Empowering Pharmaceutical Retail Storefronts: An Exploratory Study

on Classiﬁcation and Association Techniques

Humberto Tozetti Carlos

1 a

, Luciana Lee

2 b

and Mateus Barcellos Costa

1 c

Postgraduate Program in Applied Computing (PPComp),

Instituto Federal do Esp

ırito Santo, Serra, Brazil

CEUNES - Universidade Federal do Esp

ırito Santo, S

ao Mateus, Brazil

Keywords:

Recommendation Systems, Association Rules, Classiﬁcation Algorithms, Retail Analytics.

Abstract:

This work presents a study of association and classiﬁcation algorithms to support sales in retail stores through

recommendation systems. The study aimed to evaluate these algorithms in terms of their ability to provide

contextual information relevant to sales in retail storefronts. To achieve this goal, two primary objectives

were deﬁned. The ﬁrst was to explore methods for relating sales items. For this approach, experiments

were conducted using association rule and clustering algorithms. The second objective was to evaluate the

capability of classiﬁcation algorithms to identify classes of interest present within the data universe. The

experiments utilized a dataset from the pharmaceutical sector. In the case of association rule algorithms,

given the absence of data to enable recommendations based on collaborative ﬁltering, the purpose was to

identify patterns of item associations derived from customer shopping basket data. For the classiﬁcation

algorithms, the goal was to identify sales with and without medical prescriptions, a fundamental aspect to

monitor consumer behavior regarding the use of drugs. For identifying sales with medical prescriptions, the

MultiLayer Perceptron algorithm achieved the best results. For predicting items based on the shopping basket,

the best results were obtained by combined use of the K-Means, K-Prototype, and FP-Growth algorithms.

1 INTRODUCTION

Even after the normalization of the post-Covid-19

pandemic scenario, physical retail stores, as well as

various virtual commerce models that offer direct cus-

tomer service, face challenges with their frontline

staff, e.g., monotony, insufﬁcient incentives, and in-

adequate training, leading to disengagement and high

turnover rates (Ahmed et al., 2024).

In order to mitigate frontline retail staff hindrances

to dealing with product information, many retail sec-

tors seek support from recommendation systems with

functions such as the suggestion of items based on a

customer’s purchase history or according to the user’s

proﬁle. For the frontline retail staff, however, contex-

tual knowledge is crucial for current purchase (Mo-

hamed et al., 2019; Knyazev and Oosterhuis, 2023).

In this context, this article examines item asso-

ciation algorithms, which focus on generating rec-

ommendations for complementary items in a sale, as

https://orcid.org/0009-0009-2052-4083

https://orcid.org/0000-0003-1335-074X

https://orcid.org/0000-0002-4235-5411

well as classiﬁcation algorithms designed to identify

sales based on relevant classes within the applica-

tion scenario. However, our focus is on strengthen-

ing the salesperson, pharmacist in the case of drug-

stores, with tools to support decision-making. In the

case of medicines, in addition to mitigating risks, as

it directly involves the consumer’s health, using tools

such as recommendation systems can empower an in-

experienced seller to provide better service.

Both approaches investigated were based on the

analysis of similarity between individual sales and

their items through the evaluation of sale receipts,

along with supplementary data such as the Hu-

man Development Index (HDI) of the store’s loca-

tion and the time of day during which the sale oc-

curred. The study considered the MultiLayer Per-

ceptron classiﬁer with Deep-Learned Embeddings to

predict whether a sale involves a medical prescrip-

tion and the FP-Growth algorithm for product-based

recommendations. The algorithms were implemented

and validated using a dataset from the pharmaceutical

sector, encompassing four geographically dispersed

units within a single metropolitan region in Brazil.

326

Carlos, H. T., Lee, L. and Costa, M. B.

Empowering Pharmaceutical Retail Storefronts: An Exploratory Study on Classiﬁcation and Association Techniques.

DOI: 10.5220/0013429000003929

Paper published under CC license (CC BY-NC-ND 4.0)

In Proceedings of the 27th International Conference on Enterprise Information Systems (ICEIS 2025) - Volume 1, pages 326-334

ISBN: 978-989-758-749-8; ISSN: 2184-4992

The results demonstrate that the proposed algorithms

achieve high levels of accuracy, effectively associat-

ing sales items and providing reliable recommenda-

tions.

The remainder of this paper is organized as fol-

lows: Section 2 provides the conceptual background.

Section 3 outlines the experimental process and the

solutions considered for item association and classi-

ﬁcation. Section 4 presents and analyzes the exper-

imental results. Section 5 examines the ﬁndings in

the light of related work, and Section 6 concludes the

paper with suggestions for future research directions.

2 BACKGROUND

Recommendation systems to support sales in retail

environments aim to establish relationships between

individuals and products. This relationship building

process explores data sources, extracting meaningful

insights through approaches commonly referred to as

ﬁltering. Four types of ﬁltering are frequently em-

ployed: collaborative, content-based, demographic,

and hybrid ﬁltering (Mohamed et al., 2019).

Various methods and different information

sources can support ﬁltering. In virtual buying and

selling environments, where information is abundant,

recommendations for a buyer are typically based on

previous evaluations of their purchasing activities,

either through their preference history or by analyz-

ing similar purchases. Recommendations in such

contexts are generally constructed based on three

mechanisms: cross-selling, direct recommendations,

and gift centers (Lakshmi and Lakshmi, 2014).

In physical retail sales, the amount of informa-

tion, usually available in virtual environments, is of-

ten absent. Recommendations can therefore lead to a

scarcity of data directly related to purchase choices.

Available data, such as sales receipts, may be highly

dispersed, requiring appropriate pre-processing. In

the pharmaceutical retail sector, direct recommenda-

tions are particularly notable, often bypassing ﬁlter-

ing methods due to the lack of evaluative data about

purchases.

By leveraging sales receipts as a source of ﬁlter-

ing information, such methods can be adapted and

employed. This approach is commonly referred to

as Market Basket Analysis (MBA). MBA employs

data mining techniques to identify relationships be-

tween products frequently purchased together. This

technique proves highly useful in scenarios without

customer data, such as a ﬁrst-time purchase or the ab-

sence of purchase records. To generate and identify

these relationships, MBA uses algorithms that pro-

duce association rules, incorporating metrics such as

conviction and conﬁdence to establish relationships

between products (

Unvan, 2021).

2.1 Association Rule Algorithms

Association rule algorithms aim at identifying rela-

tionships between items in a dataset, revealing co-

occurrence trends among products even without ex-

plicit correlations (Patil et al., 2014).

The common notation to represent association

rules is (A ⇒ B),where A is the antecedent, i.e., the

item or event that precedes, and B is the consequent,

i.e., the item or event that is likely to occur in the pres-

ence of A.

To evaluate the strength of an association, three

commonly used metrics are support, conﬁdence and

lift (Yazgana and Kusakci, 2016; Sagin and Ayvaz,

2018):

• Support: measures the proportion of transactions

containing both A and B.

• Conﬁdence: evaluates the probability of B be-

ing purchased given that A has already been pur-

chased.

• Lift: This measure evaluates the correlation be-

tween two items, A and B. A value below 1 in-

dicates a negative correlation, implying that the

presence of the ﬁrst item suggests the absence of

the second. Conversely, a value greater than 1 sig-

niﬁes a positive correlation, meaning that the oc-

currence of one item suggests the occurrence of

the other.

The FP-Growth algorithm is an efﬁcient alterna-

tive, particularly for large datasets, to generate associ-

ation rules. The algorithm operates in two main steps:

First, it constructs an FP-Tree by ﬁltering out infre-

quent items based on a minimum support threshold

and organizing the remaining items in descending or-

der of frequency, optimizing storage by sharing com-

mon paths. Then, it mines the FP-Tree by generating

conditional FP-Trees from the least frequent nodes to

efﬁciently extract frequent patterns.

2.2 Classiﬁcation

Classiﬁcation algorithms have broad applicability,

widely used in pattern recognition and prediction

tasks. In the context of retail storefronts, which served

as the environment for this study, classiﬁers were em-

ployed to enable the qualiﬁed use of available sales

data. In this sense, they were utilized as an analytical

method to provide deeper insights into sales patterns

and behaviors.

Empowering Pharmaceutical Retail Storefronts: An Exploratory Study on Classiﬁcation and Association Techniques

327

According to (Dahouda and Joe, 2021), key met-

rics for evaluating classiﬁcation algorithms include:

• Accuracy: The proportion of correct predictions

out of the total predictions made. While intuitive,

it may produce misleading results in imbalanced

datasets.

• Precision: The proportion of true positives results

out of all positives predicted by the classiﬁer, re-

ﬂecting its ability to avoid false positives.

• Recall: The proportion of true positives out of all

actual positives samples, assessing the classiﬁer’s

ability to identify true positives.

• F1-Score: A harmonic mean of precision and re-

call, providing a balanced measure of these met-

rics.

An MultiLayer Perceptron (MLP), in addition to

being a supervised classiﬁcation method, is a type

of artiﬁcial neural network inspired by the function-

ing of biological neurons (Murtagh, 1991; Savalia and

Emamian, 2018). It consists of interconnected layers,

organized into three or more levels:

• Input Layer: Receives input data, where each

neuron represents a feature in the dataset (Yaacob

et al., 2013).

• Hidden Layer(s): Comprises one or more levels

of neurons that process information from the input

layer. The number of hidden layers and neurons

are conﬁgurable parameters (Yaacob et al., 2013).

• Output Layer: Outputs the results of the model,

with each neuron representing a predicted class

(Savalia and Emamian, 2018).

To introduce non-linearity into the model, MLP

uses an activation function that applies a weighted

sum of the inputs for each neuron (Murtagh, 1991;

Yaacob et al., 2013). Learning is typically achieved

using the backpropagation algorithm to update con-

nection weights during training, minimizing predic-

tion errors (Murtagh, 1991).

Categorical data, such as demographic attributes,

are challenging for machine learning algorithms due

to their lack of numerical values. MLP facilitates the

handling of categorical data by using techniques like

Deep-Learned Embeddings, which convert such data

into vector representations suitable for the algorithm.

2.3 Clustering

Clustering is an unsupervised machine learning tech-

nique, meaning it does not rely on predeﬁned labels

to identify groups (B

ejar Alonso, 2013).

Clustering can be applied to a wide variety of ap-

plications. In the context of retail, it is particularly rel-

evant for customer segmentation, allowing customers

to be grouped based on their purchasing behavior

(Jridi et al., 2020). It can also be applied to sales

data to group transactions according to the items in

the shopping cart.

There are several algorithms for clustering, in-

cluding:

• K-Means: This algorithm groups data by initializ-

ing k centroids randomly and uses Euclidean dis-

tance to assign points to the nearest centroid. The

centroids are updated iteratively as the mean of

the points in each group, continuing until they no

longer change (Kuzelewska, 2014). It is simple

and efﬁcient for numerical data.

• K-Prototype: An extension of K-Means, it han-

dles mixed data types (numerical and categor-

ical). It combines distance measures for each

type: Euclidean distance for numerical and 0 for

equal, 1 for different values for categorical at-

tributes. Centroids are updated with the mean for

numerical and the most frequent value for categor-

ical attributes (Jridi et al., 2020; Haj et al., 2015;

ejar Alonso, 2013).

To optimize clustering, especially in selecting the

optimal number of clusters, the elbow method is fre-

quently employed. This technique plots the within-

cluster sum of squared errors (WSS) against the num-

ber of clusters. While WSS decreases as more clus-

ters are added, the improvement eventually plateaus,

creating an “elbow” in the graph that shows the ideal

number of clusters (Syakur et al., 2018).

3 EXPERIMENTS

This section illustrates the proposed experiments and

the database used.

3.1 Dataset

For the experiments, a real-world sales transaction

database was utilized, representing a network of phar-

macies geographically distributed across a large and

diverse metropolitan area in Brazil, with 282,513

sales, carried out in 2019, in 4 stores, with a total

of 2,546 unique products. The experimental dataset

comprises anonymized data from this database, free

of sensitive information and thoroughly cleaned. The

columns described in Table 1 are included.

In addition to Table 1 columns, the dataset also

includes one column for each item sold at least once.

ICEIS 2025 - 27th International Conference on Enterprise Information Systems

328

Table 1: Dataset Columns.

Column Description

quarter Identiﬁes the sales quarter

day of week Identiﬁes the day of the week

period Identiﬁes the time shift (morning, afternoon,

or evening)

has prescription Indicates whether a medical prescription was

issued.

hdi Human Development Index (HDI) of the

stores location

These columns use binary values to indicate the pres-

ence of an item in a sale. Each row in the dataset cor-

responds to a transaction conducted within the phar-

macy network. The values for these columns are de-

ﬁned as follows:

1. A value of 0 indicates that the item is not present

in the sale;

2. A value of 1 indicates that the item is present in

the sale.

3.2 Prediction of Items Based on the

Market Basket

To enhance product recommendations at retail store-

fronts and support decision-making, association rules

were applied. Their primary purpose was to uncover

correlations between products in transaction records.

In the experiment, the technique was speciﬁcally used

to develop a recommendation system that relies exclu-

sively on product transaction history, without requir-

ing customer-speciﬁc data.

To generate the association rules, the FP-Growth

algorithm was employed to create a model capable of

suggesting the next item for a customer’s cart based

on the current basket.

Given the variety of items that may be present in

sales, as well as the numerous combinations of these

items in each transaction, clustering techniques were

applied to group transactions with similar character-

istics into the same cluster. This facilitated the gener-

ation of association rules among the products.

Several libraries support clustering algorithms,

with the primary requirement being the deﬁnition of

the number of clusters to generate. To determine this

number, the elbow curve method was used. Based on

its results, clustering was performed to create ﬁve dis-

tinct groups. The elbow curve method graph was gen-

erated using the K-means clustering algorithm. This

choice was made due to the need to analyze the items

speciﬁcally—since they are the objects of associa-

tion—and because K-means is well-suited for numer-

ical data. However, for grouping transactions, the K-

Prototype algorithm was used to better handle the cat-

egorical data in the dataset. In the case of transaction

clustering, categorical data can help identify sales pat-

terns based on period, location, or even day of the

week. These measures were implemented to enhance

the richness and accuracy of transaction clusters.

Following this, the data were processed, as the as-

sociation rules required only item-level information

without necessarily considering attributes like HDI,

day of the week, or other details. While these at-

tributes were relevant initially for clustering, they

were not essential for generating the associations.

The FP-Growth algorithm was used to create the

association rules, as it is a strong alternative for han-

dling the cardinality and complexity of the dataset.

For generating associations, a conﬁdence threshold of

30% was used as an input parameter. This algorithm

was implemented for both the clustered and original

datasets to maximize the potential relationships be-

tween items, thereby increasing the possibilities for

recommendations.

Given the number of association rules generated

using the original database, it was found that most of

the items did not have rules to present results in the

model. In order to achieve parity between the models,

the support value for generating the association rules

was reduced in order to obtain a sufﬁcient quantity to

compare the results of the recommendation models.

Three recommendation models can be generated:

the ﬁrst model is achieved from association rules de-

rived from the dataset without clustering . The sec-

ond, from association rules derived from the clustered

dataset. Finally, the third model is obtained from the

integration of both previous strategies. The Figure 1

illustrates the workﬂow of the model generation pro-

cess in terms of a BPMN diagram.

Figure 1: Association Experimentation ﬂow.

3.3 Identiﬁcation of Sales with Medical

Prescriptions

The MultiLayer Perceptron (MLP) classiﬁer was used

in combination with Deep-Learned Embeddings in

order to address the problem of identifying medi-

cal prescriptions in sales. The objective was to cre-

ate a model capable of predicting, based on the in-

put shopping cart, whether the sale involves a medi-

cal prescription, even when a prescription is not pro-

Empowering Pharmaceutical Retail Storefronts: An Exploratory Study on Classiﬁcation and Association Techniques

329

vided. The choice of MLP was motivated by its inte-

gration with Deep-Learned Embeddings, enabling the

use of categorical data present in the dataset to en-

hance the classiﬁcation process. Moreover, the Mul-

tiLayer Perceptron has demonstrated its performance

when compared with statistical approaches, as pro-

posed by (Murtagh, 1991). This information is not

only valuable for understanding consumer behavior

regarding medications but can also be useful for ﬁl-

tering items to recommend to the customer.

To enable the combined use of these algorithms,

initial preprocessing of the dataset was necessary.

This requirement stemmed from the presence of cat-

egorical data in the transaction records and an im-

balance in the has prescription class. Each sale

record included not only the products sold but also

data categorizing the day, time period, location, and

quarter in which the sales occurred. Incorporating

this information enhances the pattern recognition pro-

cess among transactions. To integrate these categor-

ical features into the classiﬁer, a transformation was

performed using Deep-Learned Embeddings, with the

resulting embeddings serving as input for training the

classiﬁcation model.

In addition to categorical data, it was necessary

to address the imbalance in the has prescription

class, as only approximately 11% of sales include a

registered medical prescription. Although this indi-

cates the presence of such records in the dataset, it

does not mean the sale was conducted with a prescrip-

tion, as the registration of this information is optional

in cases where the prescription is not mandatory. Im-

plementing the classiﬁer and model with such a dis-

parity between classes would result in poor outcomes.

To mitigate this, the compute class weight function

from the Python sklearn library was employed. The

output of this function is used to create a dictionary,

which is further used by the ﬁt function of the classi-

ﬁer model.

For building the model, the MLP algorithm was

implemented using Python and the Keras library, as

its solution for the MultiLayer Perceptron works ide-

ally with the preprocessed categorical data. The in-

put parameters included a maximum of 50 epochs for

training, a split of 70% of the records for training and

30% for testing, and an early stopping mechanism to

accelerate model training when improvements cease.

Following this process, the classiﬁcation results

were evaluated, achieving an accuracy of 97.97%.

With the classiﬁcation process completed and

yielding favorable results, it was possible to develop

a prototype. This prototype receives data from a sale,

identifying the items present in a transaction and cal-

culates the likely class of the sale. Figure 2 illustrates

the ﬂowchart of the described process.

Figure 2: Classiﬁcation Experimentation Flow.

The experiments were conducted on a workstation

equipped with an Intel i7 11th generation processor,

32GB of RAM, and a GeForce RTX 3060 GPU with

6GB of memory. All algorithms were implemented

in Python using libraries such as NumPy, Pandas,

Scikit-Learn, Seaborn, Matplotlib, Pyarrow, Kmodes,

OS, Mlxtend, GC, TensorFlow, Random, and Yellow-

Brick.

4 ANALYSIS OF RESULTS

This section discusses the results of the algorithms

and solutions implemented in the experimental phase.

4.1 Prediction of Items Based on the

Market Basket

In order to validate the product recommendation

model, the ten most frequently sold items were se-

lected. For each item, sales transactions that included

the product and at least one additional item were ﬁl-

tered. These transactions were used to assess recom-

mendations based on historical sales data. The model

also allows setting the number of items to recom-

mend. In this experiment, the top ﬁve recommenda-

tions were analyzed.

Three approaches were employed to rank the most

recommended items. The ﬁrst approach measured the

correlation strength between items. The second ap-

proach relied on the conﬁdence level of the associa-

tion. Finally, the third approach used rule support as

the ranking criterion.

In addition to ranking methods, three sources of

association rules were used to generate recommen-

dations. The ﬁrst source derived rules directly from

the database without clustering. The second source

generated rules from ﬁve clusters created using the

K-Prototype algorithm. Finally, the third source com-

bined the rules from the previous two methods into a

uniﬁed set of association rules.

To evaluate the performance of the recommenda-

tion models, the fulﬁllment rate was employed. The

ICEIS 2025 - 27th International Conference on Enterprise Information Systems

330

following equation demonstrates its calculation.

Fulﬁllment Rate (%) =

(1)

Where:

• N

: Number of transactions containing recom-

mended items.

• N

: Total number of transactions.

The resulting fulﬁllment rates for each dataset are

presented in Tables 2, 3, and 4.

Table 2: Fulﬁllment rates by product based on orderings by

Conﬁdence, Lift, and Support in the original dataset.

Product

Conﬁdence Lift Support

(%) (%) (%)

Product 1 7.59 7.59 6.24

Product 2 31.28 31.28 42.73

Product 3 40.86 40.86 53.10

Product 4 15.24 15.24 7.58

Product 5 50.21 50.21 46.86

Product 6 22.43 22.43 54.51

Product 7 6.44 6.44 13.95

Product 8 6.37 6.37 6.12

Product 9 27.84 27.84 38.99

Product 10 14.83 14.83 30.73

Table 3: Fulﬁllment rates by product according to the

metrics of Conﬁdence, Lift, and Support in the clustered

dataset.

Product

Conﬁdence Lift Support

(%) (%) (%)

Product 1 6.09 7.59 5.13

Product 2 28.66 20.61 45.45

Product 3 50.19 40.82 53.10

Product 4 20.61 18.32 13.29

Product 5 42.54 38.42 47.42

Product 6 43.98 36.71 54.51

Product 7 13.08 13.08 7.68

Product 8 4.95 7.55 8.29

Product 9 18.51 23.13 40.76

Product 10 14.58 26.80 34.26

Table 4: Fulﬁllment rates by product according to the met-

rics of Conﬁdence, Lift, and Support in the merged dataset.

Product

Conﬁdence Lift Support

(%) (%) (%)

Product 1 9.94 7.59 6.74

Product 2 20.52 5.24 45.22

Product 3 50.19 40.82 53.10

Product 4 17.47 18.32 13.29

Product 5 39.17 38.42 47.42

Product 6 43.98 36.71 54.51

Product 7 6.09 6.09 7.68

Product 8 6.06 6.37 8.29

Product 9 27.11 19.05 40.76

Product 10 14.58 22.36 34.26

As can be observed, regardless of the source of the

rules, ordering by support consistently delivers better

results. This is evident when analyzing the number

of items with higher fulﬁllment rates when sorted by

Support across each dataset.

The support metric also exhibits a higher average

fulﬁllment rate by products across all datasets, as il-

lustrated in Figure 3.

Figure 3: Comparative graph of the average fulﬁllment rate

for each dataset, sorted by metric.

In addition to analyzing the sorting metric with

the best results, the relevance of each database was

assessed. This evaluation considered the highest-

ranking values from each source to generate associa-

tion rules for each product. Speciﬁcally, only the best

fulﬁllment rate per item was used for each ranking.

Table 5 presents the highest-ranking values from each

source and product.

Table 5: Highest Values Among Metrics by Database and

Product.

Product

Original Clustered Merged

(%) (%) (%)

Product 1 7.59 7.59 9.94

Product 2 42.73 45.45 45.22

Product 3 53.10 53.10 53.10

Product 4 15.24 20.61 18.32

Product 5 50.21 47.42 47.42

Product 6 54.51 54.51 54.51

Product 7 13.95 13.08 7.68

Product 8 6.37 8.29 8.29

Product 9 38.99 40.76 40.76

Product 10 30.73 34.26 34.26

By identifying the best fulﬁllment rates for each

item across databases, it is possible to determine, for

a given item, the scenarios in which each database

would deliver the best performance. For this analysis,

a database was counted as the best performer when-

ever its fulﬁllment rate was the highest among all

databases. Building on the results from Table 5, Table

6 was created to summarize the number of items for

which each database showed the greatest relevance.

It is worth noting that in cases like Product 3,

where all databases exhibit the same fulﬁllment rate,

Empowering Pharmaceutical Retail Storefronts: An Exploratory Study on Classiﬁcation and Association Techniques

331

the best-performing status is attributed to each of

these databases. This approach ensures a comprehen-

sive view of which databases would most effectively

satisfy the majority of recommendations.

Table 6: Predominance Count by Database.

Dataset

Predominance

Count

Original 4

Clustered 7

Merged 6

Based on these ﬁndings, the clustered database

demonstrates the highest relevance, even outperform-

ing the merged strategy, which integrates rules from

both the clustered and original databases. This superi-

ority arises from the fact that, in percentage terms, the

clustered database delivers more satisfactory results

in most scenarios. Conversely, the original database

exhibits the weakest performance, showing low rele-

vance in the majority of cases. Thus, clustering tech-

niques have proven to be a valuable tool for uncover-

ing meaningful association rules, thereby contributing

to the development of more effective item recommen-

dation models.

4.2 Identiﬁcation of Sales with Medical

Prescriptions

In addition to accuracy, precision, recall, and F1-score

metrics were used to assess the model’s performance.

Figure 4 provides the report summarizing these three

metrics, while Figure 5 displays the confusion matrix

produced by the model.

The Classiﬁcation Report highlights precision, re-

call, and F1-score, offering a detailed assessment of

the model’s performance. In this experiment, the

model showed stronger performance in classifying

class 0 but also achieved solid metrics for class 1,

which represents sales with reported revenue. Al-

though the false-positive rate was higher for class 1,

this outcome is expected, as many sales in this cate-

gory were not labeled at the retail storefront.

Figure 4: Classiﬁcation metrics results.

The methodology for validating the generated

Figure 5: Confusion matrix generated by the model.

classiﬁcation model consisted of two stages. In the

ﬁrst stage, manual testing was conducted with ex-

pected results to verify whether the model performed

as anticipated. Upon obtaining positive outcomes, the

second stage was initiated, which focused on assess-

ing the model’s ability to generalize and classify sales

as medical prescription transactions.

In the manual testing stage, several sales trans-

actions excluded from the training and testing pro-

cesses were selected and manually input into the

model to observe the results. The primary objective

was to validate whether transactions involving prod-

ucts requiring mandatory prescription retention—i.e.,

those that cannot be sold without a prescription un-

der any circumstances—were correctly classiﬁed as

prescription-required. Additionally, the goal was to

verify if the model’s behavior in this controlled envi-

ronment aligned with its performance during training.

For transactions requiring mandatory prescrip-

tions, the model accurately identiﬁed all such cases

as prescription-required. For other transactions, the

model’s performance was consistent with its training

results, correctly classifying sales as either requiring

or not requiring a prescription.

A test was conducted with two speciﬁc products to

assess the model’s ability to classify transactions as

prescription-required, even when initially labeled as

non-prescription. These products, while not manda-

torily prescribed, are seldom sold without a doctor’s

recommendation due to their use in treating hyper-

tension and diabetes. In this test, 314 transactions

involving both items were analyzed, of which 243

lacked prescription records, accounting for approxi-

mately 77.4% of the sales.

Focusing on the 243 transactions without prescrip-

tion records, the model’s predictions were further ex-

amined. It determined that 184 of these transactions

should have included a prescription, accounting for

approximately 75.8% of the sales.

The model suggests that, instead of the observed

22% of transactions involving a prescription, the ac-

tual proportion is likely closer to 81%, offering a more

ICEIS 2025 - 27th International Conference on Enterprise Information Systems

332

precise representation of expected medication con-

sumption patterns.

5 RELATED WORK

Yoosofan et al. (2015), Kusumo et al. (2021) and

Unvan (2021) have applied association rules in the

retail sector using the FP-Growth and A Priori algo-

rithms. These techniques aimed to support storefront

operations by organizing inventory and shelves and

associating items to design sales campaigns. Yooso-

fan et al. (2015) and Kusumo et al. (2021) applied

association rules to medication databases, neither de-

veloped recommendation systems. In contrast, Guru-

dath (2020) proposed a collaborative ﬁltering-based

recommendation system that, in addition to utilizing

association rule algorithms, leverages item relation-

ship information through techniques such as market

basket analysis.

Lakshmi and Lakshmi (2014), Mohamed et al.

(2019) and Ricci et al. (2015) address challenges and

solutions in recommendation systems, outlining types

of ﬁltering, application examples, and their limita-

tions. Mohamed et al. (2019) and Ricci et al. (2015)

discuss the relevance of these systems in enhancing

customer satisfaction and increasing revenue through

boosted sales. Additionally, Mohamed et al. (2019)

and Ricci et al. (2015) emphasize the importance of

techniques such as classiﬁcation, clustering, and as-

sociation rules in the context of recommendation sys-

tems, which were fundamental to the work presented

in this study.

Nistal-Nuno (2022) discusses recommendation

systems for pharmaceutical e-commerce based on

Bayesian User Modeling to suggest items to cus-

tomers. The justiﬁcation lies in providing conve-

nience for customers to ﬁnd all necessary medications

on the website. The application was user-oriented

rather than product-oriented, employing clustering

techniques to group similar users. Notably, associ-

ation rule techniques were not applied. Conversely,

Tercan et al. (2021) implemented recommendation

systems using embedding techniques and product

similarity, in addition to artiﬁcial neural networks

trained with the user’s browsing history.

Murtagh (1991) examines the approach of the

Multilayer Perceptron algorithm, describing its con-

cepts, functionality, and applications in various ﬁelds,

while comparing its efﬁciency by referencing other

authors and studies.

In the work discussed by Dahouda and Joe (2021),

the use of Deep-Learned Embeddings is demon-

strated, comparing its efﬁciency to other techniques

for handling categorical data, such as the One-Hot En-

coding algorithm.

Potolea and Lemnaru (2011) address the effect of

imbalance classes and how they impacted the perfor-

mance of classiﬁers. According to the authors, the use

of balancing methods was crucial for developing the

proposed model. The authors also discuss the effect

of class imbalance and its impact on the performance

of classiﬁers. In our work, the use of balancing meth-

ods was also crucial to identify the presence of med-

ical prescriptions in medication sales, given the class

imbalance as discussed in Section 3.

6 CONCLUSION

This study explored the use of data mining and ma-

chine learning algorithms to support retail pharmacy

storefront operations. The primary objective was to

provide actionable insights to employees, particularly

customer service staff, by analyzing customers’ shop-

ping baskets. For instance, classifying a sale as orig-

inating from a medical prescription enables profes-

sionals to guide customers more conﬁdently or follow

speciﬁc procedures as required.

The application of association rule techniques,

leveraging the FP-Growth algorithm supported by

clustering mechanisms, proved effective in develop-

ing a recommendation system, achieving an average

success rate of over 30% when recommending items

based on past transactions. Additionally, the use of a

MultiLayer Perceptron classiﬁer in combination with

Deep-Learned Embeddings to label sales based on

their items yielded an accuracy of 97.97%.

Currently, the prediction model relies on associa-

tions between items already present in the database.

While this is feasible due to the database’s com-

prehensiveness, it may become a limitation as the

database evolves, with items being added or removed.

To address this, implementing fuzzy comparison tech-

niques—generating similarity scores for item descrip-

tions—could be a viable solution.

For further improving storefront support, accurate

and well-contextualized item predictions are critical.

Incorporating additional relevant information to en-

hance the model’s ability to suggest subsequent items

is key. A potential improvement to the proposed rec-

ommendation model involves integrating sales classi-

ﬁcation into the recommendation process. This would

allow for generating recommendations tailored to sce-

narios involving or not involving medical prescrip-

tions, better aligning item associations. The perfor-

mance of this enhanced model could then be com-

pared to the one presented in this study.

Empowering Pharmaceutical Retail Storefronts: An Exploratory Study on Classiﬁcation and Association Techniques

333

ACKNOWLEDGEMENTS

We would like to thank the Graduate Program in Ap-

plied Computing at Ifes Serra (PPCOMP) and the

Capixaba Open University Program (UnAC) of the

Secretariat for Science, Technology, Innovation, and

Professional Education (SECTI) of the Government

of the State of Esp

ırito Santo, Brazil, for their support

in the development of this work. We also extend our

special thanks to the Esp

ırito Santo Research and In-

ovation Support Foundation (FAPES) for its ﬁnancial

support of this study.

REFERENCES

Ahmed, M. N., Akaak, A. S., et al. (2024). Integrated role

of human resource management in mitigating the post-

covid-19 challenges in retail. Business Ethics and Lead-

ership, 8(1):1–14.

ejar Alonso, J. (2013). Strategies and algorithms for clus-

tering large datasets: a review.

Dahouda, M. K. and Joe, I. (2021). A deep-learned embed-

ding technique for categorical features encoding. IEEE

Access, 9:114381–114391.

Gurudath, S. (2020). Market basket analysis & recommen-

dation system using association rules. Master of Science

in Big Data Management and Analytics, Grifﬁth College,

Dublin.

Haj, K. M. B., Essoussi, N., et al. (2015). “mapreduce-

based k-prototypes clustering method for big data. In

2015 IEEE International Conference on Data Science

and Advanced Analytics (DSAA). Paris: Campus des

Cordeliers.

Jridi, H., Ben HajKacem, M. A., and Essoussi, N. (2020).

Parallel k-prototypes clustering with high efﬁciency and

accuracy. In Big Data Analytics and Knowledge Dis-

covery: 22nd International Conference, DaWaK 2020,

Bratislava, Slovakia, September 14–17, 2020, Proceed-

ings 22, pages 380–395. Springer.

Knyazev, N. and Oosterhuis, H. (2023). A lightweight

method for modeling conﬁdence in recommendations

with learned beta distributions. In Proceedings of the

17th ACM Conference on Recommender Systems, pages

306–317.

Kusumo, H., Marlina, D., Novita, M., and Anwar, M.

(2021). Analysis of transaction patterns at drug store

with apriori algorithm. In Journal of Physics: Confer-

ence Series, volume 1869, page 012070. IOP Publishing.

Kuzelewska, U. (2014). Clustering algorithms in hybrid

recommender system on movielens data. Studies in logic,

grammar and rhetoric, 37(1):125–139.

Lakshmi, S. S. and Lakshmi, T. A. (2014). Recommenda-

tion systems: Issues and challenges. International Jour-

nal of Computer Science and Information Technologies,

5(4):5771–5772.

Mohamed, M. H., Khafagy, M. H., and Ibrahim, M. H.

(2019). Recommender systems challenges and solutions

survey. In 2019 international conference on innovative

trends in computer engineering (ITCE), pages 149–155.

IEEE.

Murtagh, F. (1991). Multilayer perceptrons for classiﬁca-

tion and regression. Neurocomputing, 2(5-6):183–197.

Nistal-Nuno, B. (2022). Medication recommendation sys-

tem for online pharmacy using an adaptive user interface.

Computer Methods and Programs in Biomedicine Up-

date, 2:100077.

Patil, P. H., Thube, S., Ratnaparkhi, B., and Rajeswari, K.

(2014). Analysis of different data mining tools using

classiﬁcation, clustering and association rule mining. In-

ternational Journal of Computer Applications, 93(8).

Potolea, R. and Lemnaru, C. (2011). A comprehensive

study of the effect of class imbalance on the performance

of classiﬁers. In ICEIS (1), pages 14–21.

Ricci, F., Rokach, L., and Shapira, B. (2015). Recom-

mender systems: introduction and challenges. Recom-

mender systems handbook, pages 1–34.

Sagin, A. N. and Ayvaz, B. (2018). Determination of asso-

ciation rules with market basket analysis: application in

the retail sector. Southeast Europe Journal of Soft Com-

puting, 7(1).

Savalia, S. and Emamian, V. (2018). Cardiac arrhythmia

classiﬁcation by multi-layer perceptron and convolution

neural networks. Bioengineering, 5(2):35.

Syakur, M. A., Khotimah, B. K., Rochman, E., and Satoto,

B. D. (2018). Integration k-means clustering method and

elbow method for identiﬁcation of the best customer pro-

ﬁle cluster. In IOP conference series: materials science

and engineering, volume 336, page 012017. IOP Pub-

lishing.

Tercan, H., Bitter, C., Bodnar, T., Meisen, P., and Meisen, T.

(2021). Evaluating a session-based recommender system

using prod2vec in a commercial application. In ICEIS

(1), pages 610–617.

Unvan, Y. A. (2021). Market basket analysis with asso-

ciation rules. Communications in Statistics-Theory and

Methods, 50(7):1615–1628.

Yaacob, H., Abdul, W., and Kamaruddin, N. (2013). Clas-

siﬁcation of eeg signals using mlp based on categorical

and dimensional perceptions of emotions. In 2013 5th

International Conference on Information and Communi-

cation Technology for the Muslim World (ICT4M), pages

1–6. IEEE.

Yazgana, P. and Kusakci, A. O. (2016). A literature survey

on association rule mining algorithms. Southeast Europe

Journal of soft computing, 5(1).

Yoosofan, A., Ghajar, F. G., Ayat, S., Hamidi, S., and

Mahini, F. (2015). Identifying association rules among

drugs in prescription of a single drugstore using apriori

method. Intelligent Information Management, 7(5):253–

259.

ICEIS 2025 - 27th International Conference on Enterprise Information Systems

334