A Map of the Research About Lighting Systems in the 1995–2024 Time Frame

Abstract

Lighting Systems (LSs) are a key component of modern cities. Across the years, thousands of articles have been published on this topic; nevertheless, a map of the state of the art of the extant literature is lacking. The present review reports on an analysis of the network of the co-occurrences of the authors’ keywords from 12,148 Scopus-indexed articles on LSs published between 1995 and 2024. This review addresses the following research questions: (RQ1) What are the major topics explored by scholars in connection with LSs within the 1995–2024 time frame? (RQ2) How do they group together? The investigation leveraged VOSviewer, an open-source software largely used for performing bibliometric analyses. The number of thematic clusters returned by VOSviewer was determined by the value of the minimum number of occurrences needed for the authors’ keywords to be admitted into the business analysis. If such a number is not properly chosen, the consequence is a set of clusters that do not represent meaningful patterns of the input dataset. In the present study, to overcome this issue, the threshold value balanced the score of four independent clustering validity indices against the authors’ judgment of a meaningful partition of the input dataset. In addition, our review delved into the impact that the use/non-use of a thesaurus of the authors’ keywords had on the number and composition of the thematic clusters returned by VOSviewer and, ultimately, on how this choice affected the correctness of the interpretation of the clusters. The study adhered to a well-known protocol, whose implementation is reported in detail. Thus, the workflow is transparent and replicable.

1. Introduction

Lighting Systems (LSs) are everywhere: inside cities, in homes, in offices, in public buildings, and along streets and highways. The advent of LSs has sped-up human progress, improving, at the same time, the quality of life [1]. Both industry and academia are engaged in the advancement of this relevant category of applications. From the industry side, it can be observed that the lighting industry worldwide is expected to reach a projected revenue of USD 223.3 million by 2030. A compound annual growth rate of about 7.0% is expected for the worldwide lighting industry from 2025 to 2030 (source: https://www.grandviewresearch.com/horizon/outlook/lighting-market-size/global, accessed on 20 March 2025). From the academic side, it can be observed that the research domain of LSs has been active for many years, and this has produced a huge number of studies published in journals, conference proceedings, and book chapters [2].

Bibliometric Analysis (BA) is a popular and rigorous method for exploring and analyzing large volumes of scientific data (hundreds, if not thousands) [3,4,5]. BA comprises two complementary techniques: Performance analysis and Science mapping. The present work focuses on keyword co-occurrence analysis, which is one of the pillars of Science mapping [6,7,8]. Co-keyword analysis, combined with network analysis, can uncover the knowledge produced by research in a given scientific field [9]. The VOSviewer open-source software is used to construct easy-to-interpret bibliometric networks [10]. VOSviewer is largely used in bibliometric studies [11,12,13,14,15,16,17,18,19,20].

The present study has the following merits:

• It provides a map of the state of the art of research in the domain of LSs, within a 1995–2024 time frame. The reported bibliometric research is somewhat urgent because, as will be proven subsequently, the extant literature about LSs is becoming huge; nevertheless, as far as we know, our research is the first study in such a category. In recent years, a recurrent refrain has been that the Internet of Things (IoT) and Artificial Intelligence (AI) are the technologies to rely on, so that traditional LSs can make a qualitative leap by becoming smart and at the same time sustainable. This study highlights that the state of the art of LSs confirms this widespread thinking, but at the same time it tells us that the relevance of AI is still marginal. Such a conclusion confirms previous reports from both academia and industry.
• It meets the rigor requirement for this type of literature review. This objective was pursued by adopting a well-known protocol, whose implementation is reported in full detail. Thus, the investigation is transparent and replicable.
• It delves into the impact that the use/non-use of a thesaurus of authors’ keywords had on the number and composition of the thematic clusters returned by VOSviewer and, ultimately, on how this choice affected the correctness of the interpretation of the clusters. Published articles adopting keyword co-occurrence analysis can be split in two groups: the smaller group mentions the usage of a thesaurus of authors’ keywords [9,11,17,18,21,22,23], while the larger group does not. References [7,12,15,16,20,24,25,26,27,28,29]. Among the seven works belonging to the first group, references [11,17,23] just mentioned that a thesaurus was used, reference [22] focused on the relevance of adopting a thesaurus for data clean-up, finally, references [9,18,21] detailed the steps needed to build a thesaurus of authors’ keywords that complies the VOSviewer’s format. Ref. [18] listed, in addition, the 34 terms which comprised the thesaurus.
• The choice of the number of clusters of authors’ keywords was obtained by balancing the numerical value of four independent methods belonging to the family of so-called “clustering validity approaches” [30,31,32,33,34] with what the experts (the authors on this occasion) judged to be a meaningful partition of the input dataset. Such an approach was recommended in [35,36] for overcoming the issue pointed out in ref. [37]. In the latter work, the authors observed the absence of a reason behind their choice, common to most published articles talking about bibliometric analyses, of the value of the minimum number of occurrences for an authors’ keyword to be included in the BA.
• Regarding the latter two points, it follows that the present study also gives two recommendations from a methodological perspective. First, it points out that keyword co-occurrence analyses must be driven by a robust thesaurus of keywords. Second, the VOSviewer parameters must be optimized.

The remaining part of this work is structured as follows: Section 2 introduces the background necessary to delve into later sections. Section 3 describes the Research Method adopted in the study. Section 4 presents the results of the bibliometric analysis and then replies to the research questions. Section 5 analyses potential threats to the validity of the findings of the present study, while Section 6 concludes the paper. Two appendices complete this paper. Abbreviations lists the used acronyms, while Appendix A shows the entire thesaurus of authors’ keywords entered as input to VOSviewer.

2. Background

This section first recalls the major features of the VOSviewer software, then it focuses on methods that previous research suggested as necessary to accomplish a correct estimation of the number of clusters of authors’ keywords.

2.1. VOSviewer Terminology [38]

This software allows one to create, visualize, and explore networks. A network is composed of nodes and links. Nodes are the objects of interest (authors’ keywords in the present study), while links express a co-occurrence between pairs of nodes. VOSviewer provides a Network Visualization, an Overlay Visualization, and a Density Visualization. Below, we spend a few words about the first two visualization categories.

In Network Visualization, nodes are represented by a circle (the default) and by their label (i.e., the name of the keyword). Nodes are grouped into non-overlapping clusters, labeled as positive numbers. Clusters are colored, the color of a node is the same of the cluster it belongs to. This type of visualization is drawn according to the values of three numeric weight attributes: Links, Total link strength, and Occurrences. The Occurrences attribute denotes the number of papers in which a keyword occurs. The value of this attribute for a node indicates the relevance of the latter in the network. In the visualization, nodes (as well as the associated label) with higher Occurrences are larger than nodes with lower Occurrences. Let us call n a generic node, the value of the Links and Total link strength attributes indicate, respectively, the number of links involving n and the total strength of those links. VOSviewer represents the strength of a link as a positive numerical value. The higher this value, the stronger the link. The thickness of the link between two nodes represents the strength of the link. The strength of a link indicates the number of publications in which two keywords occur together. VOSviewer takes a distance-based approach to visualizing bibliometric networks. The distance between two nodes/keywords in the visualization roughly indicates the relatedness of the keywords in terms of co-occurrence links. In general, the closer two keywords are located to each other, the stronger their relatedness.

The Overlay Visualization depicts a network structurally identical to the previously described network, except that nodes are differently colored. In VOSviewer, by default, colors range from blue (lowest score) to green to yellow (highest score). Colors are shown in the bar located in the bottom right corner of the visualization. The colors are determined by the value of three numeric score attributes: Avg.pub.year, Avg.citations, and Avg.norm.citations. (The latter two attributes are not relevant to the present BA, therefore they are ignored). When Avg.pub.year is selected in the Scores drop down list in the right-hand panel of VOSviewer, then the color of a node/keyword indicates the average publication year of the articles in which that keyword occurs, with yellow indicating terms that mainly occur in recent publications and blue indicating terms that mainly occur in older publications.

2.2. On the Choice of the Number of Clusters

Selecting an appropriate number of clusters (let us call this $\mathcal{C}$) ensures that similar data points are grouped together, without overfitting or underfitting. A poorly chosen $\mathcal{C}$ may result in clusters that do not represent meaningful patterns in the data. In the case of networks of keyword co-occurrences, the output of the clustering stage is determined by the value of the minimum number of occurrences (briefly, the $\mathcal{T}$ threshold) for a keyword to be admitted to the BA. Most published articles on bibliometric analysis do not provide any explanation about the rationale behind the choice of $\mathcal{T}$. The explanations in these works often look like the following: “the minimum occurrence threshold of authors’ keywords was set at […]” [39] or other equivalent sentences [8,40]. Lim et al. [9] suggested adjusting the $\mathcal{T}$ value in order to have a network at least of 20 nodes/keywords, but no theoretical foundation was provided to justify their proposal. Frequently, the $\mathcal{T}$ value changes with the study. For example, in the three articles just mentioned, $\mathcal{T}$ was equal to 10, 30, and 3, respectively. Such a variety of choices denotes that there is no one-size-fits-all case, which, in turn, raises the issue of how the $\mathcal{T}$ value should be set. To overcome this issue, in the present study, four independent methods were considered. They are briefly recalled in the following:

• Elbow method [41]
  This method first measures the Within-Cluster Sum of Squares (WCSS) for a varying cluster number, then it selects the cluster for which the change in WCSS starts to decrease. The mathematical definition of WCSS may be found in [33]. In simple words, we can say that WCSS measures how well the data points are clustered around their respective centroids. With respect to leveraging the Elbow method, it is necessary to note that it provides a starting point, but since, in some cases, the Elbow point may not be distinctly visible, a subjective interpretation is then required. In addition, it is useful to remember that such a method is specifically tailored for centroid-based clustering algorithms like KMeans, so it does not necessarily work well for other methods [36].
  To reinforce the clustering decision, the Davies–Bouldin index, the Silhouette score, and the Calinski–Harabasz index [33] are frequently adopted, either individually or together.
• Davies–Bouldin Index [42]
  The Davies–Bouldin (DB) index is defined as the ratio between the intra-cluster distance and the inter-cluster distance. The former distance is calculated as the mean distance between each element in a cluster to the centroid of that cluster. (It offers insights into how closely grouped the elements within a single cluster are); while the latter distance measures how far apart each cluster’s centroid is from the other clusters. (The larger this distance, the more separated the clusters).
  A lower score of DB index signifies a better cluster formation, with zero being the absolute ideal score. It has been remarked [43] that relying solely on the DB index for cluster analysis would be imprudent, because such an index is only effective with clusters of convex shape. The DB index should be used in conjunction with other metrics for a more holistic evaluation.
• Silhouette score [44]
  The Silhouette score (Sscore) of an input dataset measures how dense and well separated the clusters are. The mathematical definition of this metric may be found in [35]. For the purposes of the present study, it is sufficient to recall the following: Sscore is defined as [(b − a)/max(a,b)], where “a” and “b” denote, respectively, the mean intra-cluster distance and the mean nearest-cluster distance for each sample in the input dataset. To clarify, b is the distance between a sample and the nearest cluster that the sample is not a part of. Sscore takes values in the range [−1, 1]. A higher score indicates better clustering. Values near 0 indicate overlapping clusters, while negative values generally indicate that a sample has been assigned to the wrong cluster.
• Calinski Harabasz Index [45]
  The Calinski–Harabasz (CH) index is calculated as the ratio of the sum of inter-cluster dispersion and the sum of intra-cluster dispersion for all clusters (where the dispersion is the sum of squared Euclidean distances). In simple words, the CH index measures how dense and well separated the clusters are. The mathematical definition of this metric may be found in [35]. A higher CH score means better clustering.

In conclusion, it can be said that the attention that should be given to the selection of a strategy for finding the $\mathcal{C}$ value is well summarized in ref. [35]. In that article, the authors state that the selection of the $\mathcal{C}$ value should be performed by balancing the suggestions from the application of cluster validity indices with what experts judge to be a meaningful partition of the specific dataset. The same recommendation ended the more recent ref. [36].

3. Research Method

The research methodology followed a well-established protocol [4,5] comprising the following stages:

• Definition of the research objectives.
  The present work is the outcome of a cooperation between the Department of Industrial & Information Engineering & Economics of University of L’Aquila (Italy) and an Italian SME (B2B S.r.l). Recently, B2B S.r.l stated a desire to obtain an unbiased map about the state of the art of LSs. Their interest originated within national IT projects that B2B S.r.l. is involved in. The aim of the research is represented by the two Research Questions (RQs) this study aimed to answer:
  RQ1. What are the major topics explored by scholars in connection with LSs in the 1995–2024 time frame?
  RQ2. How do they group together?
• Literature search and data collection.
  In ref. [4], it was clarified that it is not necessarily convenient to use more than one scientific database in bibliometric research. A more serious issue comes from the inevitable duplication of publications, which makes the findings of the analysis debatable. In the case of Scopus and Web of Science, for example, the percentage of duplicates tends to be large, because most of journals are scanned in both databases. A previous research work by Singh et al. [46] showed that about 99% of the journals indexed in Web of Science are also indexed in Scopus. Moreover, the integration of items belonging to files obtained by querying distinct databases is time consuming, since they provide article information in a different format. In the present study, we queried Scopus rather than Web of Science due to its significantly greater coverage of published literature in the field of LSs, as will be proven subsequently.
  In ref. [47], it was remarked that a search string can be either generic or specific. We opted for the first option, since generic search strings maximize the “recall” value (i.e., the fraction of the documents that are relevant to the query that are successfully retrieved). The search string was the following: (“Lighting system” OR “Light control”). These terms are directly related to the study RQs. Previous research has stressed that to extract the literature relevant to the review aim, the terms in the search string are crucial [4]. The above search string was restricted by adding filters for publication type (articles, conference papers, review, and book chapters) and language. These filters represent what in ref. [4] the authors called the inclusion/exclusion criteria to be applied in bibliometric research. The final search string, plus the filters as written by Scopus, were as follows:
  TITLE-ABS-KEY ((“Lighting system” OR “Light control”)) AND
  PUBYEAR > 1994 AND PUBYEAR < 2024 AND (LIMIT-TO (DOCTYPE, “ar”) OR
  LIMIT-TO (DOCTYPE, “cp”) OR LIMIT-TO (DOCTYPE, “re”) OR
  LIMIT-TO (DOCTYPE, “ch”)) AND (LIMIT-TO (LANGUAGE, “English”))
  Scopus returned 12,148 documents, which were saved to the following file:
  Scopus12148.csv.
• Data screening and Preprocessing
  When multiple databases are queried, cleaning the output dataset is mandatory to ensure accuracy. Basically, it is necessary to remove duplicates and correct authors’ names. This effort was not required in our case, since we only queried Scopus. About the preprocessing, it first included the semi-automatic construction of a thesaurus of authors’ keywords, then the investigation of the best $\mathcal{T}$ value. Section 4 discusses both these arguments in connection with the clustering effectiveness. In practical terms, this review answers a further relevant research question:
  RQ3. What is the impact of the thesaurus of authors’ keywords and the $\mathcal{T}$ value on the clustering effectiveness?
• Selection of Bibliometric techniques
  The Scopus engine computes statistics for the “publication” and “citation” performance of authors, journals, institutions, and countries. These bibliometric indicators were developed by running a Performance analysis on the dataset output of the search [4]. To match the aim of the present review and, hence, to be able to answer the RQs, the Science Mapping analysis procedure was selected [4,5]. Specifically, keyword analysis was chosen, since this bibliometric method allows mapping relevant publications on a given topic ([3,25]) (in our case the LSs domain), because the authors’ keywords represented the major themes of the underlying publication (before delving into the investigation of its content through a systematic literature review). Moreover, such an analysis allows tracking the evolution of the reference research field over time, and hence recognizing emerging trends.
• Data analysis, Visualization, and Reporting
  Data analysis and visualization were carried out leveraging VOSviewer. Preliminary, the authors entered the Scopus12148.csv file (returned by Scopus) into VOSviewer. The following options were selected in sequence: (a) create a map based on bibliographic data; (b) read data from the bibliographic database file. As the counting method, full counting was selected. The interpretation of the results is the subject of Section 4, where the three RQs are also answered.

4. Results

This section goes into detail about (a) data collection; (b) data preprocessing (i.e., thesaurus construction and computation of the number of clusters); (c) analysis of the network of the keyword co-occurrences without a thesaurus; and (d) analysis of the network of the keyword co-occurrences in the presence of the thesaurus. The latter argument is split into a node analysis, link analysis, and temporal evolution of the research topics. By summarizing the results, it will be possible to answer the three RQs (Section 6).

4.1. Data Collection

As mentioned in the previous section, the Scopus reply to the input search string was the Scopus12148.csv file composed of 12,148 documents. (Queried with the same search string, Web of Science returned 5577 items, i.e., less than half of the items returned by Scopus.) In addition, Scopus provided the following statistics (built by performing a Performance analysis on the .csv file): documents by year; documents per year by source; documents by author; documents by affiliations; documents by country or territory; document by type; documents by subject area; documents by funding sponsor.

Table 1. Distribution of papers over the years.

Year	Documents	Year	Documents
2024	923	2009	301
2023	895	2008	263
2022	873	2007	245
2021	761	2006	221
2020	784	2005	172
2019	823	2004	151
2018	778	2003	116
2017	712	2002	96
2016	594	2001	89
2015	548	2000	91
2014	536	1999	85
2013	501	1998	84
2012	430	1997	74
2011	414	1996	93
2010	339	1995	82

and

Table 2. Aggregation of papers by type.

Document Type	Documents	Document Type	Documents
Conference Paper	5699 (46.9%)	Book Chapter	286 (2.4%)
Article	5786 (47.6%)	Review	377 (3.1%)

show, respectively, “documents by year” and “documents venue” to provide an overview of the research performance in the LS domain.

4.2. About Data Preprocessing

As said in Section 3, the preprocessing of the Scopus12148.csv file was carried out in two sequential steps. They are described in the following:

4.2.1. The Thesaurus

In the LSs domain, there are no publicly available thesauri of terms, hence the need to create one. It has been noted that the semi-automatic construction of thesauri using bibliometric analysis tools produces good results and in less time than completely manual creation [48]. The approach followed in this work for the construction of the thesaurus of the authors’ keywords is in line with [48].

Initially, a thesaurus of the authors’ keywords was built. The construction was semi-automatic, as explained hereafter. Using VOSviewer, a network of keyword co-occurrences was built starting from the Scopus12148.txt file. VOSviewer returned 22,999 distinct keywords. Setting $\mathcal{T}=5$ (similar values have been used in many previously published articles), 1096 keywords satisfied the constraints. Of these, 1089 were interconnected. We chose to limit our attention to these keywords. The 1089 keywords were copied into a word table. Then, the keywords column was sorted in ascending order and, subsequently a second column was added.

Eventually, the thesaurus was built manually and copied into a Thesaurus.txt file. The latter has the structure required by VOSviewer [10]. The contents of the Thesaurus.txt are given in Appendix A, which also shows the synonyms of each keyword in the thesaurus.

Table 3. The distinct keywords in the thesaurus (NoS = Number of Synonyms).

Keywords	NoS	Keywords	NoS
android application	1	LoRa	1
ANN	3	maintenance	1
arduino	4	MCU	3
artificial intelligence	1	ML	1
artificial lighting	3	mobile application	1
bluetooth	1	motion detection	4
building energy management	6	MPPT	1
cloud computing	2	MQTT	1
CNN	2	multi agent system	5
computer vision	3	museum lighting system	1
daylight	15	NB-IoT	1
deep q-learning	2	NN	2
Digital twin	1	occupant behavior	6
dimming control	4	office lighting	1
DL	2	outdoor lighting system	1
DRL	2	PIC MCU	1
edge computing	1	PIR sensor	3
embedded system	2	plant lighting	1
emergency lighting	1	predictive maintenance	1
energy management	21	PSO	1
environmental sustainability	3	public building	1
face recognition	1	public lighting system	5
fluorescent lamp	4	python	1
fog computing	1	q-learning	1
FPGA	1	raspberry pi	1
fuzzy control	4	regression analysis	1
genetic algorithm	3	remote monitoring	2
green building	4	RL	2
high pressure sodium lamp	3	security	1
home automation	1	Smart building	7
human centric lighting	2	smart city	2
hybrid lighting system	2	smart grid	2
illumination control	8	smart lighting control system	11
illumination sensor	2	smart street lighting system	6
image processing	4	STM32	1
indoor lighting system	4	street lighting control system	9
industrial lighting	1	sustainability	2
infrared sensor	2	tunnel lighting system	2
IoT	4	ubiquitous computing	2
LDR sensor	4	user comfort	8
LED	1	wireless communication	1
LED lighting system	33	WSN	8
light sensor	2	zigbee	3
lighting control system	25

lists the 87 distinct keywords that are part of the built thesaurus, together with the number of synonyms of each keyword. By considering the synonyms, it turned out that the 87 keywords in the thesaurus covered 312 keywords (i.e., the 28.65%) in the text file returned by VOSviewer. The ignored 777 keywords were either too generic or outside of scope of the present BA.

Table 4. Examples of ignored keywords.

Out-of-scope keywords	adaptive optics	cultural heritage
	agriculture	traffic control
	aircraft	traffic light control
	airports	traffic congestion
	algae	traffic light
	antioxidant	traffic light control
	COVID-19	intelligent transportation
Too-generic keywords	2D code	design
	2D materials	driver
	3D display	photosynthesis
	3D printing	visible light communication
	3D reconstruction biomass	power quality
	data-driven	light

shows examples of these two categories.

In Section 1, it was remarked that despite use of a thesaurus of authors’ keywords being highly recommended (see, for instance, ref. [10]), its adoption is rare. Refs. [18,22] are the only two articles that we found where a thesaurus about authors’ keywords was not just mentioned but also built and used.

4.2.2. Computation of the Number of Clusters

Figure 1 shows the dataset of the 87 keywords returned by VOSviewer, starting from the 12,148 documents, plus the thesaurus. The points are very dense. This statement is confirmed by the negative result of the test for finding outliers performed by calling the LocalOutlierFactor function, which is part of sklearn.neighbors.

The computation of the four methods recalled in Section 2 was performed using the Colab hosted Jupyter Notebook service that provides free-of-charge access to computing resources (namely: CPU, Python, and a rich pool of ML libraries). Preliminary, the elbow method was applied to the input dataset (Figure 1). Figure 2 shows a plot of WCSS against the number of clusters. The point on the plot where the WCSS curve starts to flatten suggests that $\mathcal{C}=5$ was the “optimal” number of clusters in our case. After this step, we made further measurements to consolidate the somehow subjective interpretation of the Elbow plot.

Table 5. Summary of the measurements for varying $\mathcal{T}$ values. (NoC = Number of Clusters; NoK = Number of Keywords; DB score = Davies–Bouldin score; Sscore = Silhoutte score; CH score = Calinski–Harabasz score).

$\mathcal{T}$	NoC	NoK	DB Score	Sscore	CH Score
5	11	87	2.428	0.034	25.090
10	9	73	2.299	0.075	26.140
15	7	63	1.774	0.138	27.104
20	7	56	1.188	0.184	29.540
25	6	50	1.184	0.170	22.042
50	5	33	0.949	0.216	22.951
60	4	29	0.695	0.305	23.381
65	4	27	0.761	0.279	20.508
70	3	25	0.818	0.401	26.500
75	3	22	0.896	0.340	16.611
100	3	15	1.338	0.089	5.621

reports the results. As the $\mathcal{T}$ value, we adopted 70. This value produced $\mathcal{C}=3$, which represented the best trade-off between the suggestion coming from the application for the three selected cluster validity indices and our judgment about a satisfactory partition of the input dataset (Section 2.2).

The next two subsections discuss, in order, the keyword co-occurrence network in the absence and presence of the thesaurus of authors’ keywords.

4.2.3. Analysis of the Network of Keyword Co-Occurrences Without a Thesaurus

Initially, the analysis of the keyword co-occurrences was performed without the thesaurus, as done in a large number of published articles. It has already been mentioned that by submitting the Scopus12148.txt file to VOSviewer and setting $\mathcal{T}=5$, the software returned 1096 keywords, of which 1089 were linked to each other. The 1089 keywords were aggregated into 16 distinct thematic clusters.

Table 6. The top five keywords of each cluster in the absence of a thesaurus of authors’ keywords. (NK = Number of keywords; Occ = Occurrences).

Cluster	NK	Top Five Keywords	Occ
1	160	traffic light control	183
		optimization	91
		smart city	86
		reinforcement learning	81
		machine learning	79
2	121	led lighting	156
		street lighting	139
		solar energy	91
		led driver	60
		power quality	59
3	112	light control	165
		solid-state lighting	32
		optogenetics	30
		energy efficient	26
		outdoor lighting	26
4		lighting system	238
		energy saving	208
		lighting control	135
		simulation	98
		energy consumption	90
5		leds	73
		light-emitting diodes	55
		photosynthesis	55
		light-emitting diode	54
		light quality	46
6		internet of things	174
		smart lighting	156
		iot	145
		microcontroller	79
		sensors	72
7		lighting	381
		light	78
		melatonin	32
		environment	31
		circadian rhythm	27
8		led	434
		energy management	49
		design	39
		tunnel lighting	35
		luminance	33
9		energy efficiency	389
		sustainability	74
		visual comfort	56
		artificial lighting	43
		energy audit	42
10		light emitting diode	60
		visible light communication	57
		intelligent lighting	54
		safety	32
		daylight harvesting	25
11		illumination	93
		energy	57
		efficiency	47
		building automation	43
		performance evaluation	22
12		sensor	46
		control system	40
		pwm	35
		neural network	31
		wireless communication	29
13		illuminance	88
		daylight	76
		light pollution	60
		dimming	45
		intelligent lighting system	33
14		lighting systems	148
		zigbee	78
		intelligent control	35
		artificial neural networks	18
		calibration	18
15		renewable energy	62
		public lighting	49
		light emitting diodes	43
		photovoltaic	37
		smart grid	33
16		oled	15
		light sources	14
		cct	7
		rare earths	7
		lifetime	5

shows the top five keywords of each cluster, according to the value of the VOSviewer Occurrences attribute. In [49], the authors also focused their discussion on the top five keywords. The table also shows, for each cluster, the number of involved keywords.

Table 7. The top 30 authors’ keywords in the absence of a thesaurus.

Keywords	Occurrences	Links
LED	434	401
energy efficiency	389	312
lighting	381	361
lighting system	238	231
energy saving	208	231
traffic light control	183	118
internet of things	174	207
light control	165	133
led lighting	156	170
smart lighting	156	173
lighting systems	148	150
Iot	145	186
street lighting	139	168
lighting control	135	164
simulation	98	122
illumination	93	122
optimization	91	129
solar energy	91	124
energy consumption	90	136
energy savings	88	125
illuminance	88	122
smart city	86	118
daylighting	85	107
reinforcement learning	81	76
machine learning	79	104
microcontroller	79	109
light	78	94
zigbee	78	103
image processing	77	88
daylight	76	104

shows the top 30 keywords (2.7%), selected according to the decreasing value of the Occurrences attribute; while Figure 3 shows the network of the thematic clusters returned by VOSviewer.

Interpretation of the 16 thematic clusters returned by VOSviewer was difficult, due to the simultaneous presence of synonyms, out-of-scope keywords, and too-generic keywords. Let us look, for instance, at Cluster 1 (Figure 4 and

Table 8. The 160 keywords in Cluster 1. (Occ = Occurrences).

Keywords	Occ	Keywords	Occ
traffic light control	183	signalized intersection	9
optimization	91	smart transportation	9
smart city	86	traffic signal	9
reinforcement learning	81	waiting time	9
machine learning	79	deep q-network	8
image processing	77	defect detection	8
fuzzy logic	66	intersections	8
deep learning	64	light sensors	8
wireless sensor network	57	smart light	8
artificial intelligence	47	smart traffic lights	8
deep reinforcement learning	47	traffic network	8
traffic control	46	cellular automata	7
intelligent transportation system	42	convolutional neural networks	7
machine vision	42	deep neural networks	7
traffic light	42	deep reinforcement learning (drl)	7
computer vision	41	fog computing	7
traffic congestion	37	fuzzy controller	7
traffic lights control	36	green wave	7
traffic signal control	32	image acquisition	7
fuzzy control	31	intelligent traffic control	7
genetic algorithm	29	intelligent traffic light	7
sumo	28	intelligent transport system	7
plc	27	intelligent transportation	7
traffic management	26	python	7
intelligent transportation systems	25	requirements engineering	7
traffic lights	25	smart streetlight	7
adaptive control	21	smart traffic light	7
artificial neural network	21	traffic light control systems	7
fpga	21	transfer learning	7
object detection	21	vhdl	7
vanet	21	visual inspection	7
digital twin	20	conflict resolution	6
neural networks	20	data fusion	6
q-learning	20	deep q-learning	6
traffic light control system	20	markov decision process	6
wsn	20	multiple intersections	6
adaptive traffic light control	19	pedestrian detection	6
traffic flow	19	predictive maintenance	6
congestion	17	real-time	6
fuzzy logic controller	17	real-time control	6
intersection	17	reflection	6
optimal control	17	scada	6
rfid	17	signal control	6
component	16	single intersection	6
its	16	smart traffic	6
multi-agent system	16	traffic control systems	6
particle swarm optimization	16	traffic monitoring	6
traffic	16	urban traffic	6
ambient intelligence	15	urban traffic control	6
embedded system	15	urbanization	6
intelligent transport systems	14	vehicles	6
intelligent transportation system	14	adaptive systems	5
multi-agent	14	adaptive traffic signal control	5
traffic simulation	14	algorithms	5
edge computing	13	arduino microcontroller	5
traffic light controller	13	bi-level optimization	5
transportation	13	building energy consumption	5
big data	12	context-aware	5
camera	12	distributed systems	5
multi-objective optimization	12	dynamic control	5
adaptive	11	emergency vehicle	5
cloud computing	11	emergency vehicles	5
embedded systems	11	image segmentation	5
intelligent systems	11	intelligent traffic system	5
intelligent traffic	11	internet of vehicles	5
motion detection	11	model	5
multi-agent reinforcement learning	11	multi agent system	5
traffic optimization	11	negotiation	5
vehicle detection	11	privacy	5
yolo	11	real-time systems	5
genetic algorithms	10	reinforcement learning (rl)	5
intelligent traffic light control	10	road traffic congestion	5
multi-agent systems	10	simulation of urban mobility	5
opencv	10	smart street lights	5
autonomous vehicles	9	style	5
clustering	9	traffic light control (tlc)	5
edge detection	9	urban intersections	5
petri nets	9	urban mobility	5
quality control	9	vehicle density	5
scheduling	9	vehicular networks	5

). The following remarks can be made: Cluster 1 is composed of 160 heterogeneous keywords. Examining Table 8, it emerges that Cluster 1 is largely focused on traffic control and traffic light control. This is testified by the following elements:

-

the top keyword is “traffic light control” (oc = 183);

-

14 keywords are synonyms of the “traffic light control” keyword (Table 9);

-

19 keywords are synonyms of the “traffic control” keyword (Table 9).

Table 9. Some synonyms of keywords in Cluster 1.

Synonyms of …	Keyword	Synonyms of …	Keyword
adaptive traffic light control	Traffic light control	congestion	Traffic control
intelligent traffic light		intelligent traffic
intelligent traffic light control		pedestrian detection
intelligent traffic system		road traffic congestion
smart traffic light		simulation of urban mobility (SUMO)
smart traffic lights		smart traffic
traffic light		SUMO
traffic light control		traffic congestion
traffic light control (tlc)		traffic control
traffic light control system		traffic control systems
traffic light control systems		traffic flow
traffic lights		traffic management
traffic lights control		traffic monitoring
traffic signal control		traffic network
		traffic optimization
		traffic simulation
		urban traffic
		urban traffic control

Table 9. Some synonyms of keywords in Cluster 1.

Synonyms of …	Keyword	Synonyms of …	Keyword
adaptive traffic light control	Traffic light control	congestion	Traffic control
intelligent traffic light		intelligent traffic
intelligent traffic light control		pedestrian detection
intelligent traffic system		road traffic congestion
smart traffic light		simulation of urban mobility (SUMO)
smart traffic lights		smart traffic
traffic light		SUMO
traffic light control		traffic congestion
traffic light control (tlc)		traffic control
traffic light control system		traffic control systems
traffic light control systems		traffic flow
traffic lights		traffic management
traffic lights control		traffic monitoring
traffic signal control		traffic network
		traffic optimization
		traffic simulation
		urban traffic
		urban traffic control

In addition, Cluster 1 contains keywords addressing transportation either directly (“intelligent transportation systems”, “intelligent transportation system”, “intelligent transport systems”, “intelligent transportation system”, “its”, “transportation”, “intelligent transportation”, “smart transportation”, and “vehicles”) or indirectly (“urban intersections”, “urban mobility”, “vehicle density”, “vehicular networks”, “camera”, “context-aware”, “vehicle detection”, “image acquisition”, and “image segmentation”). There was a further issue that made the interpretation of the thematic clusters returned by VOSviewer difficult in the absence of a thesaurus of keywords. In this latter case, it frequently happened that synonyms belonged to distinct thematic clusters.

Table 10. Examples of synonyms belonging to distinct thematic clusters.

Synonym Keywords	Cluster
lighting system	4
lighting systems	14
led	8
leds	5
led lighting	2
light-emitting diode	5
light-emitting diodes	5
light emitting diode	10
light emitting diodes	15

shows a few examples. Last but not least, it is worth noting that most of the keywords mentioned above are outside the scope of the present BA that focused on LSs. In summary, Cluster 1 raises more questions than it answers.

4.2.4. Analysis of the Network of the Keyword Co-Occurrences in Presence of the Thesaurus

Using VOSviewer, a keyword co-occurrence network was built starting from the Scopus12148.csv file, while providing the software with the Thesaurus.txt file as input. This time, $\mathcal{T}=70$ (see Section 4.2.2). VOSviewer returned a network composed of 25 keywords. The 25 keywords were partitioned into three distinct thematic clusters (Figure 5). The study by Triantafyllopoulos et al. [37] adopted a thesaurus of authors’ keywords that was used together with the Elbow method to decide the number of thematic clusters. Incidentally, in that study, the authors also obtained $\mathcal{C}=3$ clusters. As in all networks, keyword networks comprise nodes and links between pairs. Next, the nodes are analyzed.

• Node analysis

Table 11. The 25 keywords returned by VOSviewer for $\mathcal{T}=70$, sorted by cluster number. (TLS = Total link strength; Occ = Occurrences).

Keywords	Cluster	Links	TLS	Occ
led lighting system	1	22	593	1282
lighting control system	1	24	556	862
energy management	1	24	481	522
daylight	1	20	292	372
illumination control	1	19	187	222
user comfort	1	16	135	130
public lighting system	1	15	99	113
sustainability	1	13	52	88
artificial lighting	1	12	58	72
dimming control	1	13	74	71
iot	2	24	364	382
smart lighting control system	2	22	292	310
street lighting control system	2	22	189	236
smart building	2	20	165	166
wsn	2	16	153	146
fuzzy control	2	19	96	120
arduino	2	14	95	95
mcu	2	17	79	95
zigbee	2	16	108	89
smart city	3	18	146	148
computer vision	3	13	59	101
image processing	3	15	56	95
rl	3	9	30	86
ml	3	14	54	80
dl	3	15	53	71

and

Table 12. The 25 keywords returned by VOSviewer for $\mathcal{T}=70$, sorted by Occ value. (Occ = Occurrences).

Keywords	Occ	Keywords	Occ
led lighting system	1282	public lighting system	113
lighting control system	862	computer vision	101
energy management	522	arduino	95
iot	382	mcu	95
daylight	372	image processing	95
smart lighting control system	310	zigbee	89
street lighting control system	236	sustainability	88
illumination control	222	rl	86
smart building	166	ml	80
smart city	148	artificial lighting	72
wsn	146	dimming control	71
user comfort	130	dl	71
fuzzy control	120

list the 25 keywords returned by VOSviewer for $\mathcal{T}=70$. In the first table, the rows are sorted by cluster number and, within each cluster, by decreasing value of the Occurrences attribute; in the second table, the 25 rows are sorted by decreasing value of the Occurrences attribute.

The following is devoted to analyzing the three clusters composing the keyword co-occurrence network (Figure 5) returned by VOSviewer in the presence of the thesaurus. The three clusters are represented in different colors (red, green, and blue, respectively) based on the link strength between the keywords. Figure 6 shows only the keywords/nodes of the network to better figure out the boundary of the clusters.

The keywords inside each cluster are arranged in a logical way to form statements that capture the essence of the research and scope of the clusters. This approach is becoming popular. Two examples of recent studies are [25,26]. The former study conducted a bibliometric investigation of the state of the art regarding entrepreneurship and religions; while the latter reported on the co-occurrence of keywords in the domain of customer engagement. In 2024, this an approach was called “sensemaking” [50]. Carrying out the sensemaking step enables developing a preliminary understanding of the field’s intellectual structure, highlighting the connection of various research topics, before delving into a reading of the full papers (i.e., before starting a systematic literature review). It has been remarked ([3,25]) that keyword co-occurrence analysis reflects the present of the research domain under investigation. In this specific case, through the present BA, this review aimed to answer RQ1 and RQ2 (Section 3).

• Cluster 1 (Table 11, Figure 5 and Figure 7)

Figure 7. Cluster 1 for $\mathcal{T}=70$.

Figure 7. Cluster 1 for $\mathcal{T}=70$.

The “led lighting system” keyword stood out in Cluster 1, since its value for the Occurrences VOSviewer attribute was the highest among the 25 keywords (oc = 1282). This assertion is confirmed by the findings in [2,51], where the authors stated that LSs are predominantly LED-based today. Moreover, from the cluster, it was reveled that lots of research has been devoted to “lighting control system(s)” (oc = 862) with the aim of reducing energy consumption by mixing “daylight” with “artificial lighting” to make the public LSs of cities more sustainable, ([52,53]) increasing user comfort at the same time [51].

• Cluster 2 (Table 11, Figure 5 and Figure 8)

Figure 8. Cluster 2 for $\mathcal{T}=70$.

Figure 8. Cluster 2 for $\mathcal{T}=70$.

The nine keywords that compose Cluster 2 tell us that, so far, a lot of papers have investigated the adoption of IoT technology for the development of smart LSs, both for outdoors (the public domain) and indoors (i.e., the smart building domain). Refs. [2,51] confirm this a claim. The fifth keyword in the ranking (i.e., “WSN”) finds an explanation in the review by Gupta et al. [54], where the authors clarified that IoT technology, when incorporated with WSNs, extends its communication capabilities. The adoption of fuzzy logic for smart lighting control (i.e., “fuzzy control”) dates to 2000, when Wu et al. [55] proposed a fuzzy-logic-driven converter for photovoltaic-powered LS applications. More recently, Mohandas et al. [56] investigated a fuzzy logic controller finalized to control the light luminance level in street LSs based on human presence and light condition. Lighting sensors were used as input. The solution combined fuzzy logic with the ANN model. The seventh and eighth keywords refer to the hardware largely used in IoT-based LSs, while “zigbee” is one of the most widely used communication protocols, since it enables low-cost and low-power wireless IoT networks.

• Cluster 3 (Table 11, Figure 5 and Figure 9)

Figure 9. Cluster 3 for $\mathcal{T}=70$.

Figure 9. Cluster 3 for $\mathcal{T}=70$.

The six keywords in Cluster 3 tell us that to implement future smart cities, both academia and industry are investigating strategies to leverage AI methods (i.e., “computer vision”, “image processing”, and “ML”). This claim is in line with the findings in [2,51].

In light of the above considerations, the following it can be said: The reduced number of keywords and their thematic cohesion allow us to identify, for each of the three clusters, the basic theme that unites the research described by the authors’ keywords in the cluster. Cluster 1 is characterized by research adopting LED-based smart LSs, both for outdoor and indoor applications. Cluster 2 collects studies that propose solutions leveraging IoT technology. Lastly, the third cluster emphasizes the role of Artificial Intelligence in the implementation of smart cities. As a synthesis, it can be said that, from the BA, it emerges that most of the research efforts describe ways to deploy sustainable smart LSs in our cities by combining three technologies, namely LED, IoT, and AI (Figure 10). This finding is in line with the conclusions of two recent systematic literature reviews [2,51].

• Link analysis

After commenting on the nodes of the network of authors’ keyword co-occurrences, it is worth spending a few words on the links among them.

Table 13. Number of links among the 25 keywords.

Keywords	Links	Keyword	Links
lighting control system	24 (100%)	user comfort	16 (66.7%)
energy management	24 (100%)	zigbee	16 (66.7%)
IoT	24 (100%)	public lighting system	15 (62.5%)
led lighting system	22 (91.7%)	image processing	15 (62.5%)
smart lighting control system	22 (91.7%)	DL	15 (62.5%)
street lighting control system	22 (91.7%)	Arduino	14 (58.3%)
daylight	20 (83.3%)	ML	14 (58.3%)
smart building	20 (83.3%)	computer vision	13 (54.2%)
illumination control	19 (79.2%)	sustainability	13 (54.2%)
fuzzy control	19 (79.2%)	dimming control	13 (54.2%)
smart city	18 (75.0%)	artificial lighting	12 (50.0%)
MCU	17 (70.8%)	RL	9 (37.5%)

lists the 25 keywords returned by VOSviewer for $\mathcal{T}=70$, in decreasing order of the value of the Links attribute. From graph theory, we know that a graph with 25 nodes can comprise at most 300 edges (25 × 24/2). From Table 13, it emerges that the total number of links among the 25 keywords is 420. After removing repetitions, the number of distinct links becomes 210. Therefore, the degree of completeness of the keyword co-occurrence network is 70%. This means that, on average, each keyword co-occurs with the 70% of the remaining keywords. Of course, there are keywords above the average and keywords below the average. The following considers the nine top keywords in Table 13:

“Lighting control system”, “energy management”, and “IoT” are linked with all the remaining 24 keywords, hence they are also linked to each other. This tells us that all the articles returned by the Scopus engine (and mentioning these three keywords) adopted an IoT-based architecture to control lights, with the final aim of managing their energy consumption. The next three keywords are linked to 91.7% of the remaining keywords. This value reaffirms that almost all the selected studies referred to LED-based smart street lighting control systems. The next three keywords affirm that, in about 80% of instances, “illumination control” included daylight in order to implement “smart buildings”. An examination of the bottom side of Table 13 tells us that, so far, the number of studies proposing the adoption of AI/ML is not optimal. Accentuating research in which IoT and AI are the pillars of future LSs is a priority, both for industry and academia, to speed up the development of sustainable cities.

• Temporal evolution of the research topics

Below, we complete the discussion of the results (returned by the analysis of the co-occurrences of authors’ keywords) by examining the temporal evolution of the research topics in the articles about LSs. This information comes to light by reading column Avg.pub.year in

Table 14. The average publication year of the articles included in the BA (AVG = Avg.pub.Year).

Keyword	AVG	Keyword	AVG
image processing	2015.5158	street lighting control system	2017.9873
wsn	2015.9795	public lighting system	2018.4159
lighting control system	2016.2193	artificial lighting	2018.5139
dimming control	2016.2394	smart lighting control system	2018.5613
illumination control	2016.3649	smart building	2018.6265
daylight	2016.5108	smart city	2020.0405
computer vision	2016.7030	sustainability	2020.1477
user comfort	2016.9538	iot	2020.5288
zigbee	2016.9551	arduino	2020.7263
led lighting system	2016.9633	rl	2021.1047
energy management	2017.1494	ml	2021.3750
mcu	2017.4421	dl	2022.0704
fuzzy control	2017.6500

, where the values of the Avg.pub.year score attribute are sorted in ascending order. As described in Section 2.1, the generic value of this attribute denotes the average publication year of the articles in which the keyword on the same row of Table 14 occurs, with yellow indicating terms that mainly occur in recent publications. Therefore, for example, it can be observed that the research topics on lighting control systems initially focused on the use of LEDs (Avg.pub.year = 2016) and also making use of “daylight” (Avg.pub.year = 2016), with the primary intent of containing the energy consumption of such systems (“energy management”; Avg.pub.year = 2017). Then, research turned towards the production of “street lighting control systems” (Avg.pub.year = 2017), “smart buildings” (Avg.pub.year = 2018), and “smart cities” (Avg.pub.year = 2020). The use of the IoT was consolidated in the same period (Avg.pub.year = 2020). Regarding the presence of AI in the LS domain, the overlay visualization (Figure 11) shows that “image processing” (Avg.pub.year = 2015) and “computer vision” (Avg.pub.year = 2016) were investigated first, while ML (Avg.pub.year = 2021), RL (Avg.pub.year = 2021), and DL (Avg.pub.year = 2022) came later.

Table 15. Link strength of the “lighting control system” keyword with the remaining 24 keywords.

Keyword	Link Strength	Keyword	Link Strength
LED Lighting system	129	Zigbee	10
Energy management	99	MCU	9
daylight	45	Sustainability	8
illumination control	44	arduino	8
smart building	33	computer vision	8
user comfort	29	dimming control	8
smart lighting control system	26	artificial lighting	7
IoT	25	RL	3
WSN	24	smart city	3
fuzzy control	11	ML	2
image processing	11	street lighting control system	2
public lighting system	10	DL	2

shows the link strength of the “lighting control system” keyword (which represents the heart of the investigation domain of the present work) with the other 24 keywords. The strength of a link indicates the number of publications in which two keywords occur together (Section 2.1). In this case, one of the two keywords was “lighting control system”. The rows of Table 15 are sorted by decreasing values of the VOSviewer Link strength attribute. A comment restricted to the first three rows of the table follows: In the extant literature about LSs, there is a consolidated connection between the “lighting control system” keyword and LED technology to limit energy consumption. Such a goal has been largely investigated by combining artificial light with “daylight”.

4.2.5. Comparison of the Two Paths

Hereafter, Path 1 denotes the approach without the thesaurus, while Path 2 is the approach with the thesaurus. The comparison of the two paths with respect to the number of thematic clusters and the number of keywords that composed the thematic network returned by VOSviewer is given as follows: 16 clusters (Path 1) vs. 3 clusters; and 1089 keywords (Path 1) vs. 25 keywords. The strong reduction in the number of keywords in Path 2 with respect to Path 1 differently impacted on the values of the Links and Occurrences VOSviewer attributes (

Table 16. Comparison of the 20 top authors’ keywords in the two paths. (Occ = Occurrences).

Without Thesaurus	Links	Occ	With Thesaurus	Links	Occ
led	401	434	led lighting system	70	1282
energy efficiency	312	389	lighting control system	69	862
lighting	361	381	energy management	58	522
lighting system	231	238	iot	63	382
energy saving	231	208	daylight	42	372
traffic light control	118	183	smart lighting control system	57	310
internet of things	207	174	street lighting control system	45	236
light control	133	165	illumination control	34	222
led lighting	170	156	smart building	44	166
smart lighting	173	156	smart city	42	148
lighting systems	150	148	wsn	34	146
iot	186	145	user comfort	33	130
street lighting	168	139	fuzzy control	35	120
lighting control	164	135	public lighting system	25	113
simulation	122	98	computer vision	24	101
illumination	122	93	image processing	23	95
optimization	129	91	arduino	27	95
solar energy	124	91	mcu	36	95
energy consumption	136	90	zigbee	35	89
energy savings	125	88	sustainability	16	88

). Indeed, in the second case compared to the first, the value of the Links attribute dropped considerably, while the value of the Occurrences attribute increased significantly. The explanation of these numbers is given in the previous two sub-sections.

5. Limitations

The present BA suffers from two limitations. Firstly, the conclusions were shaped by the bibliometric data retrieved from the Scopus database that it relied upon. The reasons behind such a choice have been explained in Section 3 and supported by prior works; nevertheless, the possibility of missing out LS articles that may have been published in locations not indexed by Scopus cannot be excluded. The other limitation came from restricting the BA to the co-occurrences of the authors’ keywords. The latter technique, by definition, misses information about future research directions. This shortcoming is not relevant from the perspective of the stakeholders of the present study, while it is often an integral part of BAs.

6. Conclusions

The Scopus database was queried to reveal the research on LSs within the 1995–2024 time frame. A total of 12,148 items were retrieved. To map this huge production of knowledge, a co-occurrence analysis of the authors’ keywords was conducted. Three thematic clusters were returned by VOSviewer. Their interpretation allowed answering the following two research questions.

RQ1. What are the major topics explored by scholars in connection with LSs in the 1995–2024 time frame?

RQ2. How do they group together?

Inspired by previous research, this review also aimed to answer a third relevant research question:

RQ3. What is the impact of a thesaurus of authors’ keywords and the $\mathcal{T}$ value on clustering effectiveness?

The summary of the key findings of the present work follows:

• (Answer to RQ1) The trade-off between the numerical value of four independent methods belonging to the family of clustering validity approaches and the authors’ evaluation of a meaningful partition of the input dataset resulted in three independent clusters, collecting 25 authors’ keywords. The most relevant topics in Cluster 1 were “led lighting system” and “energy management”; while the most relevant topics in Cluster 2 were “IoT”, “smart building”, “street light control system”, “WSN”, “Arduino”, and “MCU”. Finally, the most relevant topics in Cluster 3 were “smart city, “computer vision”, “image processing”, “ML”, “DL”, and “RL”.
• (Answer to RQ2) Most of the research about LSs within the 1995–2024 time frame was devoted to proposing ways to deploy sustainable smart LSs by combining LED technology with the IoT and AI. This finding is in line with the conclusions of two recent systematic literature reviews [2,51].
• (Answer to RQ3) It was proven that, in absence of a thesaurus of authors’ keywords, the interpretation of the thematic clusters returned by VOSviewer became difficult due to the simultaneous presence of synonyms, out of scope keywords, and too-generic keywords. The only way to prevent this situation consisted in making use of the thesaurus. Regarding the choice of the $\mathcal{T}$ value, this work implemented recommendations from previous research in the field of clustering validity approaches.
• Regarding the construction of the thesaurus, it is worth noting that this is an iterative process. The factors to be taken into account in the workflow are the following: It is fundamental to keep the $\mathcal{T}$ value low, in order to include a large number of documents. In the present review, the authors started from 12,148 documents retrieved from Scopus. Here, the issue that must be addressed is making the realization of a thesaurus manageable, since the task is not fully automatic. In the present study, $\mathcal{T}$ was set to 5. By setting $\mathcal{T}$ = 3, the number of keywords to be considered in the construction of the thesaurus rose from 1096 to 4693, which made the manual work much more tedious and error-prone.