This paper delivers up-to-date bibliometric research dealing with data transformation used in deep learning. The study presents a theoretical approach that is developed by the bibliometric maps. The aim of this paper is the identification and exploration of the research gap. The study illustrates the relationships between different science areas, such as applied economics and finance, computer science, deep learning, and machine learning. Transformation in the literature is also defined as verification, integration, validation, cleaning, labeling of data. Creating a consistent picture of data and correcting erroneous data. Data transformation methods eliminate inconsistencies, add and aggregate data, and integrate data. In the literature on machine learning and deep learning, the term data transformation generally refers to the various operations performed on data before it is used to teach models. There is no breakdown of data transformation methods. Data transformation is a general term and there is no explanation of it in the context of deep learning. To fill this gap, this study uses VOSviewer software to explore queries' results in the Scopus database and to illustrate the knowledge about data transformation in deep learning. The results show that: (I) data transformation in deep learning is a process that is carried out in the first stage (data preparation) of building deep learning models; (II) decision support using deep learning is widely used in various business fields such as human resource management, market analysis, trading, insurance, risk analysis, logistics, industrial diagnostics, forecasting markets and economics, and business process automation; (III) there is a need to accurately distinguish between transformation and data augmentation methods to avoid potential problems associated with their misapplication in machine learning and deep learning; (IV) there is a need for further research on developing automated data transformation methods, especially through integration with feature selection or extraction methods, and testing their effectiveness in various business contexts. This scientific paper serves as a starting point for future studies dedicated to data transformation in deep learning.
Deep learning are machine learning methods based on deep neural networks, which are built
from multiple layers of neurons that implement nonlinear transformations. These layers map
successive levels of abstraction to form a hierarchical model [
Currently, there are few articles in the literature on data transformation in deep learning in
the area of management. They mainly concern the medical field (medical images [
Existing publications mainly deal with automatic data transformation methods in the field of
image analysis (techniques such as scaling, rotation, image rotation). Publications do not address
issues related to automatic data transformation in the area of management with regard to
decision-making [
The main conclusions of the article are as follows: (I) data transformation in deep learning is a process that is carried out in the first stage (data preparation) of building deep learning models; (II) decision support using deep learning is widely used in various business fields such as human resource management, market analysis, trading, insurance, risk analysis, logistics, industrial 0000-0003-0447-4038 (A. Kozina); © 2024 Copyright for this paper by its authors.
Use permitted under Creative Commons License Attribution 4.0 International (CC BY 4.0).
CEUR Workshop Proceedings (CEUR-WS.org) diagnostics, forecasting markets and economics, and business process automation; (III) there is a need to accurately distinguish between transformation and data augmentation methods to avoid potential problems associated with their misapplication in machine learning and deep learning; (IV) there is a need for further research on developing automated data transformation methods, especially through integration with feature selection or extraction methods, and testing their effectiveness in various business contexts. The contributions of this work are presented in the form of bibliometric maps illustrating the researched research gaps identified in the queries. An additional feature of the network diagrams of these keywords is that they indicate the location of data transformation in deep learning.
The beginning of neural networks dates back to the 1940s. At that time, a neuron model was
developed capable of recognizing only two categories of objects based on the weights set by the
operator. In the late 1950s, the first neural network called perceptron was built [
In the 1950s. Arthur Samuel made an important breakthrough by demonstrating that
computers have the ability to transcend the simple execution of direct commands. He created a
series of programs that operated on a game of checkers. They were capable of optimizing their
movements. As a result, these programs over time reached a level of play higher than their creator
himself [
A breakthrough in the field of artificial intelligence was the creation of the Dendral expert
system in 1965 at Stanford University, which automated the analysis of chemical compounds
[
The latest breakthrough took place in 2006, along with the development of Big Data and
computer technology. It continues to this day. Deep neural networks with effective methods for
their learning are being implemented. No small part in this breakthrough was the introduction of
graphics cards with thousands of cores, which significantly reduced the learning time of deep
learning models [
The unquestionable success and rapid development of deep learning suggest that in the future
it may become a tool to achieve Artificial General Intelligence (AGI) [
AGI refers to the development of software that has generalized cognitive abilities similar to
humans. The task of an AGI system, when faced with unknown tasks, is to find their solutions[
In 2015, the nonprofit organization OpenAI was founded with a mission to conduct research
and achieve general artificial intelligence, provided it would be safe and its goals would be in line
with those of humanity [
OpenAI has released its latest model called "ChatGPT," which can communicate with humans
through conversation. It is based on deep neural networks. The release took place on November
30, 2022 [
The process of building a deep neural network model consists of five steps [
When developing deep learning models, the data preparation process is very important. It
consists of the following steps:
1. Data acquisition and cleansing: determining data sources and structure. Acquired data is
saved as a dataset and cleaned of data containing errors [
The result of this step is a data model (input vector), which is used for model development. A very important data preparation process is data transformation, which results in the transformation of the analyzed dataset into the input vector of the deep learning model.
Jason Brownlee emphasizes that it is the data preparation stage, especially data
transformation, that is the most important when developing deep learning models. The reason is
that each data set is different and highly specific to the problem being solved [
The literature lacks automated data transformation methods for deep learning models used in
decision support. Manual data transformation relies on procedures to unify data, remove
inappropriate data, and often does not consider advanced transformations. Currently, data
transformations require many manual database refinement procedures [
It should also be noted that the automation of data transformation in machine and deep
learning models has reached a certain stage of advancement, however, there is still a need for
manual analysis of variable types and dimensionality reduction process, especially in
management decision-making. This task is not only time-consuming, but also subject to the risk
of incorrect implementation (the risk of error by the expert making the selection and
implementation of data transformation methods) [
Semi-automated methods that include the support of artificial intelligence, such as GPT chatbots, can be an attractive option because of their promises to reduce decision-making time. However, in practice, they come with some challenges. Using these tools requires the transfer of code and data information, which often leads to complex analysis and matching. This is a process that can consume a significant amount of time, as suggested transformations may prove to be inappropriate or inadequate for a particular problem. Analyzing the types of variables and tailoring suggestions to a given project problem are key steps that can be time-consuming and demanding. As a result, despite promising prospects, using semi-automatic methods can be a challenge, especially in terms of fast and efficient data processing in the process of building machine learning models [28].
The results of a study on the current knowledge of data transformation in deep learning are presented. A systematic literature review of the Scopus database was conducted (Table 1).
The earliest (1960) were publications on data transformation. The first publication on deep learning was published in 1980. Since 2012, the number of publications on deep learning has increased rapidly. The two keywords combined in 2011, but it is only since 2019 that publications on data transformation in deep learning have been increasing.
Deep learning is a relatively new field that has gained popularity in recent years due to
tremendous advances in computing technology and the availability of big data [
Advances in technology and the availability of big data have made machine learning, and deep
learning in particular, indispensable in many business fields. As researchers become more
knowledgeable about these fields, the number of publications and studies on the subject is
increasing [
There is an apparent increase in interest in the topics of data transformation and deep
learning. There are several factors that may explain this trend:
1. The growing importance of data: in recent years, data has become an extremely important
resource for organizations and businesses [29].
2. Advances in computing technologies: significant advances in computing technologies
have been made in recent years, such as the development of better algorithms, the availability
of powerful computing systems (e.g., GPUs) and the development of cloud infrastructure [
In the effect of the Scopus database queries and described procedure of the bibliometric review the results in the network diagram are represented in Figure 1.
The impact of the developed automatic data transformation method on decision support
should be considered in parallel with the impact of deep learning on the process. Deep neural
networks, based on multiple layers, can effectively map relationships and patterns in data that
would be difficult or impossible to see by traditional analysis methods, and often even by
traditional machine learning methods. Deep learning offers many advanced tools that support
more rational, accurate and optimal decision-making. With its ability to analyze large data sets,
make predictions, personalize, automate and solve complex problems, deep learning has wide
applications in many fields, supporting decision-making and contributing to better performance
and efficiency [
Multilayer deep neural network architectures, such as convolutional neural networks (CNNs)
and recurrent neural networks (RNNs), have gained recognition for their ability to automatically
extract features from data. This makes it possible to recognize patterns and cause-and-effect
relationships found in large and diverse data sets [
The use of deep learning in business data analysis makes it possible to identify hidden trends and customer preferences from purchasing data and consumer behavior. This allows companies to better adjust their products and marketing strategies, increasing their competitiveness in the market [35], [36].
With advanced deep learning methods, it is possible to analyze and process large amounts of
data to make more accurate predictions, identify patterns and support key decision-making. In
management, several major areas can be mentioned where deep learning has a significant impact
on decision-making:
1. Human resource management: deep learning can help analyze large amounts of employee
data, manage employee performance, understand training needs, and optimize the
recruitment process [37].
2. Market analysis and forecasting: the use of deep neural networks in the analysis of
economic, financial and market data allows the creation of more accurate forecasts,
identification of market trends and assessment of investment risks [38].
3. Commerce and recommendations: deep learning is often used in e-commerce platforms
and streaming services to generate personalized recommendations for products, movies or
music, helping to improve user experience and increase sales. In the field of marketing and
customer management, deep learning helps create personalized offers, product and service
recommendations, tailored to individual customer preferences and behaviors. Deep
learningbased models can analyze customer behavior, identify consumer characteristics, forecast
sales, target marketing strategies, purchase histories and demographics to better understand
customer needs and preferences [
The division of data transformation methods in machine learning and deep learning depends mainly on the area of the decision-making problem: 1. Management, where examples of data transformation can be: logarithmization of customer opinion data about products (sentiment analysis) [49], standardization of employee data [50] (transforming employees' skills to a standardized scale, making it easier to compare their skills against the requirements of different projects or tasks), normalization of production management data [51] (a deep learning model can analyze machine performance data in the production process, predicting potential failures, where, for a balance between different machine parameters, data normalization can be applied to scale the data to a certain range), or logistics [52] (transforming delivery route data into graphs or matrices and then using deep neural networks, such as convolutional neural networks (CNNs), to optimize delivery routes according to traffic conditions, time and other variables, where a data normalization method can be applied and normalize data on location distances or transportation costs, to a set range). Coding of categorical variables, such as age, gender, education level, so that they can be used as features in a predictive model dealing with social behavior is also an example [53]. 2. Finance, where examples of data transformation can be: logarithmizing stock prices to transform non-linear price changes into more linear trends [54], applying Label Powerset (LP) transformation in investment risk classification. For example, for three types of risk (market, credit, liquidity), where each can appear as “low,” “medium” and “high,” LP will transform the problem into a multi-class problem, where each combination of risks (e.g., “low market risk,” “medium credit risk,” “high liquidity risk”) will be treated as a separate class [55]. 3. Economics, where examples of data transformation may include: discretizing macroeconomic indicators by, for example, defining ranges to determine unemployment levels (low, medium, high) and assigning unemployment values to the corresponding ranges based on specified thresholds [56]. Demand forecasting is also exemplified by transforming temporal data, such as daily order quantities, into seasonal variables to account for seasonal demand patterns [57]. 4. Industry, where examples of data transformation can be: standardization of manufacturing process parameters such as temperature, pressure, humidity, to compare and analyze process stability [58], [59]. 5. Other areas of decision-making problems such as: Medicine, where examples of data transformation include: normalizing patients' laboratory results to account for differences in units of measurement and reference ranges [60], thresholding of radiological images to isolate areas of interest, such as tumors or other pathologies [61]. Natural language processing, where examples of data transformation include removing unnecessary Stopwords (frequently used words such as “a,” “the,” “is”) from text to improve the quality of text analysis [62]. Geographic data analysis, where examples of data transformation could be: normalizing geographic coordinates, such as longitude and latitude, to improve the comparability of spatial data [63].
Methods for dealing with missing data are called data augmentation. Data augmentation are methods of generating new data by altering the original data. This is often used to deal with missing data or to increase the number of samples in a learning dataset [64]. Data transformation involves changing or transforming the original data without generating new data. In the literature, sometimes the terms "augmentation" and "transformation" are used interchangeably, which can lead to confusion. However, many authors point out the need to divide these methods [64], [65], [66], [67], [68].
Misapplication of data augmentation and transformation methods in machine learning and
deep learning can lead to serious problems, such as model overfitting, as well as degradation of
prediction or classification quality. For example, if excessive data augmentation is applied, it can
cause the model to learn overly detailed patterns that are specific only to the training set but do
not reflect the overall structure of the problem, leading to overlearning. On the other hand, if an
inappropriate data transformation is applied, it can result in loss of information or disruption of
the data structure, which will negatively affect the model's ability to generalize and be effective
against new data. Therefore, it is important to always properly select and apply data
augmentation and transformation methods depending on the specifics of the problem, datasets
and variables used [
This study presents an exploration of interdisciplinary connections between various scientific domains using VOSviewer software for bibliometric analysis. The aim is to systematically guide readers through the process of conducting bibliometric data analysis and generating bibliometric maps as networks. The analysis encompasses author collaboration and keyword-based queries in the Scopus scientific database. By following a step-by-step approach, the paper enables novice users to navigate VOSviewer effectively. Through this study, the authors facilitate data analysis and identification of research gaps within the scientific landscape. The novelty lies in demonstrating bibliometric data analysis pertaining to distinct scientific fields using VOSviewer.
The methodology involves conducting bibliometric analysis to visualize keyword maps based on predetermined topics. The analysis encompasses 216 documents published between 2011 and 2024. Data transformation in the context of deep learning is often neglected, when in fact it is a key element in achieving success and getting good model results. This issue is often overlooked or underestimated, yet proper data transformation can significantly impact the effectiveness of deep models. Through proper data preparation, such as cleaning, normalization or transformation, one can ensure that models have access to high-quality and consistent data, which in turn will improve their generalization ability and prediction accuracy. Therefore, it is important to pay attention to the data transformation process and understand its fundamental importance for the successful implementation of deep learning.
Based on the literature review, it is evident that there is a need for distinguishing between methods of data transformation and augmentation.
Further research could focus on, for instance, utilizing transformation methods from the feature selection or extraction group in constructing an automatic data transformation method, testing their effectiveness in various business contexts, and comparing them with existing manual methods. Additionally, future studies could concentrate on developing a unified terminology and precise definitions for data transformation and augmentation, which would enhance understanding of the role of these methods and their impact on the quality of decision-support deep learning models.
I would like to thank my supervisors Professor, Eng. Marcin Hernes and Dr. Artur Rot from the University of Economics and Business for their research assistance in my dissertation.