Understanding the Rise of Automated Machine Learning: A Global Overview and Topic Analysis

Abstract

Automated Machine Learning (AutoML) has become an important area of modern artificial intelligence, enabling computers to automate the selection, training, and tuning of machine learning models and offering exciting opportunities for enhanced decision-making across various sectors. As the global adoption of machine learning technologies grows, it has been observed that also the importance of understanding the development and proliferation of AutoML research continues to grow, as highlighted by the increased number of scientific papers published each year. The present paper explores the scientific literature associated with AutoML with the aim of highlighting emerging trends, key topics, and collaborative networks that have contributed to the rise of this field. Using data from the Institute for Scientific Information (ISI) Web of Science database, we analyzed 920 papers dedicated to AutoML research, extracted based on specific keywords. A key finding is the significant annual growth rate of 87.76%, which underscores the increasing interest of the academic community in AutoML. Furthermore, we employed n-gram analysis and reviewed the most cited papers in the database, providing a comprehensive bibliometric overview of the current state of AutoML research. Additionally, topic discovery has been conducted through the use of Latent Dirichlet Allocation (LDA) and BERTopic, showcasing the interest of the researchers in this area. The analysis is completed by a review of the most cited papers, as well as discussions of the papers in the research areas associated with this AutoML. These findings offer valuable insights into the evolution of AutoML and highlight the key challenges and opportunities addressed by the academic community in this rapidly growing field.

1. Introduction

Machine learning and artificial intelligence have come to define the 21st century. The evolution of these technologies has generated a huge impact on the global economy, changing the rules in almost every industry [1]. The most impactful and resounding successes have been in the areas of education [2], communication [3], economics [4], transportation [5], and health [6] due to the intensive research efforts. The accelerated hardware advancement has played an important role in this evolution, storing and pre-processing the data used to train the network being crucial in order to be able to optimize any machine learning application [7]. Thus, the advancement of these technologies has also led to the discovery of other emerging technologies, such as AutoML.

AutoML arose from the need to streamline the machine learning process in a context where both the volume of the data and the complexity of the systems are continuously increasing, and according to He et al. [8], AutoML is a combination between automation and machine learning, being defined as a process of automatically building an ML workflow, respecting certain constrained budget limits.

One of the first AutoML systems, called Auto-sklearn, was introduced by Feurer et al. [9] in 2015 and was based on Bayesian optimization and meta-learning, which automatically selects the machine learning algorithm and its hyperparameters for a given dataset.

In 2017, Kotthoff et al. [10] introduced Auto-WEKA 2, a system with several substantial improvements, such as support for regression, expanded metrics, and parallel runs.

Zoph and Le [11] wrote another important paper, which was centered on Neural Architecture Search (NAS), that significantly advanced the field by demonstrating that reinforcement learning can be used to design neural network architectures automatically.

The main advantage of AutoML is that it can be useful for both researchers in the ML sphere and people with no experience in this field. First of all, the academic community could benefit from it because it reduces their effort in searching for optimal solutions. It can also be used in removing and correcting human errors that can occur at different moments of model building. Secondly, it facilitates the access for people from different fields to build ML models without requiring advanced knowledge of mathematics or statistics [12].

Due to the advantages listed above, AutoML has had a remarkable expansion in recent years, both on the research and on the industrial level, with many companies developing such tools. Tuggener et al. [13] presents different use cases of AutoML that have brought added value in branches such as medicine for the identification of different diseases [14], in banking for the estimation of the risks of lending for different people [15], and also in finance for making different predictions about future profits for companies or future stock prices [16]. Recent studies revealed several key advancements related to the financial field. ML models can improve trading performance, particularly in volatile markets (like the Bombay Stock Exchange), can effectively adapt to market changes and make profitable trades, and have been successful in outperforming traditional buy-and-hold strategies [17]. Various ML algorithms have been used for predicting cryptocurrency prices [18] and for detecting anomalies in cryptocurrency returns [19]. The integration of AutoML is crucial in autonomous vehicle (AV) applications, as it improves the safety, efficiency, and adaptability of these systems [20].

Thus, although it is still in its infancy, many private companies [21] (such as Google, Microsoft, and Amazon) have invested heavily in this field and have managed to develop tools that are already used in the industry, including AutoML Solution, Amazon Sage-Maker, Autopilot, etc. These platforms are cloud-based and provide scalable resources and tools with user-friendly interfaces to build, train, and deploy ML models without needing extensive infrastructure or deep technical expertise.

Another popular route among researchers and practitioners is the use of open-source frameworks, such as Auto-sklearn, TPOT, Auto-Keras, Auto-Pytorch, and H2O, which provide both flexibility and customization [22]. These applications encourage community contributions and continuous improvement.

AutoML has a wide range of applications across various industries. AutoML can be used in business for sales forecasting, customer segmentation, product recommendations, as well as supply chain optimization [23]. The healthcare systems can also benefit from these tools, as they can automate image analysis, streamline hospital operations, and reduce costs [23]. In manufacturing, AutoML is utilized to forecast equipment failures, to schedule maintenance, for quality control, and for supply chain management [24]. In construction, AutoML is implemented to optimize the use of materials and labor, to predict and prevent safety incidents, and to estimate project costs [25].

As expected, along with all these advantages that AutoML brings, there are also some negative effects that can be generated if misused. As mentioned in the article written by Sandu et al. [26], fake news is one of the biggest problems of today’s society, a problem that increased dramatically along with the development of ML. By making it possible for people with no scientific background to use such technologies, the number of people who want to create propaganda and misinforming news will increase. The areas in which deepfake has emerged the most are conspiracy theories [27,28], climate change [29,30], and politics [31,32,33]. Another question discussed in multiple articles is related to the reliability and robustness of these types of systems. There are also questions regarding the use of AutoML in some high-risk branches, such as medicine, where every mistake made can have a major impact on people’s health.

Thus, taking into account the current context, in order to understand the impact of AutoML on both the scientific community and the general public, bibliometric analysis remains imperative. Another benefit of bibliometric studies is the analysis of trends, investigating the contributions to the field and the topics of interest. Moreover, we will analyze the main authors, their collaborations, their educational institutions, as well as the most relevant publications in the field of AutoML. Last but not least, bibliometric analysis is used to identify the different research gaps and emerging trends, helping researchers focus on underexplored areas and contribute to the advancement of the area.

The main goal of this bibliometric work is to perform a detailed investigation into the articles published in the field of AutoML, considering the databases used in specialized articles, authors, affiliations, and collaborations between them, etc. Therefore, this study provides a systematic overview of the AutoML literature that helps researchers, practitioners, and policymakers quickly identify key themes, facilitating a better understanding of the current state of research.

As a result, through this paper, the answers to the following questions are aimed to be uncovered:

Q1: Taking into consideration the number of published papers and the H-index impact, which are the journals that generated the greatest influence around the AutoML domain?

Q2: Who are the most prolific authors in this field, and how can we characterize their influence on the scientific community in the field of AutoML?

Q3: What conclusions can be drawn by observing both the level of collaboration between scientists and the level of collaboration between countries, and how are these collaborations correlated with the progress of the scientific field?

Q4: Which countries and affiliations generate the highest number of published papers, and how can we characterize the annual production’s growth?

Q5: Which are the main ideas presented in the top 10 most cited papers, and how is this evidence validated further by words’ analysis and thematic maps?

Q6: Which are the topics discovered through LDA and BERTopic?

The main contributions of the current study are the following:

• a systematic overview of the existing literature on AutoML and its current applications;
• identification of leading authors, institutions, and countries that drive innovation in this area, guiding new researchers in identifying potential collaborators;
• revealing main themes and patterns in AutoML, such as growth rate of publications, shift in focus areas, and emerging topics, helping researchers in aligning their studies with current and future research areas;
• facilitating interdisciplinary collaboration by mapping the connections between AutoML and other fields, leading to innovative approaches and solutions based on the expertise of researchers from different disciplines;
• establishing a framework for future studies by synthesizing the existing knowledge and by identifying key areas for future research.

The remainder of this paper is organized as follows. In the first section, the general details for the readers are presented, offering a broad perspective on the topic. In the second section, the methodology is described, while in the third section, the results of the bibliometric investigation are presented. Section four highlights the most valuable findings of this study. Furthermore, a focus on the limitations of this work is provided. The last section concludes this paper by underlining the key points and addressing possible directions of research.

2. Materials and Methods

As a complete bibliometric analysis should always include thorough documentation of the materials and methods used, this chapter aims to clarify essential aspects of the steps taken in the process of collecting and filtering the dataset used in this study. Moreover, this chapter seeks to answer readers’ questions regarding data collection and processing in order to obtain the results presented in the following chapters.

This section of the analysis can be divided into two phases, as shown in Figure 1: Dataset Extraction and Bibliometric Analysis. These approaches are known in the literature and have been used in similar studies such as [34,35,36].

First of all, dataset extraction refers to the data that we have collected from the analyzed papers. The database from which the papers were extracted is the Web of Science Core Collection database (also known as WoS) [37]. Various filters have been applied to this database, including relevant keywords, the document type, the publication year, and the language in which the papers were written.

Bibliometric analysis is the second phase presented in Figure 1. This is a detailed analysis of the dataset that we presented in the previous phase. Based on the results obtained from various perspectives (sources, authors, literature, and mixed analysis), we will discuss possible improvements and future research ideas for interested researchers or institutions.

In the following, we will provide a more detailed explanation of each of the 2 discussed phases.

2.1. Dataset Extraction

As mentioned above, data extraction is a defining step for any bibliometric analysis. In addition to the database that is chosen to extract the articles, the filters we use can also affect the results obtained in this study. Thus, in the following, we will discuss the reasons behind our choices regarding the selection of works and their filtering.

For this paper, we chose to extract the articles for the analysis from only one database, called the Web of Science Core Collection or WoS, which is one of the most used sources in the literature [38,39,40]. One reason why this database was chosen is the paper elaborated by Singh et al. [41], demonstrating that WoS is a much more selective database and the articles have a greater impact on the academic world compared to Scopus [42,43]. Birkle et al. [44] also emphasized its long historical record of citation data, being established in 1964, but also its extensive and balanced coverage (around 34,000 journals). Iqbal et al. [45] reported the widespread use of WoS to evaluate publication performance and develop research policies.

Additionally, the topics that are covered in this database are broader than those of the competitors (Scopus and IEEE) by indexing a wide range of scientific, technological, and social science journals and offering high-quality, curated data, which is crucial when conducting a bibliometric analysis, a point discussed by the manuscripts of Singh et al. [41], Cobo et al. [46], and Bakir et al. [47].

WoS is considered a major tool for citation tracking, allowing researchers to measure the impact of published articles [44].

In addition, this database offers advanced search features that allow for detailed bibliometric queries [41] and the possibility of importing metadata, such as author, title, keywords, abstract, affiliation, and others, into Biblioshiny, an R application which was used for data analysis [48].

Moreover, studies such as those of Liu et al. [49] and Liu et al. [50] emphasize that, in order to have the most accurate bibliometric work, access to sources must be as comprehensive as possible. Thus, for the articles under discussion, we had access to all 10 indexes offered by WoS, namely:

• Science Citation Index Expanded (SCIE)—1900–present;
• Social Sciences Citation Index (SSCI) 1975–present;
• Arts & Humanities Citation Index (A&HCI)—1975–present;
• Emerging Sources Citation Index (ESCI) 2005–present;
• Conference Proceedings Citation Index—Science (CPCI-S)—1990–present;
• Conference Proceedings Citation Index—Social Sciences and Humanities (CPCI-SSH)—1990–present;
• Book Citation Index—Science (BKCI-S)—2010–present;
• Book Citation Index—Social Sciences and Humanities (BKCI-SSH)—2010–present;
• Current Chemical Reactions (CCR-Expanded)—2010–present;
• Index Chemicus (IC)—2010–present.

As mentioned above, after collecting the data, we continue with the data filtering process. The entire filtering process is presented in

Table 1. Data selection steps.

Exploration Steps	Filters on WoS	Description	Query	Query Number	Count
1	Title/Authors’ Keywords	Contains specific keywords related to AutoML in titles	(TI = (“automated_ machine_learning”)) OR TI = (“AutoML”)	#1	953
		Contains specific keywords related to AutoML in authors keywords	(AK = (“automated_ machine_learning”)) OR AK = (“AutoML”)	#2	1147
		Contains specific keywords related to AutoML in titles or authors keywords	#1 OR #2	#3	1619
2	Language	Limit to English	(#3) AND LA = (English)	#4	1613
3	Document Type	Limit to Article	(#4) AND DT = (Article)	#5	964
4	Year Published	Exclude 2025	(#5) NOT PY = (2025)	#6	920

, including the queries made (noted from 1 to 6 by using a “#” in front of their number) and the obtained results.

The first exploratory step we take in the filtering process is to select the papers that address specific topics, more precisely automated machine learning. Thus, papers that have specific keywords (“AutoML” and “Automated Machine Learning”) are searched in the database, resulting in a dataset of 1619 papers.

The second exploratory step involves language filtering. Only papers written in English are chosen. Thus, 6 articles that are written predominantly in another language are excluded. As a result, the dataset is reduced to 1613 papers.

The next step is to select from all the remaining papers only those that are classified as “articles”. This step is crucial because WoS classifies as “article” all papers that are considered new and original work. This method of filtering is more widely debated by Donner et al. [51], where the importance of differentiating between types of papers is also emphasized, as they can influence the number of citations. As a result of this procedure, we are left with 964 studies.

The last filter we apply is based on the publication year, being excluded from our analysis the year 2025. This decision is taken because the articles published in the last year may have a very small number of citations, thus affecting the accuracy of the bibliometric analysis.

2.2. Bibliometric Analysis

As already mentioned, the tool used to analyze data is Biblioshiny 4.2.3 (web-interface application, part of the bibliometrix package), used along with the R programming language version 4.4.1 (R Foundation for Statistical Computing, Vienna, Austria). Being an intuitive and easy-to-use tool, it has been used in many analyses by researchers [52,53,54]. It has been observed that its use in the literature has increased, as it is very useful in observing trends and possible developments in the researched field.

Figure 2 presents the bibliometric analysis facets. The first facet we are going to analyze is the dataset overview. In this stage, details about citations, sources, authors, scientific evolution, and annual scientific production evolution will be discussed. The next section analyzes the impact that the works have on the scientific world, analyzing different specific indicators such as the H-index. Another section focuses on the authors, their productivity over time, the collaboration between them, and their affiliations. The 4th fact makes an analysis of the most important works in the literature from the perspective of the number of citations. An analysis will be performed on each of these works, as well as an analysis on the keywords and scientific production based on country bigrams and trigrams. The last facet has to analyze three field plots, revealing correlations between various categories of data, offering a complex perspective on the analyzed field.

2.3. Topic Analysis

Topic analysis was conducted with the help of Latent Dirichlet Allocation (LDA) [55] and BERTopic [56].

LDA is an unsupervised topic modeling approach that can discover latent (i.e., hidden) topics by analyzing word co-occurrence patterns in a text corpus. In this study, LDA was implemented with the help of the Gensim Python 4.3.2 library and has considered for each paper the title, abstract, author keywords, and keywords plus. To improve the performance of the LDA model, a pre-processing step has been first performed. During this step, the text in the titles and abstracts has been converted to Unicode, and tokens that do not contribute to the analysis, such as URLs and citations, have been removed. The resulting text has been converted to lowercase. The text has been further pre-processed by normalizing semantically equivalent expressions such as “automated machine learning”, “auto ml”, “auto-ml” and “automl”, which would otherwise be treated as distinct entities by LDA. In the following, terms that are frequently found in research papers, such as “introduction”, “methods”, or “conclusions”, are removed. Most punctuation is removed as well, with the notable exception of hyphens, so tokens like “COVID-19” do not fuse. Finally, duplicated, leading, and trailing spaces are also removed. A simplified pre-processing approach has been used for the keywords, since in most cases they are in a form that is adequate for LDA. Thus, keywords are converted to Unicode, the text is converted to lowercase, and the semantically equivalent expressions are normalized. In the following step, bigrams and trigrams are constructed using the Phrases component of the Gensim library, which has been customized with academic-specific connector words, such as “based” or “driven”, in addition to the list of English connecter words that are commonly included in Gensim. The keywords that consist of multiple tokens are converted into n-grams. In the following step, the tokens in the title, abstract, and keywords are reduced to their canonical form though lemmatization, which converts inflected or derived words to the same canonical form. For lemmatization, the spaCy [57] Python 4.3.2 library has been used with the “en_core_sci_lg” pipeline, which provides a large vocabulary and 600 k word vectors. In the next step, for each paper, the title, abstract, author keywords, and keywords plus are combined to form the text that will be analyzed using LDA. Since a paper’s title often captures the essence of the paper, it has been repeated three times in the combined text to give the words in the title a bigger emphasis. In order to determine the best hyperparameters for LDA for the analyzed text corpus, a grid search approach has been employed, varying the number of clusters (between 2 and 5) as well as the alpha (0.01, 0.1, 1.0, “symmetric”, “asymmetric”) and eta (0.01, 0.1, 1.0, “symmetric”, “auto”) hyperparameters. A good trade-off between a small number of clusters and a good coherence of the topics has been determined when setting the number of clusters to 2, alpha to 1, and eta to 1.

BERTopic is a topic modeling technique that leverages transformer-based embeddings and clustering algorithms to generate coherent topics from text corpora. The analysis has been performed on the titles and abstracts of the papers. In comparison with the pre-processing approach used for LDA, a much simpler approach has been chosen for BERTopic that only includes normalizing semantically equivalent expressions and removing the terms that are frequently found in research papers. N-grams between 1 and 3 have been included. The parameters of the HDBSCAN algorithm used by BERTopic have been adjusted iteratively, with a final minimum cluster size of 7, while the minimum samples parameter was kept at the default value.

3. Results

This section includes the bibliometric analysis based on the previously extracted dataset regarding Automated Machine Learning. Using the tables, graphs, and illustrations generated by the Biblioshiny 4.2.3. library, which is based on the R programming language, the analysis will include five facets previously mentioned in the methodological section: dataset overview, sources, authors, papers, and mixed analysis. Multiple aspects and dimensions will be added in the forthcoming sections to create a comprehensive analysis.

3.1. Dataset Overview

An overview of the data, including information such as timespan, sources, documents, average citations, and references, is included in

Table 2. Main information about the data.

Indicator	Value
Timespan	2006:2024
Sources	517
Documents	920
Average years from publication	2.56
Average citations per document	13.35
Average citations per year per document	2.976
References	37,214

. The dataset comprises 920 documents, published within an eighteen-year timespan, from 2006 to 2024.

The relatively high number of articles published on AutoML shows increasing interest in this topic, especially in recent years, as shown by the average number of years from publication, which is 2.56.

The average citations per document and the yearly average citations per document indicate the importance and impact of this field in the scientific community. This fact is reinforced by the substantial number of references: 37,214.

Figure 3 illustrates the upward trend of published papers in the field of automated machine learning. The value for the annual growth rate is offered by Biblioshiny. The value can also be calculated using the following formula [38,58]:

$$Annual Growth Rate={\left({\displaystyle \frac{Final Year Production}{Initial Year Production}}\right)}^{{\displaystyle \frac{1}{Time Span}}}-1$$

In the Biblioshiny library, the time span refers to the period between the initial and the final years for the latest period in which all years have at least one publication [28,45]. The initial year production and the final year production refer to the production of publications for the initial and final year of the time span determined previously [28,45]. For the dataset included in this research, the annual growth rate is calculated using the following formula:

$$\begin{array}{cc}\hfill Annual Growth Rate& ={\left({\displaystyle \frac{2024 Year Production}{2016 Year Production}}\right)}^{{\displaystyle \frac{1}{Time Span between 2016 and 2024}}}-1\hfill \\ & = {\left({\displaystyle \frac{290}{1}}\right)}^{{\displaystyle \frac{1}{9}}}-1=1.8776-1=0.8776=87.76\%\hfill \end{array}$$

Article production began in 2006 with a single publication, but between 2007 and 2015, there were no recorded publications. In 2016 and 2017, only one article was released each year. The topic of AutoML became more popular in 2018, when five articles were published. From then on, the number of publications gradually increased, reaching 290 articles published in 2024.

Regarding the increase observed since 2018 onward, it should be stated that it is expected to be mostly related to the advancements made in the field of automated machine learning, concretized by a more common use of the term in the scientific literature, introduction of the cloud-based solutions for automating neural network design in industry, as well as the use of AutoML in hybrid approaches through its expansion for supporting the multimodal, tabular, text, or image data.

On another note, it should also be stated that a reduced part of the increase observed in the number of papers included in the dataset might be a consequence of the general trend observed in the last decades in science, characterized by the increase in the number of scientists [59,60], as well as the advancements in computer science [59], or the growth in the research and development expenditure [60].

Figure 4 showcases the fluctuating pattern of the average citations per year, with values ranging from 1.11 registered in 2017 to 14.22 in 2019. This variation indicates that the numerous papers from this dataset have different levels of impact in the scientific community. The lower number of citations gathered in the last years might be due to the increase in the number of papers published, as well as the reduced number of years since the papers’ publications, which could not have ensured a minimum period of time in which the papers could be read by the audience, could be cited, and the works in which the papers to be cited could be indexed in WoS.

Table 3. Document contents.

Indicator	Value
Keywords plus	1493
Authors’ keywords	2661
Authors	3964
Author appearances	4894
Authors of single-authored documents	19
Authors of multi-authored documents	3945
Single-authored documents	19
Documents per author	0.232
Authors per document	4.31
Co-authors per documents	5.32
Collaboration index	4.38

presents information on document contents, such as keywords plus (known as index terms and determined by the database through the extraction of the most common words in the titles of the papers included in the bibliography of the papers from the dataset [61]) and the authors’ keywords.

The number of keywords plus, specifically 1493, was significantly smaller than the authors’ keywords, with a value of 2661. By dividing the number of keywords plus by the number of documents, the average number of 1.62 such terms per document is obtained, demonstrating a concentrated and compact vocabulary in the articles included in the dataset analysis. Consequently, the average number of authors’ keywords per document, more specifically 2.89, is calculated using the same formula of dividing the authors’ keywords by the total number of documents.

The contrast between the number of authors and the published articles, with more than four times as many authors as papers, shows that the academics from this field have a high degree of collaboration.

Also, the discrepancy between the number of unique authors (3964) and the author appearances (4894) indicates that some academics have contributed to multiple papers.

The high level of collaboration is also proven by the number of authors of multi-authored documents (3945). Only 19 documents from this dataset were classified as single-authored documents.

The substantial degree of collaboration between researchers is further reinforced by the indicators included in this table.

The relatively small value of 0.232 documents per author, calculated by dividing the number of articles included in the dataset (920) by the number of authors (3964), along with the values registered for authors per document (4.31) and co-authors per document (5.32), and a collaboration index of 4.38, all contribute to the theory of high tendency to collaborate in this academic field.

By dividing the number of authors of single-authored documents (19) by the number of single-authored papers (19), it is confirmed that the academics who have published their papers as sole authors have done so for precisely one paper.

3.2. Sources

Figure 5 directs the attention to the top 20 most relevant journals based on the number of published articles. These journals have published at least 6 articles on the AutoML domain.

The first position is held by IEEE Access, with 28 articles, followed by Scientific Report with 26 papers, and Applied Sciences-Basel with 21 documents. Sensors has 16 publications, IEEE Transactions on Pattern Analysis and Machine Intelligence has 15 papers, while Applied Soft Computing and Remote Sensing have 13 and 12 articles.

Nine documents were published in IEEE Transactions on Neural Networks and Learning, Journal of Machine Learning Research. Energies and Machine Learning Systems each has 8 papers, while Diagnostics, Journal of Clinical Medicine, and Knowledge-Based Systems have published 7 papers. Last in this top 20 are Applied Intelligence, Engineering Applications of Artificial Intelligence, IEEE Transactions on Industrial Informatics, IEEE Transactions on Knowledge and Data Engineering, Information, and Information Sciences, with six articles.

It is worth noting that the journals that occupy the first positions are the ones that are preferred by academics to publish their research, not only because they are well acclaimed but also because they promote technological advancement.

The core sources by Bradford’s Law are showcased in Figure A1. This pattern separates the sources into three distinct zones: Zone 1 (placed in the core), which includes representative and highly cited journals, Zone 2 (placed in the center) includes journals with moderate productivity, and Zone 3 (placed externally) includes less influential journals [62].

Regarding the present analysis, the top 20 most relevant sources are also included in Zone 1, followed by 18 more journals. This finding reinforces the presumption that these journals have the highest impact and influence in the academic community, making a significant contribution to the progress of the AutoML domain.

Figure A2 presents the journals’ impact based on H-index [63]. This metric indicates the journals’ significance. In theory, the higher the value registered for this index, the greater the impact, productivity, and influence in the field’s academic community.

Scientific Reports occupies the top position, with an H-index value of 11. In this journal, at least 11 articles related to AutoML were published, each of them gathering at least 11 citations.

The second position is held by IEEE Transactions on Pattern Analysis and Machine Intelligence, with a recorded value of 9 for the H-index (at least 9 papers with at least 9 citations each). The third and fourth places are held by IEEE Access and Sensors, with H-indexes of 8 and 7, respectively. Two journals occupy the fifth and sixth places: Engineering Applications of Artificial Intelligence and IEEE Transactions on Neural Networks and Learning Systems, each with a value of 6 for the H-index. The last positions are occupied by the journals with an H-index of 5: Applied Sciences-Basel, Applied Soft Computing, Energies, IEEE Transactions on Intelligent Transportation Systems, Journal of Machine Learning Research, Machine Learning, Proceedings of the VLDB Endowment, Remote Sensing, and Sustainability.

3.3. Authors

An essential part of a bibliometric study is analyzing the most prominent authors based on the number of publications and citations.

Figure 6 explores the 14 most relevant authors based on the number of publications regarding automated machine learning. The most relevant author is Wang Y. with 12 published articles, followed by Tsamardinos I. with 11 papers. The third and fourth places are occupied by Liu L. and Yin MY., each with 10 publications, while Zhang L. takes the fifth place with 9 documents. The remaining 9 authors in the top 14 are Gao JW., Hutter F., Li Y., Lin JX., Liu XL., Moore JH., Wang J., Zhang Y., and Zhu JZ., each having 8 papers published.

This graph demonstrates the substantial participation of these 14 authors in the AutoML domain, their indisputable dedication to studying all aspects of this area, and their desire to advance technology in the examined field. These authors are regarded as major contributors in the domain, and their articles are important resources to the academic community and others.

Figure A3 illustrates the authors’ productivity based on Lotka’s Law [64]. This principle reveals a negative correlation between the number of publications and the percentage of authors: as the number of written documents increases, the percentage of authors decreases drastically: 85.9% of the authors, namely 3406, contributed to a single article, while 9.5% (377) have been involved in writing two articles, and 2.4% of the authors, precisely 97, have participated in writing three articles. The results demonstrate that the number of highly productive researchers is significantly smaller than the number of those with low productivity. This finding emphasizes the challenges authors face when writing about the AutoML domain.

Figure A4 brings to attention the top nine authors’ local impact based on H-index. The top four most cited authors, with a calculated value of 6 for the H-index, are Hutter F., with 1737 total citations, Moore JH. (331 total citations), Tsamardinos I., with 169 citations, and Wang Y. with 138 total citations. The top is continued with five more researchers, Li I., Liu ZY., Mohr F., Wever M., and Zhang Y, each with an H-index of 5 and total citations ranging from 56 to 183.

Figure 7 depicts the top 13 most relevant affiliations, based on the number of publications about automated machine learning.

By analyzing this figure, it is proven that the universities from this list made an important contribution to this field. All this was possible because of the excellent academic guidance, focused on technological advancement.

This list comprises diverse affiliations: 5 Chinese, 5 USA, 2 British, and 1 Singaporean. The Chinese Academy of Sciences occupies first place with 51 publications. The second position is held by the University of California System with 45 published articles. Sichuan University and the University of Chinese Academy of Sciences, CAS come in third and fourth place, with 23 and 22 publications. Notably, these two numbers are significantly smaller than those of the first two places. Other affiliations, based on the number of articles published, are the University of California Los Angeles, the Chinese Academy of Sciences, Soochow University–China, the University of California System, University College London, Harvard University, and the National University of Singapore. These universities have published between 15 and 20 papers.

Figure 8 reveals the top 15 most relevant corresponding authors’ countries based on the number of published articles around the AutoML field.

In this graph, Single-Country Publications (SCPs) are depicted in green while Multiple-Country Publications (MCPs) are colored in orange.

China leads the top with 246 publications, recording the highest values for Single-Country and Multiple-Country Publications: 183 and 63. The second most important country in this analysis is the USA, totaling 141 documents (SCP = 104, MCP = 37). With a significant difference, India sits in third place, having 43 published manuscripts (SCP = 36, MCP = 7). Please consult Figure 8 for the entire list.

This analysis emphasizes the topic’s global impact, the degree of collaboration between the authors, and the countries with the highest number of publications, providing a broad understanding of the international scientific community around the subject of automated machine learning. This might also assist both institutions and authorities in developing advanced strategies, pertinent decisions, and best practices in this domain [65].

The top five countries’ production over time is illustrated in Figure A5. According to the data, the USA was the first to publish papers on AutoML in 2006. The following paper was published 10 years later, in 2016, as a collaboration between the USA and Germany. The United Kingdom started to publish in 2018, China in 2019, and India in 2020. China had the highest increase in the number of publications. In second place is the USA with a similar trend. On the other hand, Germany, the United Kingdom, and India saw lower growth rates.

The top ten countries based on the number of citations are presented in Figure 9. China occupies the first position with 3307 total citations and a value of 13.4 for average article citations. In second place is the USA, totaling 2745 citations and a higher value for average article citations of 19.5. Even though Germany has a lower number of citations (2291), it recorded the highest value for average article citations: 60.3. The top is then continued with countries with lower impact: United Kingdom (348 citations), India (323 citations), Korea (303 citations), Greece (283 citations), Spain (224 citations), France (221 citations), and Brazil (203 citations).

Figure 10 illustrates the scientific production based on country, utilizing color representation as well as the collaboration between them. Countries with high productivity levels, such as China and the USA, are represented in darker shades of blue. Lighter shades of blue denote countries with medium and low productivity levels, such as Germany, India, the United Kingdom, and South Korea. Countries without published documents are highlighted in grey (e.g., Albania, Moldova, and Ukraine).

The width of the red lines showcases the collaboration intensity. This figure illustrates the diverse collaborations between countries, with China and the USA being the foremost promoters of partnerships.

Researchers from China collaborated the most with those from the following countries: the USA, the United Kingdom, Australia, Canada, Singapore, France, Germany, Pakistan, Saudi Arabia, and Norway.

The USA partnered with Canada, the United Kingdom, Germany, France, Australia, India, Mexico, Singapore, Switzerland, and Ireland to write documents related to the studied domain. For additional insights, please analyze Figure 10.

Figure 11 explores the collaboration network of the 50 most important authors, which are grouped in 11 clusters. By analyzing this graph, it is revealed that the authors’ collaboration networks are relatively small, with two exceptions: the cluster in red and the cluster in light green.

The first cluster mentioned, the one in red, includes the following authors: Hutter F., Wang J., Zhang Y., Vanschoren J., Bischl B., Li YQ., Lindauer M., Liu Y., Liu ZY., and Eggensperger K. They introduced new AutoMl approaches, which allow systems to work efficiently on large datasets. It also makes a comparison with other popular AutoML frameworks, managing to obtain significantly improved results compared to Auto-sklearn [66].

The second cluster, the one in light green, has the most authors (11) and includes Wang Y., Liu L., Yin MY., Zhang I., Gao JW., Li Y., Lin JX., Liu XL., Zhu JZ., Xu C., and Zhang RF. They have had results in medicine, settling the problem of asymptomatic COVID-19 patients [67].

The third cluster, the one in pink, features Fu Wx. and Moore Jh. They present TPOT-NN—which is a new tree-based AutoML extension. In their studies, it is demonstrated that TPOT-NN is a tool that achieves better accuracy in classification problems compared to the standard tree-based problems [68].

The fourth cluster, which is colored in purple, has also two authors, namely Felix Mohr and Marcel Wever. They analyze new optimization methods, creating Naïve AutoML, an approach that realizes an in-isolation optimization. The isolated optimization reduces search spaces, performance not being affected [69].

The fifth cluster is colored in orange and consists of the authors Moncef Garouani, Mourad Bouneffa, Adeel Ahmad, and Mohamed Hamlich, who have released a new version of self-explainable AutoML AMLBID software. This implementation makes datasets’ meta-features more informative, and it also can improve the accuracy [70].

The next cluster, colored in gray, consists of authors Wang CN., Cheri X., and Wang HZ., who propose a new method to find the best configuration of hyperparameters using an algorithm called ExperienceThinking. The final results show that the obtained results are superior to the existing ones, thus demonstrating the capability of the new algorithm [71].

The seventh blue-colored cluster includes Baratchi M. and Wang C., who make an analysis of the literature on the AutoML domain. They provide an extensive overview of the past and present, as well as future perspectives of AutoML [72].

The eighth cluster, colored in light blue, brings us back to the field of medicine. The authors (Benecke J. and Olsavszky V.) make an analysis and prediction on enterobiosis and cystic echinococcosis [73].

The yellow cluster introduces us to the banking field. The authors investigate the performance of an AutoML methodology in forecasting bank failures, called Just Add Data (JAD). They discover that JAD can increase the accuracy of the financial data analyzed [74].

The last two clusters, colored in brown and green, which are represented by the authors Rashidi HH. and Albahra S. and Czub N. and Mendyk A. The first article, led by Rashidi HH. [75], is used to determine the impact of point-of-care (POC) in burn-injured and trauma patients, while the research led by Czub N. [76] managed to develop a quantitative structure–activity relationship (QSAR) prediction model.

3.4. Analysis of the Literature

This section explores the existing literature published around the domain.

The first subsection provides an overview of the top ten most globally cited papers from the dataset. Several details of the manuscripts are presented, including paper details, first author, year of publication, journal, the number of authors who collaborated to write the paper, and region, as well as relevant indicators: total citations, total citations per year, and normalized total citations.

The second subsection is focused on presenting a review of the top 10 most cited papers included in the top. The key points of interest are included to gain a better understanding of the topics discussed.

The last subsection offers an extensive analysis of the most common terms in authors’ keywords, keywords plus, abstracts, and titles.

3.4.1. Top Ten Most Cited Papers—Overview

Table 4. Top ten most globally cited documents.

No.	Paper (First Author, Year, Journal, Reference)	Number of Authors	Region	Total Citations (TC)	Total Citations per Year (TCY)	Normalized TC (NTC)
1	Elsken T., 2019, Journal of Machine Learning Research [77]	3	Germany	1181	168.71	11.86
2	He X., 2021, Knowledge-Based Systems [8]	3	China	804	160.80	31.87
3	Feurer M., 2019, The Springer Series on Challenges in Machine Learning [78]	6	Germany	312	44.57	3.13
4	Le TT., 2020, Bioinformatics [79]	3	USA	244	40.67	10.26
5	Alaa AM., 2019, PLoS ONE [80]	5	U.K. and USA	231	33.00	2.32
6	Zöller MA., 2021, Journal of Artificial Intelligence Research [81]	2	Germany	168	33.60	6.66
7	Chen Z., 2021, Nucleic Acids Research [82]	12	UK and China	146	29.20	5.79
8	Lindauer M., 2022, Journal of Machine Learning Research [83]	9	Germany	146	36.50	12.06
9	Karmaker SK., 2021, ACM Computing Surveys (CSUR) [84]	6	USA	121	24.20	4.80
10	Xu H, 2023, Soft Computing [85]	4	Canada	109	36.33	15.03

includes the top ten most globally cited documents that address the topic of automated machine learning.

By quickly analyzing the number of authors in the papers, it can be observed that all documents are the product of collaborative work between researchers, confirming the high degree of cooperation. Two papers resulted from the association of authors from different countries (the UK with the USA, and the UK with China).

Another important finding is that the authors have diverse countries of origin: Germany, China, the USA, the UK, and Canada.

The top was generated based on the documents’ total citations (TC), total citations per year (TCY), and normalized total citations (NTC). As the TC and TCY indicators are easily comprehensible, further explanations must be made for the NTC. The NTC is calculated by dividing the average yearly number of citations of a paper by the average number of citations received by all the papers included in the dataset and published in the same year as the paper for which the NTC is calculated [54].

The first position is held by Elsken et al. [77] with 1181 total citations, 168.71 total yearly citations, and a value of 11.86 for normalized total citations. The paper produced by He et al. [8] occupies second place with 804 total citations, 160.8 total yearly citations, and 31.87 NTC. Feurer et al. [78] published the paper from third place, accumulating 312 total citations, 44.57 total citations per year, and a value of 3.13 for normalized TC.

The recorded values for total citations (from 109 to 1181) and total citations per year (from 24.2 to 168.71) confirm the papers’ relevance and increased visibility in the academic community around the AutoML domain.

3.4.2. Top Ten Most Cited Papers—Review

Table A1. Brief summary of the content of top 10 most global cited documents.

No.	Paper (First Author, Year, Journal, Reference)	Title	Data	Purpose
1	Elsken T., 2019, Journal of Machine Learning Research [77]	Neural Architecture Search: A Survey	Authors did not use data; they explained the concepts in a theoretic manner	To provide an overview of existing work
2	He X., 2021, Knowledge-Based Systems [8]	AutoML: A Survey of the State-of-the-Art	Authors used CIFAR-10 datasets and ImageNet datasets to compare the algorithms	To present a comprehensive and up-to-date review of the state of the art (SOTA) in AutoML
3	Feurer M., 2019, The Springer Series on Challenges in Machine Learning [78]	Automated Machine Learning Methods, Systems, Challenges	Authors did not use data; they explained the concepts in a theoretical manner	To illustrate overview of general methods in AutoML
4	Le TT., 2020, Bioinformatics [79]	Scaling tree-based automated machine learning to biomedical big data with a feature set selector	Authors used both simulated datasets and a real-world RNA expression dataset	To implement two new features in TPOT that help increase the system’s scalability: Feature Set Selector (FSS) and Template
5	Alaa AM., 2019, PLoS ONE [80]	Cardiovascular disease risk prediction using automated machine learning: A prospective study of 423,604 UK Biobank participants	Authors used a dataset with real records from patients	To test whether ML techniques based on an automated ML framework could improve CVD risk prediction compared to traditional approaches
6	Zöller MA., 2021, Journal of Artificial Intelligence Research [81]	Benchmark and Survey of Automated Machine Learning Frameworks 2021	Authors did not use data; they explained the concepts in a theoretical manner	To benchmark 14 most popular AutoML and hyperparameter optimization (HPO) frameworks
7	Chen Z., 2021, Nucleic Acids Research [82]	LearnPlus: a comprehensive and automated machine-learning platform for nucleic acid and protein sequence analysis, prediction and visualization	Authors use the datasets from a study by Han et al. [129]	To introduce a novel machine learning platform, iLeanPlus
8	Lindauer M., 2022, Journal of Machine Learning Research [83]	SMAC3: A Versatile Bayesian Optimization Package for Hyperparameter Optimization	Authors used the Letter Dataset for the hyperparameter optimization benchmark	To introduce SMAC3, an open-source Bayesian optimization package for hyperparameter optimization
9	Karmaker SK., 2021, ACM Computing Surveys (CSUR) [84]	AutoML to Date and Beyond: Challenges and Opportunities	Authors did not use data; they explained the concepts in a theoretical manner	To define a new classification system for AutoML
10	Xu H, 2023, Soft Computing [85]	A data-driven approach for intrusion and anomaly detection using automated machine learning for the Internet of Things	Authors used data generated by traditional networks that they improved by using Synthetic Minority Oversampling Technique (SMOTE) algorithm	To present a data-driven approach method to detect intrusion and anomaly detection

illustrates the ten most cited papers in the field of AutoML. In the following, we will review these papers, emphasizing the important discoveries that each work has made.

The first article that comes into discussion is “Neural Architecture Search: A Survey”, which addresses the automatic methods of searching for neural architectures (NAS) [86]. The authors aim to classify and compare the methods that exist in the literature, as AutoML models become more and more responsive and optimization design becomes a necessity. The methods that come into discussion include reinforcement learning-based algorithms [87] and evolutionary methods, which are inspired by genetic algorithms and gradient based modes. The results were surprising, as they discovered that NAS can lead to the discovery of more performant architecture than the current ones. However, there is a challenge at the moment, because the consumption of computational resources is quite high. For future directions, the paper aims to explore the different combinations of existing methods to achieve more robust results, as well as to develop them computationally in order to face the challenges in various fields. This article is the first in terms of citations but also in terms of citations per year, with an average of 168 citations per year.

The second article provides a comprehensive analysis of AutoML, which has become more and more studied in the literature in recent years and represents a promising solution to build a machine learning system without the need for human assistance. This paper is state of the art, introducing us to the field of AutoML with its methods (covering data preparation, hyperparameter optimization, and neural architecture search—NAS), and also summarizes the performance of NAS algorithms on two known datasets (CFAR-1o and ImageNet). Additionally, various methods for optimizing the architecture as well as the resource-aware NAS are suggested. Finally, they raise the issue of various challenges that researchers studying the AutoML field may encounter, such as very high computational costs and also low efficiency for certain practical applications. Even if in terms of the total number of citations it is weaker than the first article, with 800 citations compared to 1181 citations, the average number of citations per year is almost the same, with 160 citations annually.

The third article is quite behind the first two papers in terms of citation numbers, also having a number four times lower than the average number of annual citations. First of all, the authors talk about hyperparameter optimization (HPO) [88], initially analyzing blackbox optimization methods, followed by a discussion on Bayesian optimization. In addition, they raise the problem of high computational costs for blackbox methods, suggesting modern multi-fidelity methods for approximating the quality of hyperparameters, the costs being much lower. Once again, the authors conclude with prediction of future research directions.

The fourth paper in the list is the one written by Trang Le. In this paper, the authors introduce two new features to improve scalability in the Tree-based Pipeline Optimization Tool (TPOT) [89], namely Feature Set Selector (FSS) and Template. These two characteristics are used for large datasets, managing to reduce computation time and power. While the Template method imposes constraints using genetic programming, FSS allows subsets of features to be specified as separate datasets. Ultimately, the study shows that FSS performs better than an XGBoost [90] model, proving that research in the field significantly improves computational time and power used.

The fifth article, which is very close to the previous one regarding the number of citations (231 compared to 244), presents the use of machine learning in medicine for the prediction of cardiovascular diseases, also making a comparison with traditional detection methods. Using data from 423,604 participants, they developed a model derived from AutoPrognosis, an algorithmic tool that automatically selects and tunes ensembles of machine learning modeling pipelines. Lastly, they found that the newly created model improved the prediction of cardiovascular disease risk, from an AUC-ROC score of 0.724% to 0.774%. Thus, the potential that AutoML technologies can have in domains associated with human health is emphasized, improving the accuracy of diagnoses.

The paper by Zöller et al. [81] includes a study of current AutoML methods and a benchmark of the 14 most popular AutoML and hyperparameter optimization (HPO) frameworks on 137 real datasets from established benchmark suites. A mathematical formulation covering the complete procedure with steps for creating an ML pipeline was presented. The experiments conducted by the authors revealed that on average, all AutoML frameworks performed similarly, with a performance difference of only 2.2%, and that currently AutoML is completely focused on supervised learning. Two authors from Germany contributed to this paper, obtaining 168 total citations, with a mean value of 33.6 citations per year, while the NTC value is 6.66.

Chen et al. [82] introduced in their paper a novel machine learning platform, iLeanPlus, employing AutoML and featuring graphical and web-based interfaces. The solution includes four built-in modules and provides sequence analysis and prediction based on nucleic acid and protein sequences. The real-world performance of the solution was demonstrated in two case studies, highlighting the platform’s ability to rapidly develop accurate models that maximize its predictive performance. The research was performed by twelve authors from China, Australia, Japan, and the USA, the article reaching a total of 146 citations, with a mean yearly reference of 29.2 and an NTC value of 5.79.

Another original solution was presented by Lindauer et al. [83]: SMAC3, an open-source Bayesian optimization package for hyperparameter optimization of machine learning algorithms. SMAC3 was designed to be used in different use cases: for Low-dimensional and Continuous Blackbox Functions [91], for CASH and Structured Hyperparameter Optimization, for Expensive Tasks and Automated Deep Learning, as well as for Algorithm Configuration. SMAC3’s performance was demonstrated during a brief empirical comparison with similar tools (Dragonfly [92], BOHB [93], Hyperband [94], and random search [95]). Nine researchers from Germany worked on this paper. The document gathered 146 total citations, a value of TCY of 36.5, and an NTC of 12.06.

Karmaker et al. [84] proposed a new classification system for AutoML systems and defined seven levels based on the different levels of automation: lower levels employ less automation and more manual work, and higher levels require more automation and less manual work. Several challenges are addressed during a case study: challenges in task formulation, challenges in prediction engineering for identifying “promising” tasks, and in result summarization and recommendation. Six authors from the USA contributed to the writing of the paper, achieving 121 references, a mean citation per year of 24.2, and an NTC value of 4.8.

The paper by Xu et al. [85] is focused on intrusion and anomaly detection in the network using a data-driven approach. The model uses the Synthetic Minority Oversampling Technique (SMOTE) [96] algorithm and random undersampling to balance the dataset. The results of this study show that the proposed model predicts with an accuracy of 99.7% whether the network traffic instance is normal or if there is malicious activity going on. The model proves to be lightweight, requiring less processing time and computational resources. Four authors from China participated in the research, gaining 109 total citations, 36.33 average annual citations, and 15.03 NTC.

A summary of the content of the top ten most globally cited documents, including the datasets used and purpose, can be found in Table A1.

The analysis of the top ten most cited papers reveals that researchers in this field are concerned with both the general aspects of AutoML (by creating comprehensive surveys), with model optimization, as well as with specific applications in healthcare or security.

3.4.3. Words Analysis

In this section, a word analysis is performed to uncover the different themes, topics, and trends resulting from the dataset. Information from authors’ keywords, keywords plus, bigrams, and trigrams (in both abstracts and titles) is extracted to gain a clear understanding of the studied documents. Co-occurrence networks and thematic maps are included in the analysis to uncover more insights on the topic of automated machine learning.

Table 5. Top ten most frequent words in keywords plus and in authors’ keywords.

Words in Keywords Plus	Occurrences	Words in Authors’ Keywords	Occurrences
model	56	automated machine learning	329
classification	55	automl	297
prediction	51	machine learning	215
optimization	45	deep learning	86
algorithm	36	neural architecture search	65
selection	28	artificial intelligence	59
algorithms	25	automated machine learning (automl)	42
models	25	optimization	31
diagnosis	24	training	30
system	24	hyperparameter optimization	28

includes the top ten most frequent words in keywords plus and authors’ keywords. Regarding keywords plus, the most frequent words are “model” with 56 occurrences, “classification”—55 occurrences, “prediction”—51 occurrences, “optimization”—45 occurrences, “algorithm”—36 occurrences, “selection”—28 occurrences, “algorithms” and “models”, each with 25 occurrences, followed by “diagnosis” and “system”, each with 24 occurrences. The words reveal that the most recurrent topics in the studied papers are related to model selection and methods such as classification and optimization that achieve better predictions and medical diagnosis.

The most frequent words in authors’ keywords are: “automated machine learning” with 329 occurrences, “automl” with 297 occurrences, “machine learning”—215 occurrences, “deep learning”—86 occurrences, “neural architecture search”—65 occurrences, “artificial intelligence”—59 occurrences, “automated machine learning (automl)—42 occurrences, “optimization”—31 occurrences, “training”—30 occurrences, and “hyperparameter optimization” with 28 occurrences. As expected, the significant themes are automated machine learning, artificial intelligence, and optimization methods.

Figure 12 presents the top 50 most frequently used words based on keywords plus (A) and authors’ keywords (B). The size of the words indicates the frequency of the words included in the studied literature on automated machine learning.

In terms of keywords plus, the most prominent words are “model” with 56 occurrences, “classification” with 55 occurrences, “prediction” with 51 occurrences, and “optimization” with 45 occurrences, whereas in terms of authors’ keywords, “automated machine learning” (329 occurrences), “automl” (297 occurrences), “machine learning” (215 occurrences), and “deep learning” (86 occurrences) are the most frequently used words.

Table 6. Top ten most frequent bigrams in abstracts and titles.

Bigrams in Abstracts	Occurrences	Bigrams in Titles	Occurrences
machine learning	1315	machine learning	430
automated machine	499	automated machine	367
learning automl	288	architecture search	49
learning ml	160	neural architecture	47
deep learning	156	learning approach	46
neural network	141	deep learning	27
learning models	136	machine learning-based	26
neural networks	125	learning model	21
ml models	115	neural network	19
architecture search	110	time series	19

sheds light on the most frequent bigrams in both abstracts and titles. In the case of abstracts, the first position is occupied by “machine learning” with 1315 occurrences, followed by “automated machine” with 499 occurrences and “learning automl” with 288 occurrences. Next in the list are “learning ml” (160 occurrences), “deep learning” (156 occurrences), “neural networks” (125 occurrences), “ml models” (115 occurrences), and “architecture search” with 110 occurrences. On the right, bigrams in titles are provided, the top three being “machine learning” with 430 occurrences, “automated machine” with 367 occurrences, and “architecture search” with 49 occurrences. The list is completed with the following items: “neural architecture” (47 occurrences), “learning approach” (46 occurrences), “neural network” (19 occurrences), and “time series” with 19 occurrences.

The bigrams identified in the data collection, and showcased in this table, provide the main themes addressed in these articles but without further explanation of their connections within the paper. Therefore, only a section of the topic categorization is presented and does not ensure that the documents studied in this analysis refer to possible solutions as well as challenges connected to the domain of automated machine learning.

The top ten most frequent trigrams in abstracts and titles are presented in

Table 7. Top ten most frequent trigrams in abstracts and titles.

Trigrams in Abstracts	Occurrences	Trigrams in Titles	Occurrences
automated machine learning	496	automated machine learning	335
machine learning automl	285	machine learning approach	43
machine learning ml	157	neural architecture search	41
machine learning models	102	automated machine learning-based	23
machine learning algorithms	81	machine learning model	18
neural architecture search	76	machine learning automl	15
architecture search nas	51	machine learning models	12
machine learning methods	43	machine learning tool	10
machine learning model	42	machine learning algorithms	8
receiver operating characteristic	35	machine learning framework	8

. In the case of abstracts, “automated machine learning” occupies the first place, with 496 occurrences, followed by “machine learning automl” with 285. The list is completed with the following trigrams: “machine learning ml” (157 occurrences), “machine learning models” (102 occurrences), “machine learning algorithms (81 occurrences), “neural architecture search” (76 occurrences), “architecture search nas” (51 occurrences), “machine learning methods” (43 occurrences), “machine learning model” (42 occurrences), and “receiver operating characteristic” with 35 occurrences.

Regarding the titles, the first position is occupied by “automated machine learning” with 335 occurrences. “Machine learning approach” is second at the top, with 43 occurrences. The next trigrams are also added to the list: “neural architecture search” (42 occurrences), “automated machine learning-based” (23 occurrences), “machine learning model” (18 occurrences), “machine learning automl” (15 occurrences), “machine learning models” (12 occurrences), and “machine learning tool”, with 10 occurrences. The top is completed with “machine learning algorithms” and “machine learning framework”, each with 8 occurrences.

Figure A6 provides a visual depiction of the top authors’ keywords grouped in two clusters based on co-occurrence.

The first cluster, in red, comprises 36 keywords, including “automated machine learning”, “automl”, “machine learning”, “deep learning”, “neural architecture search”, “artificial intelligence”, “hyperparameter optimization”, “bayesian optimization”, “meta-learning”, and “classification”. For the entire list, please consult Figure A6.

The second cluster, in blue, includes 13 keywords: “automated machine learning (automl)”, “optimization”, “training”, “neural networks”, “computer architecture”, “task analysis”, “data models”, “feature extraction”, “neural architecture search (nas)”, “search problems”, “computational modeling”, “pipelines”, and “adaptation models”.

Figure A7 illustrates a thematic map based on the authors’ keywords, grouped into seven clusters and four main themes. The density represents the development degree, while the centrality represents the relevance degree. This illustration presents a comprehensive overview of the main trends in the AutoML domain. In theory, niche themes are intricate and demand advanced knowledge to be applied, while motor themes promote innovation in a specific field (in this study, the AutoML domain). Basic themes address fundamental topics, and emerging or declining themes are topics considered to become less relevant.

Cluster number one, colored in red and characterized by high centrality and low density, consists of an array of keywords including “automated machine learning”, “machine learning”, “artificial intelligence”, “bayesian optimization”, “classification”, “feature selection”, “COVID-19”, “feature engineering”, “genetic algorithm”, and “hyperparameter tuning”.

The second cluster, in blue, belongs to motor themes and is defined by high relevance and development degrees. The most frequent keywords in this cluster are “prediction”, “autogluon”, “diagnosis”, “pycaret”, “sustainability”, “xgboost”, “digital health”, and “feature importance”. Autogluon is an open-source toolkit used for training machine learning models for image and text classification, object detection, and tabular data prediction with little or no prior technical experience. Pycaret is a Python library that automates machine learning workflows. XGboost is also an open-source library that uses gradient boosted decision trees for training on large datasets.

The third cluster (in green) is located at the border of all four quadrants and includes keywords with medium centrality and density: “radiomics”, “breast cancer”, “predictive model”, and “shapley additive explanations”.

Cluster number four (colored in orange) with a high relevance degree and medium density includes an extensive list of keywords, the most frequent ones being “automl”, “deep learning”, “neural architecture search”, “automated machine learning (ml)”, “optimization”, “training”, “hyperparameter optimization”, “neural networks”, “meta-learning”, and “computer architecture”.

Two clusters with low relevance and high development degrees are included in the niche themes: cluster number seven, colored in grey, and cluster number five, in brown. While cluster number five consists of two keywords, “qsar” (an acronym for quantitative structure–activity relationship, a model used in machine learning to quantify the relationship between a chemical structure and its biological activity), and “shap” (an acronym for SHapley Additive exPlanations, a framework used for explaining the output of machine learning models), cluster number seven includes only one keyword: “3d echocardiography”.

Cluster six (in pink) belongs to the emerging or declining themes and includes only one keyword: “blockchain”.

Figure 13 presents the thematic evolution based on the 250 most prevalent authors’ keywords, separated into three intervals. The first time interval, from 2006 to 2019, has a longer time frame because, between 2007 and 2015, no articles were published, and the production for the entire interval was limited to only 32 articles, resulting in only three distinct themes: “automated machine learning”, “risk assessment”, and “intelligent transportation systems”.

In the second period, from 2020 to 2022, the researchers conducted studies in various areas of interest. The most frequent themes were “classification”, “feature selection”, “supervised learning”, “image classification”, “natural language processing”, “predictive modeling”, “timeseries forecasting”, “design optimization”, as well as “performance”.

In the third interval of the dataset (2023–2024), more specialized studies have been issued. The authors became interested in “explainable machine learning”, “sustainability”, and “predictive models”. “Blockchain”, “genetic algorithm”, “remote sensing”, and “multi-objective optimization” were the most common themes. Also, different aspects of security have attracted the researchers’ attention: “network security”, “adversarial attacks”, and “maritime safety”.

3.5. Mixed Analysis

In this section, a mixed analysis is executed to provide further insights into the AutoML domain and emphasize the connections among various categories: countries, authors, journals, affiliations, and keywords.

Figure 14 provides information on the connection between the three fields: countries (left), authors (middle), and journals (right). As expected, China sits first in the author’s country of origin list. Greece, Germany, and the USA are occupying the following positions. Zhang Y., Zhang I., and Tsamardinos I. are the top three authors based on the number of publications related to the topic of the bibliometric study. IEEE Transactions on Pattern Analysis and Machine Intelligence, IEEE Transactions on Knowledge and Data Engineering, and Journal of Machine Learning Research are the journals that the authors preferred the most to publish in. Figure 14 also depicts the high level of collaboration between authors from different countries and the authors’ tendency to publish in various journals.

The connection between affiliations (left), authors (middle), and keywords (right) is depicted in Figure 15.

Soochow University from China is the most affiliated institution. The Chinese Academy of Sciences and the University of Crete are also significant contributors to the literature connected to automated machine learning. While the most productive authors are Hutter F., Wang Y., and Yin My., the most frequently encountered keywords are “automated machine learning”, “automl”, and “machine learning”.

More information can be extracted by analyzing Figure 15. Several scientists from the top 20 have affiliations with multiple institutions, while others have no affiliation with any university from the list. The increased levels of collaboration and affiliation, in addition to the interest in diverse topics around the AutoML domain, are key elements that contribute to the scientific value of the published papers and the advancement of technology.

3.6. Topic Analysis

In terms of topic analysis, LDA has been firstly conducted on the extracted papers’ abstracts, titles, author keywords, and keywords plus. As a result of this analysis, the inter-topic distance map has provided a clear separation between two main research topics, as highlighted in Figure 16. The size of each circle indicates topic prevalence, namely Topic 1 accounts for about 54.1% of the tokens, while Topic 2 accounts for about 45.9% of the tokens. Based on the percentages provided, it can be observed that the distribution is balanced.

Considering the results obtained through LDA for Topic 1 depicted in Figure 17, a series of key terms are identified, such as network, optimization, architecture, prediction, classification, pipeline, training, accuracy, design, feature, search, selection, neural_network, hyperparameter, neural_architecture_search, automated_machine_learning, and strategy, which suggest the strong technical and algorithmic focus of AutoML research. This topic retains 54.1% of tokens and emphasizes the algorithmic and technical foundations of AutoML, including neural architecture search, optimization methods, and automated pipeline design.

As a result, it is expected that the papers associated with this topic focus on core themes such as model design, neural architecture search (NAS), optimization strategies, hyperparameter tuning, pipelines, underlying more the role of AutoML as a computational methodology for improving efficiency and performance in machine learning workflows.

LDA Topic 2 retains 45.9% of tokens and highlights the application of AutoML in real-world domains, with a particular emphasis on healthcare and clinical prediction tasks, such as patient diagnosis, disease classification, and medical image analysis. These research directions are suggested by the occurrence of the dominant keywords such as prediction, patient, accuracy, feature, analysis, classification, image, ai, score, diagnosis, disease, clinical, healthcare, and detection, identification, as depicted in Figure 18.

Given the strong connection of Topic 1 with computer science and AI development communities, it can be stated that LDA Topic 1 refers to Algorithmic Foundations and Model Optimization in AutoML, while LDA Topic 2 is more related to Applied AutoML in Healthcare and Prediction Tasks. This dual structure, pointed out by the occurrence of the two distinct topics, reflects the field’s evolution: AutoML is both a methodological innovation and a practical enabler of AI adoption in critical sectors.

Next, a BERTopic analysis is conducted. The results feature a number of four topics, distributed as depicted in Figure 19, with the five most-used words highlighted in Figure 20. In the following, a description of the four identified topics is provided based on the top ten most frequent words extracted for each topic.

First, Topic 0 features a series of keywords such as performance, prediction, accuracy, patients, analysis, algorithms, and classification, capturing both the performance and the applicability of AutoML in healthcare. Compared to LDA, it can be noticed that Topic 0 in BERTopic is similar to LDA Topic 2 but more detailed, emphasizing evaluation metrics (accuracy, performance) alongside health-related applications (patients, analysis). Also, Topic 0 encompasses the highest number of papers under its umbrella.

Topic 1 uncovered by BERTopic analysis is characterized by a series of keywords such as search, neural, architecture, NAS, neural architecture search, and network. As observed, this topic puts a highlight on Neural Architecture Search (NAS) as a distinct subfield rather than just a generic technical foundation. Considering the LDA analysis conducted previously, this topic is very close to LDA Topic 1 but more semantically focused on NAS and deep learning model design.

With extracted keywords such as traffic, detection, network, intrusion, intrusion detection, framework, and accuracy, Topic 2 focuses on cybersecurity and anomaly detection. Considering the LDA analysis, it can be observed that this topic is not captured, which was expected given the reduced dimensions of this topic. Thus, it can be noted that BERTopic succeeds in capturing some uncovered and finer-grained, semantically coherent clusters than in the case of LDA, which shows a macro-subdivision of the topics.

Lastly, Topic 3 highlights AutoML’s use in education analytics (predicting dropout, academic performance), being supported by a series of keywords such as students, student, dropout, academic, university, predicting, and digital. This topic does not find correspondence in LDA results.

Given the results offered by the BERTopic analysis, the four identified topics are as follows: Topic 0—Performance and Healthcare Applications, Topic 1—NAS and Architectures, Topic 2—Intrusion Detection and Networks, and Topic 3—Student Performance and Academic Dropout Prediction.

4. Discussions and Limitations

In this section, we will discuss the results we obtained in our study, as well as the benefits and limitations that are introduced in this paper.

4.1. Bibliometric Analysis Results and Comparison with Other Studies

Our research aims to present in a technical manner the evolution of the AutoML field in the scientific world. It can be seen that the AutoML field had an insignificant impact until 2018, when it started to have exponential growth. This rapid growth has taken place in the context where the hardware technology on which these systems are based has also had a significant advance, but also related fields such as Machine Learning and Deep Learning have been intensively researched, as other studies have shown [99].

The rapid evolution in the field can also be seen from our bibliometric analysis, with China and the United States of America being the basic pillars on which the evolution took place. This observation has also been brought to light by Angarita-Zapata et al. [100], with the USA and China having the most published works in the field, highlighting these countries’ strategic emphasis on ML research. Therefore, national policies and AI agendas of these countries play an important role in research themes and collaborations. Ilic et al. [101] also underlined the fact that the paper production is linked to national initiatives and funding priorities. The study by Alvarez-Bornstein et al. [102] indicated that public funding often leads to higher-impact publications and promotes international collaborations. Nonetheless, the importance of collaboration between the public and private sectors is essential in driving innovation and regional development [103]. The study by Jee and Sohn [104] revealed that papers authored by firm-affiliated researchers, especially those solely from firms, have had a higher impact on the knowledge trajectory in AI research. A point in common with another bibliometric analysis is with the one written by Wang et al. [105], which acknowledges the merits of Chinese universities in the field of Artificial Intelligence. As shown in our analysis, the Chinese Academy of Sciences occupies first place with 51 publications, the next two positions being occupied by the University of California System with 45 published articles and Sichuan University with 23 published articles.

Furthermore, the collaboration between countries and universities had a significant positive impact on the studied field, as the two countries mentioned above have successful collaborations with countries all over the world, notably being the collaborations with the UK, Australia, and Canada. Among the authors most involved in AutoML research, it is worth mentioning names such as Wang Y., with 12 published papers, and Tsamardinos I., with 11 papers, but also names such as Hutter F. and Moore JH., who have an impressive calculated H-Index value of 6, being also the highest H-Index in the field.

4.2. Topics Versus Themes Discussion

Considering the LDA analysis and the thematic map results, it can be noticed that two complementary perspectives on AutoML research are identified. First, the LDA analysis revealed two major topic clusters: one focused on algorithmic and methodological innovation (e.g., neural architecture search, optimization, pipeline automation), and the other on applications in healthcare and predictive modeling (e.g., diagnosis, patient data, clinical prediction). On the other hand, the thematic map confirmed this duality, with AutoML, deep learning, and neural architecture search listed as motor themes and healthcare-related applications (prediction, diagnosis, radiomics, breast cancer) also identified as central and highly relevant.

Furthermore, the thematic map provided more information for the field by capturing basic themes, which gravitate around foundational concepts such as machine learning, artificial intelligence, hyperparameter tuning; niche themes, which have a focus on specialized methods like SHAP and QSAR; and emerging themes, such as blockchain, which were less visible in the global structure provided by LDA.

As a result, taken together, the two methods provide a comprehensive view over the field of AutoML: while LDA highlights the core structural division of the field into methodological and applied research, the thematic map positions these findings within a broader evolutionary framework of themes.

Referring to the results identified as a result of BERTopic analysis in comparison with LDA analysis, it can be observed that while LDA revealed two macro-research directions in AutoML, namely methodological innovation and real-world applications, BERTopic analysis provided a more granular view.

Specifically, BERTopic identified the following four topics: Topic 0—Performance and Healthcare Applications, Topic 1–NAS and Architectures, Topic 2—Intrusion Detection and Networks, and Topic 3—Student Performance and Academic Dropout Prediction. Thus, it can be stated that while the two approaches are complementary, they show different facets of the AutoML research, both at the micro and macro levels.

Table 8. Comparison LDA Topics–BERTopics–thematic maps.

LDA Topics	BERTopics	Thematic Map Clusters	Overlap/Notes
LDA Topic 1—Algorithmic Foundations and Model Optimization (NAS, optimization, pipelines, hyperparameters)	BERTopic 1—NAS and Architectures	Cluster 4—Motor Themes (AutoML, deep learning, neural architecture search, optimization, hyperparameter tuning, meta-learning)	Strong alignment: all three approaches confirm AutoML’s algorithmic and methodological core.
LDA Topic 2—Applied AutoML in Healthcare and Prediction (patients, diagnosis, clinical, prediction, accuracy)	BERTopic 0—Performance and Healthcare Applications	Cluster 2—Motor Themes (prediction, diagnosis, autogluon, digital health); Cluster 3—Applied Themes (radiomics, breast cancer, predictive models)	Overlap across all methods: healthcare and prediction consistently emerge as key AutoML applications.
–	BERTopic 2—Intrusion Detection and Networks	–	Unique to BERTopic: cybersecurity applications (traffic, intrusion detection) do not surface in LDA or thematic map.
–	BERTopic 3—Student Performance and Academic Dropout Prediction	–	Unique to BERTopic: education-related applications absent from LDA and Thematic Map.
–	–	Cluster 1—Basic Themes (ML, AI, AutoML, Bayesian optimization, feature selection, COVID-19)	Identified only by thematic map: foundational concepts treated as background context in LDA and BERTopic.
–	–	Clusters 5 & 7—Niche Themes (SHAP, QSAR, 3D echocardiography)	Unique to thematic map: specialized subfields not detected by topic models.
–	–	Cluster 6—Emerging/Declining Themes (Blockchain)	Unique to thematic map: peripheral exploratory direction not visible in LDA or BERTopic.

provides a comparison between the results obtained in LDA Topics and BERTopics. As observed, there are a series of overlaps—as discussed above—as well as unique topics depending on the type of analysis.

4.3. Discussions of Specific Themes

In this sub-section, various topics addressed by researchers in the field are discussed in order to have a better understanding of the applicability of AutoML. Two main fields have been selected: the AutoML in medicine due to its occurrence in the analyses provided above, and AutoML in finance—an underrepresented application area compared to healthcare or cybersecurity, which dominated the topic modeling results, but an important one, as it has increased potential in future works.

4.3.1. Implications of AutoML in Medicine

This is one of the branches in which AutoML best left its mark, especially by identifying the different diseases that patients may have. Due to its ability to process large datasets and its capability for hyperparameter tuning, AutoML has proven to be very useful in medical applications. One study that explores the applicability of AutoML in medicine is the one written by Potluru et al. [106], which investigates the ability of AutoML to recognize skin cancer. They developed an application that has an accuracy of 84.4%, a surprising result considering the research on AutoML is still in its early stage. Another article that describes the benefits of AutoML in medicine is that of [80] for the prediction of cardiovascular diseases. Using an AutoML algorithm and training the network with over 400,000 images, they were able to increase the accuracy of existing algorithms.

Another study that highlights the need for further development of AutoML in medicine is [107], which compares different AutoML solutions. They compare the ability of algorithms to identify different diseases (pneumonia, skin cancer, Alzheimer’s), thus emphasizing the robustness of AutoML in medicine.

AutoML in medicine could represent a remarkable advancement, as doctors will not need the assistance of a programmer to develop various solutions but only input data to train the network.

4.3.2. Implications of AutoML in Finance

Another area in which AutoML is intensively studied is the one of finance. The power to exploit large datasets is also of great interest in this sector due to the large amount of banking data that can be processed.

One of the main uses of AutoML in finance is to calculate credit risk to predict whether a customer can honor their interest or not. This method is analyzed by Marc Schmitt et al. [108], which demonstrates that the accuracy and efficiency of predictions are improved but also fosters trust and collaboration between humans and AI systems.

Another effective way for the banking system to use AutoML is in fraud detection. As Jha et al. [109] predicted, bank frauds cost the financial industry billions of dollars. They analyze transactions to recognize buyer behavior and then identify fraudulent transactions with the help of aggregations. A similar article talks about the use of intelligent systems in personalized offers that banks can offer to customers, following the customer’s lending behavior [110].

AutoML is also used for portfolio management. By analyzing historical data, AutoML can suggest opening and closing positions to minimize investment risks. The new techniques used show that performance is greatly improved in terms of Alpha and Sharpe ratio [111].

4.4. Key Limitations to Applicability of AutoML

Even though the evolution of AutoML has been remarkable in recent years, there are some limitations that negatively affect both the evolution and the efficiency of systems that are based on such technologies. In many of these systems, there are issues regarding data confidentiality and security but also in the number of input data which is crucial in order to have conclusive results.

As we have seen before in [106], in order to increase the accuracy of the systems, no less than 400,000 input data were used to train the model. Karmaker et al. [84] discuss the lack of data in AutoML systems and how this can negatively affect the obtained results. Also, even though hardware technology has evolved a lot, both processing power and data storage can be a problem for applications with such large datasets.

Another limiting factor on data is confidentiality. As observed, AutoML is a very studied field in medicine, where data protection and confidentiality are crucial for patients. In the article written by Pang et al. [112], the risks that lack of security poses are discussed in various areas in which AutoML is used, as well as the answers to the questions of whether AutoML systems are more vulnerable than traditional machine learning systems, but also what are the possibilities to emerge.

Moreover, the algorithmic biases in AutoML must be addressed. The algorithms often reflect biases present in the training data, therefore leading to biased outcomes [113]. Models tend to perform worse on underrepresented demographics due to distributional shifts in the training data and can propagate stereotypes and health disparities [114].

Also, biases in algorithms used for loan or credit limit approvals and credit score estimation can result in unfair financial outcomes [115]. Another bias can be identified in text classification models, which can lead to unfair treatment based on identity tokens, affecting both individual as well as group fairness [116].

Therefore, the ethical, transparent, and fair use of machine learning applications can be ensured by outlining strategies and best practices for mitigating the algorithmic biases.

4.5. Limitations Related to Dataset Extraction and Used Database

In the following, we propose for analysis the limitations that exist in the extracted dataset. The first limitation that could be discussed is the extraction of articles from a single database, since we are using data only from the WoS.

The choice to extract data from a single collection of papers was made based on several reasons. The first reason why we chose only one database is because, by choosing multiple databases, a paper could appear duplicated in our analysis, and the number of citations could be different for each paper, which would produce different results. For example, the article by Elsken et al. [77] has 1181 total citations according to WoS, 1992 on Scopus, while IEEE has not indexed the paper.

Also, we chose WoS as the main database because of its reputation but also because of the special features that give us the chance to do a more complex analysis (such as keyword plus, which is specific to this database).

Biblioshiny, the used tool to conduct the analysis, supports files from only specific databases (WoS, Scopus, Dimensions, OpenAlex, Lens.org, PubMed, and Cochrane Library), making it impossible to use this tool with other databases, such as ArXiv.

Another limitation that may arise is the constraint on language. As stated at the beginning of this paper, we filtered the articles so that only those in English would appear for analysis. This decision was made because the analysis of the works in different languages could compromise the extraction of the words, and thus analysis of them would also suffer. However, the filtering on the English language excluded from the analysis only six articles, which represent a percentage of less than 0.1%, so the bibliometric analysis is not affected.

Additionally, filtering the documents to only “articles” can be viewed as a limitation, as other types of studies, such as conference papers or book chapters, can contain relevant advancements in the studied field. This measure was added to accurately compare the documents in terms of the number of citations and content.

Moreover, a limitation may arise from the selection of the most cited papers. The reasoning behind this decision is that citations have been used to measure the impact and relevance of papers, reflecting how often a study is referenced by others [117]. Citations also provide a quantitative basis for evaluating the qualitative impact of research. Therefore, papers with a higher number of citations are considered more influential than those less cited [118]. Nonetheless, Bornmann and Marx [119] argued that a high number of citations does not necessarily correlate with a high level of usefulness. Other metrics can be used to measure the impact and relevance of a paper, such as journal impact factor, the authors’ position in the author list, and H-index [120]. Altmetric Attention Score (AAS), social media mentions, as well as blog and news article mentions can also provide insights into the impact of research [121].

Lastly, the words and topic analyses were based on the keywords extracted from the titles, abstracts, author keywords, and keywords plus and not from the body of the articles. Pottier et al. [122] found that authors frequently use redundant keywords in titles and abstracts, which can limit the semantic richness of the extracted keywords. Despite this, Zhang et al. [123] reported that full-text-based extraction can introduce noise, which can decrease performance, and this point must also be considered when conducting a bibliometric analysis.

The results of the two types of topic analysis, namely LDA and BERTopic, revealed both similar and distinct topics, proving the utility of both methods. It should be mentioned that, in terms of coherence score (c_v), BERTopic outperformed LDA (0.71 for BERTopic versus 0.47 for LDA) when considering the first ten tokens per topic. Topic coherence measures how semantically consistent or interpretable the top tokens of a topic are by estimating how often the top N tokens of a topic co-occur in the same documents or within a certain window. This finding is in accordance with other studies, such as Vallelunga et al. [124], who also noticed that BERTopic achieved a better performance in terms of evaluation metrics than LDA.

5. Conclusions

This analysis represents a detailed inspection of all published articles in the AutoML field. As has been presented, the field has experienced an exponential growth in recent years, being also influenced by the current context, in which related fields such as machine learning and deep learning are expanding rapidly. As presented in the first part of the analysis, essential questions about the AutoML field were asked in order to better understand the role it plays in scientific research. As we have previously analyzed, the answers to the questions are as follows:

The journals with the greatest influence in the scientific world, taking into account the H-index that each one has, are Scientific Reports with an H-index value of 11, followed by IEEE Transactions on Pattern Analysis and Machine Intelligence and IEEE Access, with H-Index values of 9 and 8, respectively.

From the perspective of the extracted dataset, the most prolific researcher is Wang Y. with 12 published articles, Tsamardinos I. being the second in the field with 11 papers. They also share the first two places when we look at the best productivity according to the H-index.

One point that could be improved could be considered the collaborations between countries that are nevertheless at a rather low level, noting that the co-authors of a paper generally come from the same country. This observation is also suggested to us if we analyze the collaboration network, which is a fairly restricted network between researchers.

Another point analyzed was that of affiliations and the countries with the highest production of articles in the AutoML field. China and the USA are the main countries that generate articles, and from these two countries also come affiliations with the most published papers. Some significant examples of universities interested in research on this topic are Chinese Academy of Sciences, University of California System, and Sichuan University.

Overall, the combined analyses related to topic modeling, such as LDA and BERTopic, as well as the analysis related to themes, namely the thematic map, show that AutoML research is structured around two central pillars: algorithmic innovation and applied adoption, with healthcare as the most prominent domain. BERTopic analysis reveals a series of additional emerging areas such as cybersecurity and education, while the thematic map situates these developments within a broader evolutionary framework, highlighting basic, motor, niche, and emerging themes. Together, the results confirm AutoML’s dual role as both a driver of methodological progress and a catalyst for domain-specific applications.

As for the analyzed works in the field, there is extensive research in the medical area but also an emerging one in the financial sector. Many papers analyzed had impressive discoveries, the results obtained with the help of AutoML technologies being superior to those existing in the literature.

Future research should focus on several key areas. The first one is healthcare applications, as using AutoML can improve diagnostics, treatment planning, and personalized medicine [125].

Another area of research is agriculture, as AutoML can be leveraged for precision farming, crop monitoring, but also sustainable practices [126].

Future research should also focus on applying AutoML to optimize financial models, for risk assessment, customer service, and economic forecasting [127].

Other important aspects of developing AutoML applications must be explored in future studies, such as explainability, transparency, and robustness [128]. These practices ensure user data protection, trustworthiness, and compliance with regulations.