Scientia et Technica Año XXVIII, Vol. 30, No. 01, enero-marzo de 2025. Universidad Tecnológica de Pereira. ISSN 0122-1701 y ISSN-e: 2344-7214
26
I
Abstract— This research work presents the development of
a model using data mining techniques to identify financial
variables in a manufacturing company of automotive
vehicle bodies in Pereira. The study is structured into four
key phases. The first phase focuses on data preprocessing,
including characterization, normalization, and
dimensionality reduction using PCA, Relief, and
Correlation. The second phase applies unsupervised
learning with K-means and Gaussian Mixture Models
(GMM) to cluster and validate data based on a defined
target variable. In the third phase, supervised classifiers
such as Bayesian Classifier, Artificial Neural Networks,
Support Vector Machines, and KNN are employed to
predict supplier efficiency, optimizing investment and
costing processes. Finally, the fourth phase integrates
preprocessing and prediction into a practical form, using
libraries such as Plotly and Dash for detailed visualizations,
and tools like GitHub and Heroku for application
development. This study highlights the importance of
artificial intelligence in business decision-making,
demonstrating how data science techniques and
visualization tools can facilitate the interpretation and
utilization of data analysis results.
Index Terms — Data Mining, Dash Plotly, Machine Learning,
Neural Networks, Supplier Classification.
Resumen— Este trabajo de investigación presenta el
desarrollo de un modelo utilizando técnicas de minería de
datos para identificar variables financieras en una empresa
manufacturera de carrocerías para vehículos automotores
en Pereira. El estudio se estructura en cuatro fases claves.
La primera fase se centra en el preprocesamiento de datos,
incluyendo la caracterización, normalización, y reducción
de dimensionalidad mediante PCA, Relief y Correlación. La
segunda fase aplica aprendizaje no supervisado con K-
This article was submitted for review on October 10, 2024, accepted on March
26, and published on March 31, 2025. This work was supported by CIAF
University under the project entitled “Prediction Model of Financial Suppliers in
the Vehicle Manufacturing Industry in Pereira.” The study was conducted by
Diana Carolina Romero Cárdenas and Alejandro Ospina Mejía in collaboration
with Master's student Luis Ariosto Serna Cardona. For correspondence:
karol272@gmail.com, alejandro880411@gmail.com, moog22002@hotmail.com
means y Mezclas Gaussianas (GMM) para agrupar y
validar datos según una variable objetivo definida.
En la tercera fase, se emplean clasificadores supervisados como
el Clasificador Bayesiano, Redes Neuronales Artificiales,
Máquinas de Soporte Vectorial y KNN para predecir la
eficiencia de los proveedores, optimizando los procesos de
inversión y costeo. Finalmente, la cuarta fase integra el
preprocesamiento y la predicción en un formulario práctico,
utilizando librerías como Plotly y Dash para visualizaciones
detalladas, y herramientas como GitHub y Heroku para el
desarrollo de la aplicación. Este estudio destaca la importancia
de la inteligencia artificial en la toma de decisiones
empresariales, demostrando cómo las técnicas de ciencia de
datos y las herramientas de visualización pueden facilitar la
interpretación y el aprovechamiento de los resultados del
análisis de datos.
Palabras claves— Clasificación de proveedores, Dash Plotly, Machine
Learning, Minería de datos, Redes Neuronales.
I.I. INTRODUCTION
n today's digital era, organizations face a constant challenge: it
is no longer enough to store large volumes of data—they must also
be able to manage it intelligently and convert it into actionable
knowledge. This capability has become essential to maintaining
competitiveness and improving operational efficiency in an
increasingly demanding business environment [1]. In this context,
data science has emerged as a key discipline, offering advanced
tools and techniques to analyze large-scale information, identify
patterns, predict behaviors, optimize processes, and discover new
strategic opportunities [2].
This study stems from a specific need identified in a vehicle body
manufacturing company located in Pereira, Colombia. Despite its
growth in the national market and its International projection, the
company lacks advanced mechanisms to objectively and
automatically evaluate the efficiency of its financial suppliers.
Although it uses platforms such as Siesa, Power BI, and proprietary
tools, its current analysis heavily depends on Excel macros and
Prediction Model of Financial Suppliers in the Vehicle
Manufacturing Industry in Pereira.
Modelo de predicción de proveedores financieros en la industria manufacturera de
vehículos en Pereira
investigación en biomateriales
D. C. Romero –Cárdenas ;A. O. Mejia ; L. A. Serna Cardona
DOI: https://doi.org/10.22517/23447214.25692
Scientific and technological research paper
Scientia et Technica Año XXVIII, Vol. 30, No. 01, enero-marzo de 2025. Universidad Tecnológica de Pereira.
27
manual processes, limiting data integration, reducing analytical
quality, and delaying strategic decision-making [3].
To address this issue, we propose the design of a predictive model
based on data mining and machine learning techniques, capable
of classifying suppliers into two categories: type 1 (efficient) and
type 2 (less efficient). This model is built upon key information
such as product identifiers, unit of measure, inventory type, batch
type (domestic or imported), product line and subline, supplier
and buyer codes, purchased quantities, unit values, local tax rates,
and country of origin in the case of imports. Proper identification
and analysis of these variables is essential to optimizing
investments, reducing operational costs, and strengthening
procurement management [4].
This leads to the following research question: Is it possible to
design a predictive model, based on data mining and machine
learning techniques, that can classify the financial efficiency of
suppliers in a manufacturing company? To address this question,
a four-phase methodology is proposed: (i) data preprocessing and
dimensionality reduction using PCA, correlation analysis, and
Relief-F [2][3][4]; (ii) clustering using unsupervised algorithms
such as K-Means [5] and Gaussian Mixture Models
(GMM)[6][12]; (iii) classification using supervised algorithms
such as K-Nearest Neighbors (KNN), Support Vector Machines
(SVM), Artificial Neural Networks (ANN), and Bayesian
Classifiers[13][14][15]; and (iv) development of an interactive
application, designed with Dash and Plotly, that enables real-time
visualization of results to support decision-making [16].
The proposed approach is supported by existing scientific
literature. Studies such as those by Ralambondrainy (1995) [5]
and Müller & Guido (2016) [13] have demonstrated the
effectiveness of hybrid models—combining clustering and
classification techniques—for segmenting economic actors and
predicting performance. More recently, Zhang et al. (2021) [17]
applied hybrid models for supplier selection in the manufacturing
industry, while Kim and Lee (2022)[18] used clustering
algorithms to classify suppliers in complex industrial
environments. However, these models have been developed and
validated in contexts that differ significantly from the Colombian
setting, presenting a valuable opportunity to adapt them to
emerging economies such as those in Latin America.
This paper is structured as follows: Section II presents the
methodology; Section III reports the experimental results;
Section IV discusses the findings in light of previous research;
and Section V concludes the study and outlines future research
directions.
II. II. METHODOLOGY
1. Data Preprocessing
The first phase focused on how to preprocess 350,732 sample
data points taken between 2018 and 2023, segmented into 23
columns, which has become a problem for the vehicle
manufacturing company when optimizing the investment and
costing process. Therefore, this section indicated the ideal way to
classify the information, then preprocess it and normalize it to 20
characteristic variables.
This section focused on the recognition of categorical data,
identifying variables using conventional algorithms from
different methods [7].
a. Collection and Normalization
As part of the preprocessing, it was necessary to normalize the
database before applying the prediction algorithms. This was
done to organize the data into logical groups, placing all variables
on the same scale, thus allowing fair comparisons and minimizing
data dispersion [2].
For this research and based on the state of the art supporting the
current document, the normalization method used was
StandardScaler, which generates normalized data to be processed
with unsupervised and supervised learning algorithms.[8]
TABLE I.
SAMPLE OF DATA USED IN THE EXPERIMENT
Documento
Ítem
Razón
Social
Docto.
orden
Docto.
solicitud
Desc.
ítem
15921
9186
403
69
21763
1223
15922
9332
403
69
21763
4197
15926
5330
89
182
21763
2802
b. Dimensionality Reduction
Once the database was obtained, the first step was to perform
categorical encoding of the information by classifying the data
into nominal, ordinal, and numerical variables [9]. During data
preprocessing, variable selection was conducted with the goal
of identifying or determining which methods or techniques to
apply, as conventional machine learning algorithms do not
handle categorical variables efficiently. Therefore, it is
necessary to convert them to quantitative data [10].
2. Unsupervised Learning
A clustering approach was developed using unsupervised
learning with K-Means and Gaussian Mixture Models, in which
performance metrics were used to validate the algorithm's
accuracy level according to the target variable recommended by
the expert [11].
Scientia et Technica Año XXVIII, Vol. 30, No. 01, enero-marzo de 2025. Universidad Tecnológica de Pereira
28
Based on the state of the art reviewed in this research, the K-
Means algorithm was used with ordinal variables because using
nominal variables would considerably increase computational
cost. Therefore, ordinal variables were used to divide the
dataset into K distinct clusters, aiming to obtain two groups:
efficient and less efficient suppliers [5].
A. K-Means Algorithm
The following details how the K-Means algorithm operates:
Initialization:
Select k initial points u1, u2, ..., uk called centroids.
Cluster Assignment:
For each data point xi, assign it to the cluster Cj whose centroid
uj is closest.
=
{
: ||
||^2
||
||^2
,
,
1
}
(1)
Centroid Update:
Recalculate the centroid of each cluster as the mean of the
points assigned to the cluster.
=
1
|
|
(2)
Convergence Criterion:
Continue iterating between steps 2 and 3 until the centroids do
not change significantly, i.e., until:
|
|
|
|
<
(3)
Where t is the iteration number and ϵ\epsilonϵ is a small positive
value defining the tolerance for convergence.
a. B. Gaussian Mixture Model (GMM)
After generating the labels with K-Means using the 20 features
and in agreement with the expert, it was decided to use the
labels generated by this model as the Gold standard to verify
the performance of another unsupervised algorithm, the
Gaussian Mixture Model (GMM). GMMs are a probabilistic
approach to represent the presence of subpopulations within an
unlabeled dataset. Unlike the K-Means algorithm, which
assigns each data point to a single cluster, GMMs consider that
data points can belong to multiple clusters with different
probabilities [12].
The following details how the Gaussian Mixture Model
operates:
A Gaussian mixture is defined by:
1. K: The number of Gaussian components.
2. : The weight of the k-th Gaussian component where
=
1
(4)
3. u
k
The mean vector of the k-th component.
4.
: The covariance matrix of the k-th component.
The Gaussian mixture model is defined as:
(
)
=
(
|
µ
,
)
(5)
Where
|
,
is the probability density function of the
multivariate normal distribution.
3. Supervised Learning
In this phase, supervised learning algorithms were applied
through classification tasks to validate their performance. This
research utilized the Bayesian Classifier, the K-Nearest
Neighbors (KNN) algorithm, Artificial Neural Networks (ANN),
and Support Vector Machines (SVM). It should be noted that for
supervised algorithms, classification is a very important approach
for recognizing categorical data. However, there are methods for
this purpose that focus on Kernels, which have accuracy
problems and high computational costs [13]. For this reason, an
approach to identifying ordinal categorical variables was
proposed using conventional classifiers such as Bayesian
Classification, ANN, KNN, and SVM, which not only improved
accuracy levels but also offered lower computational costs.
Subsequently, the performance of the different proposed methods
was evaluated, identifying that the three standard classifiers
significantly improved the accuracy of the database without any
procedure, making this applicability an ideal way to group and
classify categorical data [14].
The following describes the supervised KNN algorithm, the most
accurate one applied in this research.
Scientia et Technica Año XXVIII, Vol. 30, No. 01, enero-marzo de 2025. Universidad Tecnológica de Pereira.
29
The K-Nearest Neighbors algorithm is a supervised classification
method that assigns a label to a new instance based on the labels
of its k nearest neighbors in the feature space [15]. The learning
process of the KNN method is based on calculating the distance
of the new element to each existing one and sorting these
distances from smallest to largest to select the group to which it
belongs. This group is the one with the highest frequency with the
smallest distances [16].
The method operates as follows:
1. The training data X = x1, x2, … xN with labels y = y1; y2,
… yN (N being the number of data samples) are stored in
memory.
2. For a new sample xi e RD, where D is the number of
attributes, the k nearest neighbors are found using a distance
d in the entire training set (k can be 1, 3, 5, 7,…).
3. A procedure is carried out to select the class of the new
sample xi.The common distances d are:
4. The common distances d are:
• Mahalanobis Distance:
(,)
=
(
)
{}()
• Euclidean Distance:
|
|
|
=
(
)
(
)
• Manhattan Distance:
(
)
=
|
(
)
( )
|
(6)
III. III. DATASET AND EXPERIMENTAL SETUP
A. Dataset
The dataset used in this study was obtained from a vehicle
manufacturing company in the city of Pereira. Table II presents
the relevant information. It is important to clarify that the
information was anonymized to maintain the confidentiality of
the processed data. The data used in this project consists of
350,732 sample records collected between the years 2018 and
2023.
TABLE II.
SAMPLE OF THE ORIGINAL DATA USED FOR THIS EXPERIMENT
Tipo
docto
Documento
Docto.
referencia
Item
Razón Social
Sucursal
EA
EA-00187457
FACT-
120731
9186
ALEJANDRO
S.A.S
1
EA
EA-00000949
FACT-21
71630
PRIETO
LUIS
1
EA
EA-00187463
FACT-
41941
5330
CAROLINA
COLOMBIA
1
A comma-separated flat file containing this information was used
for subsequent processing in Python using the Spyder
development environment. The data underwent preprocessing,
dimensionality reduction, and feature extraction relevant to the
experiment. Following this, unsupervised and supervised
algorithms were applied, measuring their levels of accuracy and
precision to determine the most accurate algorithm. This allows
the company to predict the types of suppliers it has, classified as
Rating 1 (Efficient Suppliers) and Rating 2 (Less Efficient
Suppliers).
A. B. Specifications and Training
The identification and selection of the most relevant data sources
on supplier purchases in the financial area of a manufacturing
company in Pereira were carried out using the ERP system Siesa.
The information was exported to a CSV file for analysis. The
Extraction, Transformation, and Loading (ETL) process ensured
the proper formatting of the data, including data cleaning.
Handling of missing and duplicate values and correction of
descriptive errors.
A supplier classification module was developed, validating state-
of-the-art methods such as k-means and Gaussian mixtures to
generate grouping labels, which were evaluated using
performance metrics like accuracy, precision, recall, and F1-
Score. Additionally, a user interface with a form was designed to
enter relevant data, demonstrating the applicability of the results
through the use of the supervised KNN algorithm and Dash
Plotly, with 70% of the data for training and 30% for validation.
The combination of these methods and tools facilitated effective
classification and clear visualization of the data, providing
valuable information for financial decision-making.
Future work will focus on advanced predictive analysis and
machine learning techniques to optimize the identification of
critical financial variables and improve operational efficiency.
Deep neural networks, real-time data processing, advanced
interactive dashboards, and the expansion of the model to other
business areas will be explored. A feedback loop system is also
proposed for continuous improvement, explainable artificial
intelligence techniques to ensure transparency, and fostering
interdisciplinary collaborations, thus enhancing financial
Scientia et Technica Año XXVIII, Vol. 30, No. 01, enero-marzo de 2025. Universidad Tecnológica de Pereira
30
management and establishing a solid foundation for the
application of AI and data analysis in the organization.
IV. RESULTS AND DISCUSSIONS
A. Dimensionality Reduction Algorithms
a. PCA Algorithm
The covariance matrix was used, utilizing the Eigenvalues and
Eigenvectors of the function, reducing the dimensionality of the
variables by activating the PCA algorithm, as shown in Fig. 1.
Once the process was completed, the variance method was used,
which allowed determining which variables provided the most
information to the algorithm.
Fig. 1. Explained and Cumulative Variance
b. Correlation
As part of the dimensionality reduction process and with the
objective of identifying the linear relationship and proportionality
between statistical variables, correlation analysis was performed
on the information obtained after applying principal component
analysis in the previous step [3]. In Fig. 2, the 18 features are
observed with respect to each other, where, for example, the total
correlation between column 15 and column 1 can be seen, as well
as between column 13 and column 3. The aim is to reduce
dimensionality by removing the correlation that exists between
the features.
Fig. 2 Matrix with Full Correlation
Fig. 3 Reduced Correlation Matrix with PCA
In this matrix, by eliminating 5 features, it becomes evident that
the tone is lighter, indicating less correlation between the
columns.
c. Relief F Algorithm
For the development of the Relief algorithm, the most
recommended approach in the state of the art was used. In this
section, the Relief algorithm estimated the relevance of the
features based on how neighboring instances of different classes
differ, setting the weight of each feature to 0, defining the number
of nearest neighbors K and the number of iterations m. After
executing all iterations, the features were selected based on their
highest weights as the most relevant features [17].
Scientia et Technica Año XXVIII, Vol. 30, No. 01, enero-marzo de 2025. Universidad Tecnológica de Pereira.
31
In Fig. 4, the Eigenvalues are shown to determine which values
are below the required threshold according to state-of-the-art
recommendations. In this context, it is determined that the values
below the threshold or negative values in the Eigenvalues are
features that do not contribute value to the function.
Fig. 4. Labels with Eigenvalues
After preprocessing the threshold and removing the less
important features, Fig. 5 visualizes the most influential features
in the database.
Fig. 5. Threshold with New Eigenvalues
For data preprocessing and dimensionality reduction, three
techniques were used: PCA, Correlation, and Relief F, as
recommended by the company expert. The results showed a
reduction to 12 features with PCA, 13 features with Correlation,
and 9 features with Relief F. Consequently, it was decided not to
use any of the three techniques for dimensionality reduction and
to use the complete database with all 20 features, as
recommended by the expert.
B. Unsupervised Algorithms
a. K-Means Algorithm vs. Gaussian Mixture
Model (GMM)
Based on the state of the art reviewed in this research, the K-
Means algorithm was used with ordinal variables because using
nominal variables would considerably increase computational
costs. Therefore, ordinal variables were used to divide the dataset
into K distinct clusters, aiming to obtain two groups: efficient
suppliers and less efficient suppliers [5].
The following presents the prediction data obtained by applying
GMM to the test dataset with 20 features using the labels
generated by K-Means. It is important to highlight that the labels
generated by the GMM algorithm were duly reviewed by the
company.
TABLE III.
PERFORMANCE OF THE GMM ALGORITHM WITH 20 FEATURES
Métricas
Valores
Matriz de Confusión
TP (Prov. Clase 0)
FN (Prov. Clase 0)
FP (Prov. Clase 1)
TN (Prov. Clase 1)
43.110
0
10.582
297.041
Exactitud
97%
Clase 0
Precisión
Recall
F1-Score
0.80
1.00
0.89
Clase 1
Precisión
Recall
F1-Score
1.00
0.97
0.98
Based on this, an accuracy of 97% was identified, indicating that
it is a high-performing model. Subsequently, performance
metrics were analyzed by class, revealing that class 0 had a
precision of 80% and a recall of 100%, while class 1 achieved
a precision of 100% and a recall of 97%. This led to the
conclusion that both classes demonstrated satisfactory
performance, with an F1-score of 89% for class 0 and 98% for
class 1.
Scientia et Technica Año XXVIII, Vol. 30, No. 01, enero-marzo de 2025. Universidad Tecnológica de Pereira
32
In test 2, the prediction data obtained by applying the GMM
algorithm to the test dataset with the 13 features reduced by PCA
using the labels generated by K-Means are presented. It is
important to highlight that the labels generated by the GMM
algorithm were duly reviewed by the company.
TABLE IV.
PERFORMANCE OF THE GMM ALGORITHM WITH 13 FEATURES
USING PCA
Métricas
Valores
Matriz de Confusión
TP (Prov. Clase 0)
FN (Prov. Clase 0)
FP (Prov. Clase 1)
TN (Prov. Clase 1)
43.125
0
14.303
293.305
Exactitud
96%
Clase 0
Precisión
Recall
F1-Score
0.75
1.00
0.86
Clase 1
Precisión
Recall
F1-Score
1.00
0.95
0.98
Based on this, an accuracy of 96% was identified, indicating that
it is a high-performing model. Subsequently, performance
metrics were analyzed by class, revealing that class 0 had a
precision of 75% and a recall of 100%, while class 1 achieved a
precision of 100% and a recall of 95%. This led to the
conclusion that both classes demonstrated satisfactory
performance, with an F1-score of 86% for class 0 and 98% for
class 1.
In test 3, the prediction data obtained by applying the GMM
algorithm to the test dataset with the 9 features reduced by Relief
F using the labels generated by K-Means are presented. It is
important to highlight that the labels generated by the GMM
algorithm were duly reviewed by the company.
The following presents the prediction data obtained by applying
the GMM algorithm to the test dataset with the 9 features reduced
by Relief F using the labels generated by K-Means. It is important
to highlight that the labels generated by the GMM algorithm were
duly reviewed by the company.
TABLE V.
PERFORMANCE OF THE GMM ALGORITHM WITH 9 FEATURES
USING RELIEF F
Métricas
Valores
Matriz de Confusión
TP (Prov. Clase 0)
FN (Prov. Clase 0)
FP (Prov. Clase 1)
TN (Prov. Clase 1)
307.181
441
33
43.078
Exactitud
100%
Clase 0
Precisión
Recall
F1-Score
1.00
1.00
1.00
Clase 1
Precisión
Recall
F1-Score
0.99
1.00
0.99
Within this framework, it was observed that the experiment of K-
Means with Gaussian Mixtures, using Relief F, also achieved an
accuracy of 100%. However, considering all the features, the
algorithm achieved a high accuracy rate of 97%. In light of these
results, the expert ultimately decided to retain all the labels
generated by K-Means, recognizing their relevance in the context
of the intended financial analysis.
C. Supervised Algorithms
Subsequently, the performance of SVM, ANN, Bayesian
Classifier, and KNN was evaluated, where it was identified that
all four standard classifiers significantly improved the accuracy
of the database without any preprocessing, making this approach
ideal for grouping and classifying categorical data [14].
a. SVM Algorithm
Below are the prediction results obtained by applying each
supervised algorithm.
Scientia et Technica Año XXVIII, Vol. 30, No. 01, enero-marzo de 2025. Universidad Tecnológica de Pereira.
33
TABLE VI.
PERFORMANCE OF THE SVM CLASSIFICATION MODEL
Métricas
Valores
Matriz de Confusión
TP (Prov. Clase 0)
FN (Prov. Clase 0)
FP (Prov. Clase 1)
TN (Prov. Clase 1)
8.834
5.301
26.699
29.313
Exactitud
54%
Clase 0
Precisión
Recall
F1-Score
0.25
0.62
0.36
Clase 1
Precisión
Recall
F1-Score
0.85
0.52
0.65
The confusion matrix in Table VI shows that out of 14.135
instances of class 0, 8.834 instances were correctly predicted and
5.301 were misclassified. For class 1, out of 56.012 instances,
29.313 were correctly predicted and 26.699 were misclassified,
with an accuracy of 54%.
b. Bayesian Classification Algorithm
TABLE VII.
PERFORMANCE OF THE BAYESIAN CLASSIFICATION MODEL
Métricas
Valores
Matriz de Confusión
TP (Prov. Clase 0)
FN (Prov. Clase 0)
FP (Prov. Clase 1)
TN (Prov. Clase 1)
9.270
4.865
27.503
28.509
Exactitud
82%
Clase 0
Precisión
Recall
F1-Score
1.00
1.00
1.00
Clase 1
Precisión
Recall
F1-Score
0.03
1.00
0.06
The confusion matrix in Table VII shows that out of 14.135
instances of class 0, 9.270 were correctly predicted and 4.865
were misclassified. For class 1, out of 56.012 instances, 28.509
were correctly predicted and 27.503 were misclassified, with an
accuracy of 82%.
c. Algoritmo ANN
TABLE VIII.
ANN MODEL PERFORMANCE
Métricas
Valores
Matriz de Confusión
TP (Prov. Clase 0)
FN (Prov. Clase 0)
FP (Prov. Clase 1)
TN (Prov. Clase 1)
3.848
10.287
1.422
54.590
Exactitud
83%
Clase 0
Precisión
Recall
F1-Score
0.73
0.27
0.40
Clase 1
Precisión
Recall
F1-Score
0.84
0.97
0.90
The confusion matrix in Table VIII shows that out of 14.135
instances of class 0, 3.848 were correctly predicted and 1.422
were misclassified. For class 1, out of 56.012 instances, 54.590
were correctly predicted and 1.422 were misclassified, with an
accuracy of 83%.
d. KNN Algorithm
TABLE IX.
KNN MODEL PERFORMANCE
Métricas
Valores
Matriz de Confusión
TP (Prov. Clase 0)
FN (Prov. Clase 0)
FP (Prov. Clase 1)
TN (Prov. Clase 1)
8.197
5.938
3.301
52.711
Exactitud
87%
Clase 0
Precisión
Recall
F1-Score
0.71
0.58
0.64
Clase 1
Precisión
Recall
F1-Score
0.90
0.94
0.92
The confusion matrix in Table IX shows that out of 14.135
instances of class 0, 8.197 were correctly predicted, and 5.938
were misclassified. For class 1, out of 56.102 instances, 52.711
were correctly predicted and 3.301 were misclassified, achieving
87% accuracy.
Scientia et Technica Año XXVIII, Vol. 30, No. 01, enero-marzo de 2025. Universidad Tecnológica de Pereira
34
Based on this, a 87% precision was achieved, indicating that it
is a high-performance model. Subsequently, performance metrics
by class were analyzed, finding that both precision and recall
were 71% and 90% for both classes, and an F1-score of 64%
and 92% was observed for both class 0 and class 1. This
concluded that the supervised KNN algorithm is the most
accurate among those studied in this research.
The results obtained in this research demonstrate the high
effectiveness of the KNN model, achieving an accuracy of 87%
and an F1-Score of 92% for efficient suppliers. This validates the
model’s ability to accurately classify categorical data in a
business context.
These findings are consistent with studies such as that of Zhang
et al. [17], who applied hybrid models for supplier selection in
the manufacturing industry, reporting high levels of precision.
Similarly, Kim and Lee [18] showed that clustering algorithms
applied prior to supervised learning significantly improved
segmentation performance in complex industrial environments.
Unlike those studies, which were developed in advanced
organizational contexts in Asia, this research adapts the models
to a Colombian setting using data extracted from a local ERP
system and tailored to the conditions of a real manufacturing
company in Pereira. This contextual difference explains why the
KNN model proved more effective than SVM or ANN in this
case, particularly due to its lower computational cost and its
compatibility with visual tools such as Dash and Plotly [16].
In conclusion, this study offers a methodological contribution by
demonstrating that classification models like KNN, when
supported by robust feature reduction and interactive
visualization processes, can yield results comparable to
international research and remain applicable to emerging
business environments.
V. CONCLUSIONS AND FUTURE WORK
This work presented an approach for characterizing and grouping
categorical data using ordinal variables. The variable "Línea,"
considered key by the company expert, was used for
dimensionality reduction and supplier classification.
Preprocessing techniques such as Z-score scaling were applied,
improving experimental results and providing structured data for
supervised and unsupervised algorithms.
The clustering approach included K-Means to generate labels,
showing great similarity with the target variable. The labels were
compared with the Gaussian Mixture Models (GMM) algorithm,
achieving 97% accuracy with 20 features, 96% with 13 features
(PCA), and 100% with 9 features. However, the expert preferred
using the 20 features from K-Means due to their business
relevance. Finally, an optimized KNN model with various kernels
and features was constructed, successfully identifying and
classifying suppliers with high accuracy. This model is useful for
analysts and investors in financial decision-making for the
vehicle manufacturing company in Pereira, demonstrating the
effectiveness of the K-Means and GMM algorithms in improving
data separability and reducing computational times.
Future work will focus on advanced predictive analysis and
Machine Learning techniques to optimize the identification of
critical financial variables and improve operational efficiency.
Deep neural networks, real-time data processing, advanced
interactive dashboards, and the expansion of the model to other
business areas will be explored. A feedback loop system for
continuous improvement, explainable artificial intelligence
techniques to ensure transparency, and interdisciplinary
collaborations are also proposed to enhance financial
management and establish a solid foundation for the application
of AI and data analysis in the organization.
The performance of both supervised and unsupervised algorithms
in this study is inherently conditioned by the characteristics of the
dataset, which originates from a single manufacturing company
in Pereira. As such, the predictive models may not generalize
directly to other organizations without conducting similar
analyses to determine the most suitable algorithms for each
specific context. While the current approach effectively classifies
supplier types within the studied company, its applicability to
other environments requires contextual adaptation and possible
retraining of the models.
IV. VI. ACKNOWLEDGMENTS
V.
Thanks to the Master’s program in Systems and Computing
Engineering at the Universidad Tecnológica de Pereira, to the
Director of the Institutional Corporation for Administration and
Finance (CIAF), Luis Ariosto Serna, and to Ph.D. Julián David
Echeverry.
VI. VII. REFERENCES
[1] Gonzalez Disla, R. R. (2013). Big data: El cambio en los paradigmas
de la información.
ResearchGate. https://www.researchgate.net/profile/Renato-
GonzalezDisla/publication/311950584_BIG_DATA_El_Cambio_en_el
_Paradigma_de_la_Informacion/li
nks/58645e6208ae329d6203a9d5/BIG-DATA-El-Cambio-en-el-
Paradigma-de-laInformacion.pdf
[2] Sablón, B., et al. (2019). Gestión de la información y toma de
decisiones en organizaciones educativas. Revista de Ciencias Sociales,
XXV(2), 120–130. https://doi.org/10.31876/rcs.v25i2.27341
[3] Amo Cubillo, A. (n.d.). Research data management. Universidad de
Valladolid. Recuperado el 25 de agosto de 2024, de
http://uvadoc.uva.es/handle/10324/31269
[4] Chien, C.-F., Chang, Y.-J., & Wang, W.-C. (2018). AI and big data
analytics for wafer fab energy saving and chiller optimization to
empower intelligent manufacturing. IEEE Xplore.
https://ieeexplore.ieee.org/document/8374411
Scientia et Technica Año XXVIII, Vol. 30, No. 01, enero-marzo de 2025. Universidad Tecnológica de Pereira.
35
[5] Ralambondrainy, H. (1995). A conceptual version of the K-means
algorithm. Pattern Recognition Letters, 16(11), 1147–1157.
https://doi.org/10.1016/0167-8655(95)00075-R
[6] Liang, Y., Quan, D., Wang, F., Jia, X., Li, M., & Li, T. (2020).
Financial big data analysis and early warning platform: A case study.
IEEE Access, 8, 36515–36526.
https://doi.org/10.1109/ACCESS.2020.2969039
[7] Romero, A. C., Sanabria, J. S. G., & Cuervo, M. C. (n.d.). Utilidad y
funcionamiento de las bases de datos NoSQL. Facultad de Ingeniería,
21(33), 21–32. Recuperado el 25 de agosto de 2024, de
https://www.redalyc.org/articulo.oa?id=413940772003
[8] Iturria Aguinaga, A. (n.d.). Reduction of false positives in online
outlier detection over time series using ensemble learning. Recuperado
el 25 de agosto de 2024, de https://hdl.handle.net/10481/82540
[9] Hasan, B. M. S., & Abdulazeez, A. M. (2021). A review of principal
component analysis algorithm for dimensionality reduction. Journal of
Soft Computing and Data Mining, 2(1), 1–10.
https://doi.org/10.30880/jscdm.2021.02.01.003
[10] Macqueen, J. (n.d.). Some methods for classification and analysis
of multivariate observations. Recuperado el 25 de agosto de 2024, de
https://doi.org/
[11] Karatzoglou, A., Meyer, D., Wien, W., & Hornik, K. (n.d.). Journal
of Statistical Software: Support vector machines in R. Recuperado el 25
de agosto de 2024, de http://www.jstatsoft.org/
[12] Rasmussen, C. E. (2004). Gaussian processes in machine learning.
Lecture Notes in Computer Science, 3176, 63–71.
https://doi.org/10.1007/978-3-540-28650-9_4
[13] Castrillón, O. D., Sarache, W., & Ruiz-Herrera, S. (2020).
Prediction of academic performance using artificial intelligence
techniques. Formación Universitaria, 13(1), 93– 102.
https://doi.org/10.4067/S0718-50062020000100093
[14] Kira, K., & Rendell, L. A. (1992). A practical approach to feature
selection. In Proceedings of the ninth international workshop on
Machine learning (pp. 249-256). Morgan Kaufmann.
[15] Ahmad, A., & Dey, L. (2007). A method to compute distance
between two categorical values of same attribute in unsupervised
learning for categorical data set. Pattern Recognition Letters, 28(1), 110–
118. https://doi.org/10.1016/J.PATREC.2006.06.006
[16] Moya, R. (n.d.). Selección del número óptimo de clusters.
Recuperado el 25 de agosto de 2024, de https://jarroba.com/seleccion-
del-numero-optimo-clusters/
[17] Fin de Máster, T., Tutor, Q., Pérez, J., Cotutor, Á. A., & Bernabeu,
P. (n.d.). Métodos de clasificación con Python: Aplicaciones
empresariales. Universitat Politècnica de València, Escuela Politécnica
Superior de Alcoy. Recuperado el 25 de agosto de 2024, de
https://riunet.upv.es/handle/10251/195236
Diana Carolina Romero Cárdenas:
graduated in Systems Engineering in
2006 at the National Open and Distance
University. His research interests include
Data Science, Data Analysis, Machine
Learning, and Big Data.
ORCID: https://orcid.org/0009-0006-0568-874X
Alejandro Ospina Mejía: received his
degree in Systems Engineering in 2011.
His research interests include Data
Science, Data Analysis, Machine
Learning, and Big Data. ORCID:
https://orcid.org/0009-0009-6902-8910.
Luis Ariosto Serna Cardona received his
undergraduate degree in physical
engineering (2017) and his M.Sc. degree
in engineering (2021) and PsD Student.
He is director of research at CIAF
educacation superior and researcher in the
department of engineering. Research
interests: machine learning and deep learning.
ORCID: https://orcid.org/0000-0003-3985-4014