from more data. The
cross-validation score seems to fluctuate slightly but
generally stays around 0.97, suggesting that the model is not
overfitting the training data.
Limited Data: The x-axis only goes up to 3500, which might
be a relatively small dataset size for training complex models
like SVMs.
Overall, the graph suggests that the SVC model is likely
performing well on this specific dataset. However, it's
important to note that the generalizability of this model to
unseen data cannot be definitively determined from this
graph alone.
FIGURE 14. SVC Validation Curve.
Figure 14 features a detailed graph labelled "SVC
validation curve," showcasing the performance of a model
in terms of its training and cross-validation scores. The
vertical axis of the graph is marked with scores ranging
from 0.986 to 1.000, indicating the accuracy or
effectiveness of the model under evaluation. Horizontally,
the graph extends from 0 to 1000, possibly representing the
number of estimators or a regularization parameter denoted
as "C" that influences the model's complexity.
The graph in Figure 15 and Figure 16 displays two distinct
lines, one representing the "training score" and the other the
"cross-validation score." These lines seem to illustrate how
the model's performance varies with changes in the
parameter C, with both lines starting high and showing
This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2024.3407151
This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see https://creativecommons.org/licenses/by-nc-nd/4.0/
VOLUME XX, 2017 9
slight variations as C increases. The presence of these lines
suggests a careful analysis of the model's ability to
generalize from training data to unseen data, a crucial
aspect of machine learning model evaluation.
For MLP (Neural Networks):
FIGURE 15. MLP Learning Curve.
FIGURE 16. Heat map of the accuracy scores
VIII. SEQUENTIAL PATTERN MINING
The SPMF (Sequential Pattern Mining Framework)
algorithm is a collection of algorithms designed for
discovering sequential patterns in sequential data. This type
of algorithm is commonly used in analyzing transactional
datasets, such as those found in retail or e-commerce, to find
patterns in the order in which events occur. In the context of
the provided dataset, which contains transactional data, we
can use SPMF algorithms to discover sequential patterns in
customers' purchase behaviour.
• Data Preprocessing: Prepare the dataset in a
suitable format for sequential pattern mining. Each
transaction should be represented as a sequence of
items (products) purchased by a customer.
• Algorithm Selection: Prefix Span was chosen from
the SPMF framework based on the specific
requirements of the analysis.
• Parameter Tuning: The parameters such as
minimum support threshold, maximum pattern
length, and minimum pattern length should be
configured based on the characteristics of the
dataset and the desired patterns to be discovered.
• Pattern Mining: Apply the selected SPMF
algorithm to the preprocessed dataset to mine
sequential patterns [14].
The output of the algorithm will be a set of
sequential patterns representing common sequences
of items purchased by customers.
• Pattern Analysis: The discovered sequential
patterns should be analyzed to gain insights into
customer behaviour and preferences.
Identified frequent sequences of items that are often
purchased together or in a particular order.
Used visualization techniques to present the
discovered patterns in an understandable format.
• Interpretation and Action: Interpret the
discovered patterns to derive actionable insights for
business decision-making.
Utilize the insights to optimize product
recommendations, marketing strategies, and
inventory management processes.
Algorithm 1 Large-transaction generation algorithm using
prefixspan for extracting large 1-sequential patterns from
route database
Input:
RD (Route Database)
min_sup_count (Minimum support count)
Output:
r (Set of large 1-sequential patterns)
Method:
1: for each event (Bi, fi itemset;) in RD
2: if itemset, is not empty then
3: for each z itemset, // z is non-empty subset of
itemset,
4: if < B1, z > is not exist in I then
5: add y = < B1, z > to I', and set its sup_count
to 1;
6: else
7: increase sup_count of < Bj, z > by 1;
8: end if
9: end for
10: end if
11: end for
12: for each y = < Bj z > in I
13: if sup_count of y min_sup_count then
14: give z a unique symbol and save to I';
15: end if
16: end for
This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2024.3407151
This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see https://creativecommons.org/licenses/by-nc-nd/4.0/
VOLUME XX, 2017 9
IX. RESULTS AND DISCUSSION
A. PREDICTIVE MODEL RESULTS
Before presenting the classification reports for the Support
Vector Machine (SVM) and Multi-Layer Perceptron (MLP)
models, it's essential to highlight the significance of
predictive model evaluation in understanding the
performance of these algorithms. Evaluation metrics such as
precision, recall, and F1-score provide valuable insights into
the models' ability to correctly classify instances across
different classes [15]. These metrics, along with accuracy,
macro-average, and weighted-average scores, offer a
comprehensive view of the models' effectiveness in handling
the given dataset.
The following are the results from the predictive models
built:
TABLE II. Classification Report for SVM.
precision
recall
f1-score
support
high
0.99
1
0.99
312
low
1
1
1
232
medium
1
0.99
0.99
508
accuracy
0.99
1052
macro
Avg
0.99
0.99
0.99
1052
Weighted
avg
0.99
0.99
0.99
1052
TABLE III. Classification Report for MLP.
precision
recall
f1-score
support
high
1
1
0.99
312
low
1
1
1
232
medium
1
1
0.99
508
accuracy
0.99
1052
macro Avg
1
0.99
1
1052
Weighted
avg
0.99
1
0.99
1052
Table II and Table III show both the Support Vector
Classifier (SVC) and the Multi-Layer Perceptron (MLP)
models demonstrated excellent performance across all
classes with high precision, recall, and F1-scores. The
accuracy of both models was 0.99, indicating that they were
able to correctly classify instances into their respective
classes with a high degree of accuracy. The MLP model
showed slightly better performance in terms of the macro
average F1-score. Overall, both models are suitable for the
classification task and can be considered effective for
predicting customer segments based on the provided dataset.
B. SEQUENTIAL DATA MINING RESULTS
The following are the results obtained from the prefixspan
method:
>/home/Akila Bala/akila.bala24@gmail.com/OnlineRetail-
Master/spmf.jar
===== PREFIXSPAN 0.99-2016 – STATISTICS =====
Total time ~ 756 ms
Frequent sequences count: 419
Max memory (mb): 257.60943603515625
minsup = 159 sequences.
Pattern count: 419
Post-processing to show results in terms of string values.
Post-processing completed.
Based on the results obtained from the PrefixSpan algorithm
run:
• Total Time: The algorithm took approximately 756
milliseconds to complete.
• Frequent Sequences Count: There are 419 frequent
sequences found in the dataset.
• Minimum Support (minsup): The minimum support
threshold is 159 sequences.
• Pattern Count: A total of 419 patterns were
discovered.
• Max Memory: The maximum memory used during
the algorithm's execution was approximately
257.61 megabytes.
These results suggest that the PrefixSpan algorithm
successfully mined frequent sequential patterns from the
dataset with the specified minimum support threshold. The
discovered patterns can provide valuable insights into the
sequential behaviour of the data, which can be further
analyzed and interpreted for various purposes such as market
basket analysis, recommendation systems, or customer
behaviour analysis.
X. CONCLUSION
In the analysis conducted, we delved into a transactional
dataset to uncover valuable insights into customer behaviour
and preferences. Through thorough exploratory data analysis
(EDA), we identified trends such as popular products,
customer demographics, and seasonal sales patterns. Data
cleaning and preprocessing were crucial steps to ensure the
dataset's quality, including handling missing values and
removing irrelevant transactions like cancellations.
Leveraging RFM (Recency, Frequency, Monetary) features,
we employed predictive analysis techniques with Support
Vector Machine (SVM) and Neural Network classifiers to
accurately predict customer segments. Both models
exhibited impressive performance metrics, highlighting their
effectiveness in classifying instances into the correct
segments. Additionally, sequential pattern mining using the
PrefixSpan algorithm revealed frequent sequences of
customer purchase behaviour, offering valuable insights for
targeted marketing and personalized recommendations. By
integrating these analyses, businesses can optimize strategies
for inventory management, customer engagement, and
overall operational efficiency, ultimately driving growth and
enhancing customer satisfaction.
This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2024.3407151
This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see https://creativecommons.org/licenses/by-nc-nd/4.0/
VOLUME XX, 2017 9
REFERENCES
[1] Chen, D., Sain, S. L., & Guo, K. (2012). Data mining for the online
retail industry: A case study of RFM model-based customer
segmentation using data mining. Journal of Database Marketing &
Customer Strategy Management, 19, 197-208.
[2] Agarwal, P. (2014). Benefits and issues surrounding data mining and
its application in the retail industry. International Journal of Scientific
and Research Publications, 4(7), 1-5.
[3] Kumar, M. R., Venkatesh, J., & Rahman, A. M. Z. (2021). Data
mining and machine learning in retail business: developing
efficiencies for better customer retention. Journal of Ambient
Intelligence and Humanized Computing, 1-13.
[4] Li, H. (2005, June). Applications of data warehousing and data mining
in the retail industry. In Proceedings of ICSSSM'05. 2005
International Conference on Services Systems and Services
Management, 2005. (Vol. 2, pp. 1047-1050). IEEE
[5] Kohavi, R., Mason, L., Parekh, R., & Zheng, Z. (2004). Lessons and
challenges from mining retail e-commerce data. Machine Learning,
57, 83-113.
[6] Hormozi, A. M., & Giles, S. (2004). Data mining: A competitive
weapon for banking and retail industries. Information systems
management, 21(2), 62-71.
[7] Muley, P. A. (2022). Application of Data Mining Technique for Retail
Industry. In Sentimental Analysis and Deep Learning: Proceedings of
ICSADL 2021 (pp. 973-981). Springer Singapore.
[8] Zhang, X., Edwards, J., & Harding, J. (2007). Personalised online
sales using web usage data mining. Computers in Industry, 58(8-9),
772-782.
[9] Ahmeda, R. A. E. D., Shehaba, M. E., Morsya, S., & Mekawiea, N.
(2015, April). Performance study of classification algorithms for
consumer online shopping attitudes and behaviour using data mining.
In 2015 Fifth International Conference on Communication Systems
and Network Technologies (pp. 1344-1349). IEEE.
[10] Srikant, R., & Agrawal, R. (1996, March). Mining sequential patterns:
Generalizations and performance improvements. In International
conference on extending database technology (pp. 1-17). Berlin,
Heidelberg: Springer Berlin Heidelberg.
[11] Magnini, V. P., Honeycutt Jr, E. D., & Hodge, S. K. (2003). Data
mining for hotel firms: Use and limitations. Cornell Hotel and
Restaurant Administration Quarterly, 44(2), 94-105.
[12] Ritbumroong, T. (2015). Analyzing customer behaviour using online
analytical mining (OLAM). Integration of data mining in business
intelligence systems, 98-118.
[13] Hemalatha, M. (2012). Market basket analysis–a data mining
application in Indian retailing. International Journal of Business
Information Systems, 10(1), 109-129.
[14] Huang, C. K., Chang, T. Y., & Narayanan, B. G. (2015). Mining the
change of customer behaviour in dynamic markets. Information
Technology and Management, 16, 117-138.
[15] Research Synthesis by Systematic Literature Review and Data Mining
Techniques" Punpukdee, A., Wattana, C., Punpairoj, W.,
Srichuachom, U., Yaklai, P., & Trongtortam, S. (2021).
Dr. Kavitha Dhanushkodi is currently
working as an Assistant Professor in the
School of Computer Science and Engineering
(SCOPE) at Vellore Institute of Technology,
Chennai Campus, Chennai, Tamil Nadu, India.
She has completed PhD and Master of
Engineering in Computer Science and
Engineering from Anna University, Chennai,
Her research interests are mainly focused on
Software Security, Internet of Things and
Cyber Security. She has an overall teaching
experience of 14 years in various academic
institutions. She has published more than 40 Research papers to her
credit in reputed international journals and conferences with high
impact factors.
Akila Bala is currently a student Professor in
the School of Computer Science and
Engineering (SCOPE) at Vellore Institute of
Technology, Chennai Campus, Chennai,
Tamil Nadu, India. She is pursuing her
Master's degree which specializes in Business
Analytics, her research interests are mainly
focused on Network Security, Artificial
Intelligence and Data Science. She has
appeared in many conferences and published
one journal paper.
Nithin Kodipyaka is currently a student in the
School of Computer Science and Engineering
(SCOPE) at Vellore Institute of Technology,
Chennai Campus, Chennai, Tamil Nadu, India.
He is pursuing his Master's degree which
specializes in Business Analytics, His
research interests are mainly focused on
Information Security, Deep Learning and Data
Analytics. He has appeared in many
conferences and published one journal paper.
Shreyas V is currently a student in the School
of Computer Science and Engineering
(SCOPE) at Vellore Institute of Technology,
Chennai Campus, Chennai, Tamil Nadu, India.
He is pursuing his Master's degree which
specializes in Business Analytics, His
research interests are mainly focused on
Robotics Deep Learning and Data Mining.
This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2024.3407151
This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see https://creativecommons.org/licenses/by-nc-nd/4.0/
