You are on page 1of 29

Scientometrics

https://doi.org/10.1007/s11192-019-03181-1

Territorial innovation models: to be or not to be, that’s


the question

David Doloreux1 · Jose Gaviria de la Puerta2 · Iker Pastor‑López2 ·


Igone Porto Gómez3   · Borja Sanz2 · Jon Mikel Zabala‑Iturriagagoitia4

Received: 21 November 2018


© Akadémiai Kiadó, Budapest, Hungary 2019

Abstract
Industrial agglomerations are key in explaining the development paths followed by territo-
ries, particularly at sub-national levels. This field of research has received increasing atten-
tion in the last decades, what has led to the emergence of a variety of models intended to
characterize innovation at the regional level. Moulaert and Sekia (Reg Stud 37:289–302,
2003) introduced the concept of ‘Territorial Innovation Models’ (TIMs) as a generic name
that embraced these conceptual models of regional innovation in the literature. However,
the literature does not help to assess the extent to which convergence or divergence is found
across TIMs. In this paper we aim to clarify if there are clear boundaries across TIMs, so
each TIM has particular characteristics that make it conceptually different from others, and
hence, justify its introduction in the literature. Based on natural language processing meth-
odologies, we extract the key terms of a large volume of academic papers published in peer
review journals indexed in the Web of Science for the following TIMS: industrial districts,
innovative milieu, learning regions, clusters, regional innovation systems, local production
systems and new industrial spaces. We resort to Rapid Automatic Keyword Extraction to
identify the associations between the topics extracted from the previous corpus. Finally, a
configuration to visualise the results of the methodology followed is also proposed. Our
results evidence that the previous models do not have a unique flavour but are rather similar
in their taste. We evidence that there is quite little that is truly new in the different TIMs in
terms of theory-building and the concepts being used in each model.

Keywords  Territorial innovation models · Bibliometric analysis · Natural language


processing · Regional development

* Igone Porto Gómez


igone.porto@deusto.es
1
Department of International Business, HEC Montréal, Montreal, QC, Canada
2
Faculty of Engineering, University of Deusto, Avenida de las Universidades, 24, 48007 Bilbao,
Spain
3
Deusto Business School, University of Deusto, Avenida de las Universidades, 24, 48007 Bilbao,
Spain
4
Deusto Business School, University of Deusto, Camino de Mundaiz 50,
20012 Donostia‑San Sebastián, Spain

13
Vol.:(0123456789)
Scientometrics

Introduction

Over the last two decades, scholarly work has enhanced our understanding of how inno-
vation matters for regions and how sub-national levels matter for innovation (Asheim
et al. 2016). The interest to better understand the role of regions as vital elements in sup-
porting innovation dynamics lies in innovation being an important element in economic
development.
Theory and research on innovation at the regional level have been analysed under dif-
ferent conceptual lenses. These approaches encompass a wide range of territorial innova-
tion models, different types of regional industrial agglomerations, a considerable literature
on geography studies and in the broader stream of evolutionary and institutional econom-
ics, also the literature on innovation studies. Common to these multiple streams of work,
scholars have devoted attention to how regions build their competitive advantages through
localised learning procedures (Asheim et al. 2011). Processes of innovation of firms and
sectors, as well as the importance of geographical proximity for knowledge exchange
and interactive learning as well as regions’ governance processes are at the centre of this
research (Asheim et al. 2016).
Originally introduced by Moulaert and Sekia (2003), the term ‘territorial innovation
models’ (TIMs) embodies different types of regional industrial agglomerations, includ-
ing concepts such as milieux innovateur, industrial districts, regional innovation systems,
new industrial spaces, local production systems and learning regions. Each of these territo-
rial models highlights different mechanisms underpinning dynamism and innovativeness
whereby regions provide or generate the conditions necessary to support innovation. For
instance, innovation milieu emphasises the importance of local culture and trust for knowl-
edge exchange (Crevoisier 2004); regional innovation systems puts emphasis on local
institutions and culture (Doloreux and Porto Gómez 2017); industrial districts draw upon
the idea of Marshall and give importance to labour market and trust-based relationships
(Becattini 1979); learning regions emphasise regional capacities to promote and capture
knowledge and talented workers (Florida 1995); clusters emphasise the concentration of
interconnected firms and various support organizations as well as institutions and policies
that enable coordinated learning among key actors and improve capabilities in a spatial
context (Porter 1990).
As a result of their literature review, Moulaert and Sekia concluded that TIMs could be
characterised by a set of distinctive features: core of innovation dynamics, role of institu-
tions, regional development, culture, types of relations among agents and types of relations
with the environment (p. 295).1 Likewise, Scaringella and Radziwon (2018) considered
that the previous models had various common features: the territory, the values shared,
the stakeholders involved, the economic development, the society, the knowledge and the
outcome.
However, in spite of this unifying framework, Moulaert and Sekia also argued that TIMs
suffered from ‘conceptual ambiguity’ (p. 289) in explaining how regions shape innovation,
and argued that this could have clear implications for a coherent theory building, a claim
similar to that made by Miettinen (2002). Other authors claim that TIMs, whilst differ-
ent, are also connected (e.g. Olsen 2012; De Marchi and Grandinetti 2014; Scaringella and

1
 Similar categorizations have also other been provided by scholars such as O’Gorman and Kautonen
(2004), Moulaert and Nussbaumer (2005) or Cainelli (2008) to mention a few.

13
Scientometrics

Chanaron 2016). For instance, according to Tödtling and Trippl (2005, p. 1206), “clusters
are central elements of the knowledge application and exploitation subsystem, whilst the
[regional innovation system] RIS is a wider concept in the sense (1) that there are usually
several clusters and many industries in a RIS and (2) that institutions play a larger role”.
Dotti (2014) pointed out that TIMs have more similarities than differences, and that their
differences are mainly driven by the schools of thought in which each model has emerged.
In this sense, Dotti underlined the central role played by three schools in the advancement
of TIMs thinking: the Italian industrial district school, the North American school leading
the debate around Porter’s cluster, and the Nordic school, mainly centred on innovation
systems.
However, when reading the contributions that have and still are being published in the
field of TIMs, one can observe that the different TIMs are often put as synonyms and used
in an exchangeable manner (e.g. Muscio 2006). It is thus reasonable to believe that we are
embedded in a world of ‘fuzzy terms’ (Markusen 2003). These fuzzy ‘umbrella models’
are more dependent on the schools of thought in which they were defined rather than on
real differences in the TIMs themselves (Dotti 2014). They are charged with symbolism,
but usually lack theoretical content, and are present in the community as long as a particu-
lar fashion lasts (e.g. knowledge bases, institutional entrepreneurship, social innovation,
resilience, related variety). However, when a new fashion emerges, these trendy concepts
are left to oblivion in order to give way to new concepts that are present in the literature
albeit by different names and reinventing the wheel once more (e.g. systemic and trans-
formative innovation policies, entrepreneurial and innovation ecosystems).
The paper focuses on the following research questions: Are there clear boundaries across
the different TIMs? Does every TIM have particular characteristics that make it unique,
and hence conceptually different from others? To reach this goal, a bibliometric analysis
has been conducted on all the TIMs considered by Moulaert and Sekia (2003), including
the cluster model, based on co-word science mapping (Santos et al. 2012; Liu et al. 2015;
Bornmann et al. 2018) of articles published in peer-reviewed journals and indexed in the
Thomson Reuters Web of Science. This has led to an analysis of the full content of more
than 1540 articles published between the 1990s and 2017. One of the main benefits of
conducting a literature review is that it helps identify relationships among concepts (Jes-
son et  al. 2011; Hart 2014), and hence, the use of bibliometric methods such as natural
language processing seems adequate to respond to the research question posed in the paper.
We need to acknowledge that the use bibliometric methods is not new in the field of
territorial innovation (see García-Lillo et al. 2017). For example, Sun and Grimes (2016)
used bibliometric performance analysis to understand the evolution of national innovation
studies. Lee and Su (2010) also conducted a bibliometric analysis of regional innovation
systems. Both studies offered a static view of the articles published up to 2008, focusing on
the relationships between the researchers publishing those articles, their schools of thought
and the countries they belong to.
The paper contributes to the literature in various respects. This is the first paper where
the entire set of TIMs is compared. Besides, the time span considered in the analysis cov-
ers the entire period in which scientific contributions for every TIM are found. Secondly,
the paper relies on bibliometric methods that have not been used before in the literature
dealing with innovation at the sub-national level. This is particularly noteworthy, since the
application of bibliometric methods not only allows for the replicability of the methodol-
ogy, but it also facilitates the management of large datasets that emanate from the scientific
analysis of full texts, whilst reducing researchers’ subjective influence—an issue that is
often mentioned in the literature on systematic reviews (Tranfield et al. 2003). Finally, the

13
Scientometrics

methods used in the paper have helped us to identify the main characteristics of each TIM,
and determine the extent to which each TIM also referred to other models (or not), and if
so, what were the concepts and features that were being contemporaneously used by the
different TIMs.
The remaining of the paper is organised as follows. The next section provides the review
of the literature as regards the TIMs under analysis (i.e. clusters, industrial districts, learn-
ing regions, milieux innovateur, new industrial districts and regional innovation systems).
“Methodology” section explains the methodology followed in the bibliometric analysis and
introduces the data used in the study. “Results” section describes the underlying dimen-
sions behind every TIM, which in turn allows us to illustrate the boundaries, similarities
and differences that emerge from the comparison of all TIMs vis-à-vis. Finally, section five
concludes the paper.

Literature review

Few scholars would dispute Alfred Marshall’s (1890) account of industrial districts as one
of the essential foundations of current theory and research on TIMs. For Marshall, labour
market externalities represented a critical aspect of the ‘economies of specialisation’, given
the importance that he attributed to skilled labour as a key (regional) resource. It is clear
that he regarded the movement of labour from firm to firm, together with local processes
of skill development, as one of the primary vehicles for knowledge spill-overs. It is fair to
say that there is considerable continuity between Marshall’s seminal ideas on industrial
districts and ‘modern’ writing on TIMs.
However, it is only in the 1970s when western economies moved from a ‘Fordist indus-
trial model’ to a Post-Fordist economy that there is a resurgence of the region (Markusen
2003). At this time, we observed that some regions managed their transitions better than
others, often calling upon local resources and knowledge (Piore and Sabel 1984). This was
rapidly formalised into a series of similar ideas and (territorial) models to describe and
explain the processes whereby regions drew upon their internal resources and dynamics to
increase innovativeness. Since the early 1980s, a number of TIMs—many of them harking
back to Marshall’s (1890) idea—were born and have since attracted widespread interest
from scholars, policymakers and even the mass media (Cruz and Teixeira 2010).
This section describes the TIMs that have been considered in the paper. We hereby
focus on the TIMs considered by Moulaert and Sekia (2003), but also decide to include the
cluster model, since it is one the most popularised (if not the most) TIMs.

TIMs: a brief description

The concept of Territorial Innovation Models (TIMs) was introduced by Moulaert and
Sekia (2003) as a general label to connect all the terms that analysed regional development
dynamics in which innovation, knowledge transfer and public private partnerships sought
to explain regional competitiveness. Initially, their contribution did not include Porter’s
(1990) cluster model, which is an oversight given that clusters should be regarded as fun-
damental in the advancement of regional development, particularly due to their impact on
policy spheres around the globe (Uyarra and Ramlogan 2016). More recently, other schol-
ars have considered the cluster model in their comparative studies. For example, Olsen
(2012) performed an analysis of the policies adopted to promote innovation in regional

13
Scientometrics

innovation systems, clusters and learning regions. Scaringella and Chanaron (2016) stud-
ied the differences in the roles played by research infrastructures and universities in indus-
trial districts, innovative milieus, regional innovation systems, new industrial spaces and
regional clusters to assess their socioeconomic and entrepreneurial impact. Porto Gómez
et al. (2016) provided a qualitative measure of the governance model and the sectoral spe-
cialization promoted by various TIMs, including clusters. More recently, Asheim et  al.
(2016) showed that regional innovation systems partly draw on the concepts of industrial
districts and clusters to understand the mechanisms behind successful regional economies.
Despite all TIMs have the aim of boosting regional competitiveness, it is possible to assert
that each TIM has been promoted under different geographical, cultural and sectoral cir-
cumstances, which justify their creation and evolution.
Industrial districts emerged from the Italian Marshallian district concept (Marshall
1890, 1919), which explained the knowledge spill-over effect. Workers in a region tended
to have similar jobs and be employed in similar value chains, which ultimately contributed
to the informal spread of useful (intangible) ideas and knowledge (Ortega-Colomer et al.
2016). The industrial district approach has been (mainly) boosted by the Italian school,
which mostly dealt with case studies in Southern Europe (Becattini 1979; Becattini et al.
2009; Boschma and Ter Wal 2007).
With a similar understanding to that of the Italian industrial district and mainly pro-
moted by French scholars, the ‘milieux innovateur’ (Crevoisier 2004) focused on firms as
the key regional players in relation to innovation. Its emphasis was put on organisational
and relational knowledge (i.e. non-technical knowledge) and its central role in the devel-
opment of innovations. In this regard, Maillat (1998) underlined the role of geographical
proximity so knowledge ties could be established among regional actors. This was the first
model in which the relationships with and among institutional stakeholders were explic-
itly mentioned beyond value chains and interactions with knowledge-related infrastructures
(see Tremblay and Rousseau 2005).
Initially based on the innovation systems’ approach, the learning region concept (Florida
1995; Morgan 1997) focused on networking and on the interactive character of innovation,
which is in turn shaped by region-specific institutions and social rules. The idea behind the
learning region is that cooperation among all regional actors can contribute to a common
learning process, which benefits all societal actors. Local actors, hence, do not only con-
tribute to innovation development but also to the promotion of social capital. There is no
school that clearly espouses this model, although collective learning has influenced other
models (see Hassink 2005).
The cluster concept is defined as a geographically proximate group of interconnected
companies and associated institutions in a particular field of activity, who benefit from
commonalities and externalities (Porter 2008). The literature on clusters shows that the
concentration of industrial activity in a region influences firms’ performance because
local competition induces firms to innovate in order to remain competitive globally (Porter
2000). The concentration of various organizations in clusters allows them to interact, which
eases the flow of knowledge and leads to more face-to-face interaction, facilitating the
exchange of explicit and tacit knowledge (Wolfe and Gertler 2004; Isaksen 2009). A num-
ber of explanations have been provided on why firms benefit from geographical concen-
tration (Fang 2015), and the outcomes of such clustering: (1) productivity and efficiency;
(2) innovation and its implementation (through firm-level cooperation and collaboration);
(3) enterprise competitiveness; (4) the growth of existing enterprises and the creation and
survival of start-ups; and (5) regional growth and economic development by promoting job
creation, new ventures, better working and living conditions, and higher wages.

13
Scientometrics

Departing from the perspective of national innovation systems (Lundvall 1988), Cooke
et  al. (1997) introduced the notion of regional innovation system (RIS). Within the RIS
literature, regions are considered as meaningful spaces because they provide the appropri-
ate institutional framework to support innovation and cooperation among nearby economic
actors and institutions (Asheim and Gertler 2005). This approach understands innovation
as a socially and territorially embedded process of interaction between firms and their
environment (Barrutia and Zabala-Iturriagagoitia 2018). RIS has been mainly empirically
studied in the Nordic Countries and across various regions in Europe (Doloreux and Porto
Gómez 2017).
The concept of new industrial spaces (Storper and Scott 1988) departs from the indus-
trial district model in that it argues that social capital plays a significant role in territorial
cohesion. The underlying logic behind this concept is that the community of stakeholders,
more than the territory, is the essential determinant of territorial dynamics. In this TIM, the
sectoral specificities of regions are justified by the (generally) homogeneous dedication of
productive firms and the ties established around value chains in which firms interact (Scott
1988). Examples of new industrial spaces include the British Motor Industry (Henry et al.
1996) or the initiatives to boost regional development in Norway during the 1970s and
1980s (Isaksen 1996). However, no school has promoted this approach specifically.
Finally, localized production systems (Courlet and Soulage 1994) adopt a more confined
approach than previous ones, as it is focused on municipalities. The rationale behind this
configuration is similar to that of industrial districts, as it argues that sectoral specificity
is due to the industrial expertise of firms in urban contexts. It should be noted that despite
Moulaert and Sekia categorized local production systems as a relevant TIM, the number of
articles that develop and apply this approach is very limited as to date (see Table 1).2

TIMs: a brief comparison

There are few studies that compare TIMs among them. Olsen (2012) compared the clus-
ter, regional innovation system and learning region models, and discussed how the region
might not only be an active agent but also an object. The learning region and the RIS
would contend the region as an agent, while clusters would embrace both considerations.
As Olsen (2012, p. 1791) discusses “the region as an “active agent” is referred to when
the TIM in question refers to regional policy-makers as highly independent in the develop-
ment of public policies, while the region is referred to as an “object” in cases where the
administrative region is considered to be mainly subject to higher level decision making”.
This differentiation would turn into contrasting policies adopted in each case to push sub-
national innovation and competitiveness.
Dotti (2014) is also among the few who have analysed innovation support policies from
the perspective of three different TIMs: the dominant US model of clusters, the Nordic
model of innovation systems, and the Italian model of industrial districts. Furthermore,
Dotti also studied the relation that the latter two TIMs could have with the learning region
concept. His conclusion was that these TIMs have several similarities involving concepts
that are being equally treated, whereby innovation is regarded as a collective issue, and
knowledge sharing is key to boost the competitiveness and innovation of the firms in a
region.

2
  We used both the English ’innovative milieu’ and the French ’milieux innovateur’ as search keywords.

13
Scientometrics

Table 1  Territorial innovation models under analysis. Source: own elaboration


TIMs # Of articles in the WOS # Of articles included in the % Of articles included in the # Of meta-keywords resulting from
(a) analysis (b) analysis (c) the articles analysed in each model (d)

Clusters 1716 989 57.6 3410


Regional innovation systems 221 197 89.1 1182
Learning region 190 170 89.5 679
Industrial district 149 126 84.6 514
Milieux innovateurs 38 34 89.5 173
New industrial spaces 17 15 88.2 102
Local production systems 13 9 69.2 64
Total number of articles 2307 1540 66.75 6124

13
Scientometrics

Based on a narrative literature review of the theoretical streams behind each TIM,3
Porto Gómez (2014) provided a taxonomy where she considered the sectoral specialisa-
tion, the institutional role and the participation of regional stakeholders as the three main
features setting the boundaries that explain the differences across TIMs. For example, local
production systems and industrial districts are centred on firm involvement and participa-
tion in territorial dynamics. However, from the perspective of sectoral specialization, a
blurred line is found between industrial districts, local production systems, clusters and
regional innovation systems. Furthermore, policies and governance are mainly driven by
competitive poles and regional innovation systems, which are more permeable to a strong
role of policy circles. In this regard, while the ‘French’ competitive poles are mainly driven
by a top-down logic, regional innovation systems are more prone to bottom-up rationales
(Crescenzi and Rodríguez-Pose 2011). Therefore, the existing literature does not help to
assess the extent to which convergence or divergence is found across TIMs. This paper
aims to contribute to this particular gap.
The previous view was shared by Ortega-Colomer et al. (2016), who argued that clus-
ter-related ideas and concepts were substituting the founding principles of industrial dis-
tricts. They concluded that the concepts behind these TIMs were being misused, not only
among scholars, but also among policy makers. To provide additional clarification, they
highlighted the differences between these two models, from their origins, to the reasons for
their co-existence and the type of policies that could be concluded thereof.
To sum up, in spite of the extant contributions in the literature regarding the similarities
and differences between TIMs, there is no study tackling this analysis in a systematic and
robust manner. To address this gap, the paper applies bibliometric methods that allow for
an analysis of the full text of all papers that have been published as to date on each TIM in
order to identify the similarities and/or differences among them.

Methodology

Natural language processing is a part of artificial intelligence methods, which encompass a


wide variety of topics involving the computational processing and understanding of human
languages. Historically, this field has required a high deal of manual work to extract rules
of human language (Lounsbury et al. 1979). In the last years though, the use of machine
learning techniques has reduced this manual work and improved the performance of these
systems (Chen et al. 2017). In our case, we turned into natural language processing (Onan
et al. 2016) to contrast the content of the articles published in peer review journals indexed
in the Web of Science, and that focused on the following TIMs: clusters, regional inno-
vation systems, learning regions, industrial districts, milieux innovateurs, new industrial
spaces and local production systems.
As mentioned in the introduction, bibliometric techniques and natural language pro-
cessing have already been used in the literature on regional development. For example,
Hervas-Oliver et al. (2015) aimed to clarify the emerging sub-fields in clusters and indus-
trial districts. Similar analyses have also been performed by Lee and Su (2010), Toivanen
and Ponomariov (2011), Uriona-Maldonado et al. (2012) or Liu et al. (2015) on regional

3
  In her analysis, Porto Gómez (2014) also included the competitive poles (i.e. Pôles de compétitivité) as a
TIM.

13
Scientometrics

Fig. 1  Description of the methodology. Source: own elaboration

innovation systems. These evidences let us confirm that natural language processing can be
regarded as a useful technique to establish differences among topics, as it is the case in this
paper.
Figure 1 below provides a graphical explanation of the procedure followed. The analysis
was based on six stages: (1) the paper-dataset to be analysed was selected and downloaded;
(2) all the text from the accessed articles was extracted and the parts of the manuscript that
were not accessible as text (e.g. graphs) were removed; (3) using the Rapid Automatic Key-
word Extraction (RAKE) algorithm (Rose et al. 2010) the extraction of the keywords from
each article was performed. These keywords were not the ones identified by the author(s)
in each manuscript (usually in the title page), but rather the words that were identified as
being central to the understanding of the manuscript according to the RAKE algorithm; (4)
the previous keywords were classified according to categories identified by Moulaert and
Sekia (2003) and were labelled as ‘meta-keywords’. This stage was necessary to adapt the
results of the present paper to the largest extent possible to those by Moulaert and Sekia in
their seminal contribution; (5) the relations between these meta-keywords were established
and (6) a visualisation of the results was created to facilitate the understanding of the rela-
tionships among TIMs.
The previous methodology allowed us to conduct a systemic and systematic analysis
of the literature. Some of the steps (i.e. steps two and four) were performed manually, as
it was necessary to be acquainted with the literature on TIMs, so the keywords derived
from the bibliometric analysis could be coherently grouped into the categories identified by
Moulaert and Sekia (2003). In this respect, tools like SciMat (Cobo et al. 2012), facilitated
the manual processing. This is certainly one of the aspects that requires further exploration
in the future so as to reduce the potential bias that this manual and subjective allocation

13
Scientometrics

has on the final results. The remaining steps (i.e. steps three, four and five) were automated
through computer algorithms.4 The following paragraphs will describe the details of each
stage.

Articles included in the analysis

The criteria to assess the eligibility of the papers to be included in the study were the
following:

1. They had to deal with a specific TIM as the core analytical concept under which the
empirical investigation was carried out;
2. They had to be published before 2018 in a peer-reviewed journal indexed in the Web of
Science.5 Other publication forms (e.g. conference proceedings, books, book chapters,
working papers, etc.) and platforms (e.g. Scopus, PROQUEST, Google Scholar) were
not considered, and
3. Both empirical and/or conceptual papers could be included.

The search query on the Thomson Reuters Web of Science was performed on February
28th 2018, with December 31st 2017 chosen as the last date considered in the search. The
previous search returned 2307 articles for the entire set of TIMs (see Table 1, column a).
However, not all of these articles were used for the bibliometric process (see column b).
Two of the authors of the present paper went through this initial screening and chose the
articles that could be of relevance for the purposes of this paper. For example, those papers
that did not deal with the analysis of innovation from a sub-national perspective or those
not explicitly taking into account any of the above TIMs were excluded from the analysis.
The data cleaning also included the removal of those papers for which the full text could
not be extracted. The final number of papers considered in the analysis amount to 1540.

Text extraction

For this second stage, an automatic tool for the extraction of the text was developed,6 which
was applied to each of the previous 1540 papers. In this step, the text was not pre-processed
(i.e. filtering symbols, page numbers, title headings, etc.) and the bulk of words identified
in each paper was considered. This automatic tool allowed for a large number of pdf files to
be processed in a fast, automatic and homogeneous manner.

4
  The details of the procedure followed in each of these algorithms are provided in “Topic extraction”,
“Topic aggregation” and “Topic association” sections below.
5
  The choice of the Web of Science as the main source of data is due to the quality and peer review process
of the journals included in it (Archambault et  al. 2009; Bergman 2012; Amara and Landry 2012). Since
2004, Google Scholar or Scopus have also emerged as relevant sources of scientific information. However,
since many of the articles in the TIMs literature date back to the 1980s and 1990s, we have resorted to the
Web of Science.
6
  This tool was developed for the European Project DANTE. DANTE (Detecting and analysing terrorist-
related online contents and financing activities) aims to deliver more effective, efficient, automated data
mining and analytics solutions, and an integrated system to detect, retrieve, collect and analyse huge
amount of heterogeneous and complex multimedia and multi-language terrorist-related contents, from both
the Surface and the Deep Web, including Dark nets. See: http://www.h2020​-dante​.eu/.

13
Scientometrics

Topic extraction

The aim of this third stage was to extract the key words from each paper. Once these
keywords are identified, then it will be possible to check whether any relationship exists
among them. To achieve this goal, each paper was considered as a collection of ‘topics’.
The most repeated topics in a paper would be indicative of the main themes tackled in each
paper. An automatic algorithm was required to recollect these key topics into what was
coined as ‘keywords’. In recent years, the literature has singled out RAKE as an interesting
tool to extract keywords from unique documents (Thushara et al. 2017 since it extracts the
keywords of a document by parsing the text, and not by considering the typical stop words.
To identify the keywords in each paper RAKE uses some parameters to refine the search
detection. The first of these parameters is that any word with the potential to characterise
the contents of any paper needs to have a minimum of three letters. The second param-
eter is the length of the sentence that the word is included, this value is selected because
papers often resort to passive and subjunctive grammatical structures, so the longitude of
the sentences is usually high. Finally, the last parameter points to the number of minimum
repetitions that a certain word must have in the entire paper, so the keyword becomes rep-
resentative of the topic under analysis inside the document. In the context of this paper, this
threshold was set in six repetitions. As a result of this phase, 6124 keywords were obtained
from the 1540 papers analysed (see Table 1, column d).7

Topic aggregation

In this step, a new classification was created which groups into ‘meta-keywords’ the dif-
ferent keywords of the texts identified in the previous stage. These meta-keywords were
created based on the categories provided by Moulaert and Sekia (2003), namely, core of
innovation dynamics, role of institutions, regional development, culture, types of relations
among agents and types of relations with the environment. To these key categories of inno-
vation, we added the dimensions of governance, theory, methodology and TIMs, given the
recent contributions that have been carried out in relation to the political and governance
dimension of innovation at the regional level (e.g. Doloreux and Porto Gómez 2017).8
Governance is regarded as an important element due to the significance of institutional
settings and policies adopted to support territorial competitiveness based on innovation
(Tödtling and Trippl 2005; Isaksen and Trippl 2017). Secondly, theory was included as a
new dimension given that the differences among TIMs may be due to the theoretical posi-
tioning of the main schools (Dotti 2014). Methodology was also considered to capture the
different procedures and approaches used in each of the TIMs, particularly regarding the
empirical contributions. Finally, the TIM dimension was included as a control capture the
potential cross-references across TIMs.
After converting the keywords into meta-keywords, the number of meta-keywords in
each TIM and their relative weight as compared to the total number of meta-keywords was
created (see “Appendix 1” for further details). The most relevant meta-keywords were: core
of innovation dynamics, regional development, role of institutions and relationships among

7
 “Appendix 1” provides the details for the keywords of each TIM.
8
  The Tables showing how the original keywords extracted from the text of the pdfs for each TIM corre-
spond to the meta-keywords identified in Table 1 are available upon request.

13
Scientometrics

agents. Further on, new topics that are recently appearing in the literature are identified,
such as culture, governance linked to the policies adopted and the type of relations with the
environment.

Topic association

Once the previous information was pre-processed, the Apriori algorithm (Hahsler and
Chelluboina 2011) was used to extract the most frequent meta-keywords. This algorithm is
applied in two steps, the FP-Growth and the Association Rules. The FP-Growth step allows
to obtain the set of rules included in the pre-processed information. This method was effi-
cient and scalable for mining both long and short frequent patterns. The algorithm can be
described as a recursive elimination scheme: a pre-processing step deletes all items from
the relationships that are not frequent individually (i.e. do not appear in a user-specified
minimum number of transactions), and defined by the user as a parameter of the algorithm
called minimum support value (Borgelt 2005). 0.15 was chosen as the minimum threshold,
which means that the relationships not repeated more than 15% were considered as not rep-
resentative and discarded (Elgaml et al. 2015). The FP-Growth tree (Hong et al. 2013) was
constructed as an output of this, and a tuple of {keywords, frequency} was obtained, which
was used as the input for the next step.
Secondly, the Association Rules (Agrawal et al. 1993) was run to find relationship pat-
terns among the meta-keywords. This algorithm provides the strength of the relationships
between two meta-keywords, which range between 0 and 1, 1 being the highest possible
strength.
Frequency(Keyword A ∪ Keyword B)
Confidence(Keyword A, Keyword B) = (1)
Frequency(Keyword A)
The previous algorithms have been implemented in R9 using the software package
arulesViz.

Visualisation

To compare our results with those of Moulaert and Sekia (2003), graphical illustrations
of the meta-keywords were created.10 Circular graphs were chosen as the most allur-
ing method, as they allow presenting in a clear and concise way the relationships among
meta-keywords.
In these circular graphs, every meta-keyword has a particular colour, and receives a
scale from which different arrows depart. For example, in the case of the cluster model (see
Fig.  2), the meta-keyword role of institutions includes both the relationships that depart
from this meta-keyword to others (scale 0–4, inner part of the graph), and the relationships
that go to this meta-keyword departing from others (scale 4–12, outer part of the graph).
The relationships that depart from the meta-keyword role of institutions receive the same
colour as that meta-keyword, while the relationships that go to this meta-keyword receive
the colour of the respective meta-keywords they depart from.

9
 See https​://www.r-proje​ct.org/.
10
  The graphical representations of the keywords associated to each TIM are available upon request.

13
Scientometrics

Fig. 2  Relationships among meta-keywords in the cluster literature. Source: own elaboration

Table 2  Colour relationship TIM Key letter Colour


legend between TIMs and key
letter for visualization. Source:
Core of innovation dynamics A Black
own elaboration
Culture B Yellow
Governance—politics C Purple
Methodologies D Pink
Regional development E Red
Role of institutions F Grey
Theories G Blue
TIM H Green
Type of relations with the environment I Brown
Types of relations among agents J Orange

13
Scientometrics

For a good understanding of the figures and a clear reading of the texts in the images,
we have proceeded to make a relationship between the TIMs and a key letter for each one
and a specific colour. This relation is specified in Table 2, below:

Results

The following sub-sections will evidence the results achieved with the application of the
previous methodology to each of the TIMs considered in the paper.

Clusters

After the topic extraction phase and the subsequent division of keywords into the meta-
keywords, 3410 meta-keywords were obtained for the cluster model. The most representa-
tive meta-keywords in this TIM were core of innovation dynamics, regional development,
role of institutions and types of relationships among actors, while the meta-keywords of
TIMs, methodologies, theories, types of relationships with the environment and govern-
ance were less central.
It is particularly worth emphasising the strength of the relationships between the core
of innovation dynamics and the types of relationships among actors, which take place bi-
directionally. It should also be ntoed that both meta-keywords are the most significant and
central to this TIM (i.e. largest number of repetitions). The same bi-directionality can also
be observed between the role of institutions and the types of relationships among actors,
and between the role of institutions and the core of innovation dynamics.
The first observation that can be achieved as a result of this circular graph is that the
meta-keywords theories, TIMs, types of relations with the environment, culture, and meth-
odologies only have relationships from them to the other groups, but do not receive any
relationships. In turn, the remaining meta-keywords (i.e. types of relations among agents,
core of innovation dynamics, governance, regional development and role of institutions)
receive relationships from other meta-keywords, but also dispatch their own relationships
to other meta-keywords. Hence, it can be concluded that the latter are particularly central
when explaining the inner characteristics of the cluster model.
The second observation is that when the papers dealing with the cluster model made a
reference to contributions of other TIMs, these cross-references were mainly used to high-
light such features as the types of relations among agents, the core of innovation dynam-
ics, regional development and the role of institutions. However, it is worth stressing how
the clusters model does not convey to other TIMs to draw any type of policy recommen-
dations (i.e. no relationships between the TIM and governance meta-keywords). A very
similar conclusion is also found regarding the relationships of the culture meta-keyword,
as it shows relationships with other meta-keywords, but not with the one on TIMs. It is
also noteworthy observing how culture shows no relationships with the theories on which
the cluster model is usually based (i.e. culture does not play any role in the cluster model).
Finally, it also becomes clear how the cluster model does not have among its goals the
analysis of the territory, as most other TIMs do. This is evidenced in the figure below by
the fact that the meta-keyword types of relationships with the environment has relation-
ships with many of the other meta-keywords, but in turn, it does not receive any links from
the others.

13
Scientometrics

Fig. 3  Relationships among meta-keywords in the regional innovation systems literature. Source: own elab-
oration

Regional innovation systems

In the case of regional innovation systems, 1182 meta-keywords were identified. The first
observation is that the meta-keywords types of relations with the environment, culture and
methodologies did not receive any relationships from the other meta-keywords (see Fig. 3).
Second, it is remarkable that the meta-keyword theory does not emerge as a significant one
from the analysis. This is quite surprising, given the innovation systems framework has
clear institutional and evolutionary roots, and several scholars in the economic geography
stream have contributed to the systems thinking in the last decades.
As regards to the relationship with other TIMs, the regional innovation systems model
is probably the one in which more references are made to other TIMs. These cross-refer-
ences cover basically all the dimensions identified by Moulaert and Sekia (2003), with the
only meta-keywords that are related to the TIMs feature being methodology and culture.
The core of innovation dynamics, regional development and the role of institutions are the

13
Scientometrics

Fig. 4  Relationships among meta-keywords in the learning region literature. Source: own elaboration

meta-keywords that have the largest relevance in the case of the TIMs, as they are the ones
that dispatch and receive the most links with the remaining meta-keywords. Also note that
the meta-keyword types of relations among agents is also quite relevant. In the opposite
extreme, it is possible to find the dimensions of culture and methodology which would act
as the most trivial ones.
It is also important to emphasise the relationships that the meta-keyword governance-
politics shows with the remaining meta-keywords, and in particular the strength of the link
between governance and the role of institutions. This should come as no surprise, since the
regional innovation systems model has, since its origins, claimed that sub-national institu-
tions are the ones with the largest influence on the policies defined in relation to territorial
dynamics based on innovation.

13
Scientometrics

Learning region

After analysing the literature on the learning region model, this TIM could be character-
ised by 679 meta-keywords. In this case, the dimensions of types of relations with the
environment, governance, methodology and TIMs did not receive any relationships from
the other meta-keywords (see Fig. 4). Out of them it is particularly worth emphasising the
limited relevance assigned to the types of relations with the environment and to regional
development.
As discussed in “Literature review” section, the learning region model aims, a priori, to
explain innovation dynamics as a consequence of networking activities (i.e. networking for
learning and learning for innovation) and other social specific attributes. However, when
the full text of the papers dealing with the learning region model are analysed, this conclu-
sion is not supported, despite the dimension of types of relations among actors emerges as
a central one in this TIM.
When analysing the cross-references of the literature on the learning region model with
other TIMs, these mainly refer to the role of institutions, regional development and the core
of innovation dynamics. However, here too, as it was the case in the clusters model, there
are no cross-references to any aspects related to governance or policy implications. This
lack of policy influence of the learning region model may be partly explained by the fact
that in it, economic dynamics are understood as a path-dependent process where learning,
rather than policy intervention, would be the main driver of firm and territorial diversifica-
tion and expansion (e.g. Hassink 2005).
The most central meta-keyword in this model is the role of institutions. Since learn-
ing is a function determined to a great extent by cultural and social factors, it comes as
no surprise that the meta-keyword role of institutions receives relationships from all the
remaining meta-keywords, and it is necessary for all the other determinants to be related
to institutions. This becomes very perceptible in that even the meta-keyword methodology
only has a relationship established with the role of institutions.

Industrial district

The industrial district model was characterised by 514 meta-keywords. From all the TIMs
addressed in this paper, the industrial district model arguably shows the most balanced pat-
tern across meta-keywords. Methodologies, governance and culture are the three meta-key-
words that do not receive any links from the remaining meta-keywords (See Fig. 5). The
remaining meta-keywords are, however, comprehensively related among themselves. First
of all, the literature on industrial districts provides a significant number of cross-references
to other TIMs, particularly regarding their implications in terms of the role of institutions,
core of innovation dynamics and types of relations among agents.
Despite the origins of the industrial district model go back to Marshall (1890), it is
worth observing how the theoretical contributions of the papers in this TIM are miss-
ing after analysing their content through the application of the methodology explained in
“Methodology” section. Similarly, as it was also the case for some of the TIMs discussed
above, here too the policy implications that the industrial district model shows on the gov-
ernance-policy dimension is quite confined.
One of the characteristics of the industrial district concept is that of culture, as this
feature is very much dependent on the sectoral specialisation of a territory. However, the

13
Scientometrics

Fig. 5  Relationships among meta-keywords in the industrial district literature. Source: own elaboration

results from the bibliometric analysis show how the culture meta-keyword establishes rela-
tionships with other features such as core of innovation dynamics, role of institutions, TIMs
or the types of relations among actors, but does not receive any links itself. This reduces its
relative relevance in explaining the contents of the literature on this TIM. The other unex-
pected result that comes from the analysis, at least if one sticks to what the contributions on
industrial districts seem to suggest, is that for territories to be specialised around industries,
they should be aware of the value chains that those industries could contribute to or be part
of. In other words, it seems necessary that territories gain awareness of the characteris-
tics of the environments that could be affecting them (Barrutia and Zabala-Iturriagagoitia
2018). However, the meta-keyword interactions with the environment does not result as a
relevant dimension from the analysis, pointing to a certain paradox in the interpretation of
the results that emanate from this TIM.

13
Scientometrics

Fig. 6  Relationships among meta-keywords in the milieux innovateurs literature. Source: own elaboration

Milieux innovateur

After the initial screening of the papers in the literature and their succeeding election, the
milieux innovateurs (34 papers in the analysis), new industrial spaces (15 papers) and local
production systems (nine papers) are the three TIMs that were the least representative. This
implies that the results of the bibliometric analysis conducted should be taken with caution,
since the number of meta-keywords identified as significant is much lower than in the previous
TIMs. For example, in the particular case of the local production systems, the application of
the methodology indicated in “Methodology” section led to the identification of 64 meta-key-
words. Given the limited number of these meta-keywords, the circular graphs in the case of the
local production systems does not provide any sensible and meaningful information about their
underlying characteristics, and this is the reason why they have not been included in the paper.
In the case of the milieux innovateurs, these are characterised by 173 meta-keywords.
Out of the features identified by Moulaert and Sekia (2003), the one dealing with cul-
ture can be regarded as the least significant as it shows relationships with some of the

13
Scientometrics

Fig. 7  Relationships among meta-keywords in the new industrial spaces literature. Source: own elaboration

meta-keywords but it receives no links from them. A similar conclusion can be reached
for governance-politics, which shows more uni-directional relationships with the remaining
meta-keywords than the one dealing with culture (see Fig. 6). It is also significant that the
features of theory and interactions with the environment are not considered in this TIM. On
the other hand, the features of regional development and the role of institutions have the
most central meta-keywords when explaining the content of the contributions published in
the literature regarding the milieux innovateur model.
In terms of the relationship that milieux innovateurs show with other TIMs, this is
mainly established to explain the role of institutions and the core of innovation dynamics.
However, these results may be partially biased by the limited number of observations that
are found in the literature.

New industrial spaces

The last of the TIMs reported in this paper is the one dealing with the so-called new industrial
spaces. In this case, the characterisation of the contributions that can be found in the literature

13
Scientometrics

to date number 102 meta-keywords. Out of the meta-keywords that emerge from the analysis
as the relevant ones to explain the literature on new industrial spaces, theories, TIMs, govern-
ance and culture emerge as the least central ones as they do not receive any links from the
remaining meta-keywords despite establishing relationships with them (See Fig. 7). On the
other hand, the dimensions of regional development, role of institutions, the types of relations
among agents and the core of innovation dynamics emerge as the most significant.
According to the description made in “Literature review” section, the strength of the con-
cept of new industrial spaces lies in social capital which is claimed to play a central role in
explaining territorial cohesion. As a result, it should be expected that the dimension of culture
would keep strong relationships with the dimensions related to innovation and the interactions
among actors. However, when the results of Fig. 7 are observed, it becomes clear that for the
meta-keyword culture, the core of innovation dynamics are important, but the opposite rela-
tionship does not exist: culture does not emerge as an important factor for innovation dynam-
ics. Besides, the culture meta-keyword does not show any relationships with the interactions
among actors, which a priori would be a sensible result that is not supported by the evidence.
In relation to the references that the new industrial spaces model makes to other related
TIMs, these cross-references are established to explain the relationships among actors, the
role of institutions, regional development and the core of innovation dynamics. Finally, the
sectoral specificities of regions are justified by the (generally) homogeneous dedication of
productive firms and the ties established around value chains in which those firms interact.

Relationships across TIMs

After explaining the underlying concepts portraying each TIM, it is possible to conclude
that all the models embrace the meta-keywords or features identified by Moulaert and
Sekia (2003) in a similar manner. TIMs seem thus to be more similar than the scholarly
contributions seem to suggest, without clear conceptual differences among them. Figure 8
provides evidence of the relationships among the meta-keywords across TIMs. With it we
aim to be able to respond to the research question guiding this paper, and hence conclude
whether there are relationships across the entire set of TIMs.
The results show that the meta-keywords of theories, TIMs, types of relations with the
environment, culture and methodologies are the least representative, as they only have uni-
directional relationships with the remaining meta-keywords. This conclusion is very illuminat-
ing, as it points to the fact that theoretical and methodological contributions are scant across
TIMs. In addition, it can also be concluded that cross-references across TIMs are missing, and
that the analysis of the cultural dimension, which could eventually be included as part of the
role of institutions and which is fundamental in explaining territorial dynamics, is also rather
limited across TIMs. The present paper thus confirms the arguments posed in the introductory
section of the paper: many of the keywords most commonly used in the literature on TIMs
may be charged with symbolism, but lack theoretical content and methodological rigour.
The concept of TIMs was introduced to group and characterise those approaches intended
to explain the development of innovation from a territorial perspective at sub-national levels.
As a result, it seems reasonable and sensible to observe how the meta-keyword core of inno-
vation dynamics has strong relationships with regional development, while at the same time
also receives relationships from all the other meta-keywords. Innovation being a systemic and
interactive phenomenon, it would be expected that the meta-keyword types of relationships
among agents would also have a central role in most TIMs. This is clearly observed in Fig. 8,

13
Scientometrics

Fig. 8  Relationships among meta-keywords across TIMs. Source: own elaboration

as it establishes and receives relationships with most meta-keywords. Finally, the political
implications of most TIMs also become visible in the analysis, as the meta-keyword of gov-
ernance and politics also receives and establishes relationships with the other meta-keywords.

Discussion and conclusions

Several scholars have reasserted the importance and contribution of regional industrial
agglomerations in explaining the development paths followed by territories, particularly at
the regional level. Given the large scale and scope of this field of research, Moulaert and
Sekia (2003) introduced the concept of ‘Territorial Innovation Models’ as a generic name
that embraced those conceptual models of regional innovation in the literature, in which
local institutional dynamics played a significant role (p. 291). This field of research has
received increasing attention in the last decades, what has led to the emergence of a variety
of approaches and models intended to characterize innovation at sub-national levels.

13
Scientometrics

The paper aims to clarify if there are indeed clear boundaries across TIMs, so each TIM
has particular characteristics that make it conceptually different from others, and hence,
justify its introduction in the literature, or whether all TIMs are actually using the same
concepts and rely on the same theoretical foundations. The paper presents the results of a
bibliometric analysis on TIMs that aims to make a conceptual contribution to the literature
on regional industrial agglomerations, regional development and regional innovation.
To reach this goal we have resorted to natural language processing techniques, which allow
us to extract the full corpus of all articles in the literature. The first advantage of this novel
methodology is that it is suitable for those cases in which the content of a big number of docu-
ments is to be analysed. Second, it does not only optimise the ‘time to process’ the extraction
of keywords and the categorization of a large amount of documents, but it also allows the
researcher to control the exploration process and adapt this categorization to the particularities
of the research question. As a result of the meta-keywords that emanate from the application of
the previous methodology, we develop a set of circular graphs that allow determining not only
the relationships across these meta-keywords, but also the strengths of those relationships.
Accordingly, we are able to determine the meta-keywords that are more central to explaining
the contributions that can be found in the literature concerning every TIM and the relation-
ships among them, and hence respond to the research question that guides the paper.
Our results point in two main directions that may put into question the contributions
found to date in the literature. On the one hand, many TIMs argue that they are specialized
into particular features, such as industrial districts focusing on sectoral specialization, the
learning region focusing on social issues in order to foster networking activities leading
to innovation, or the RIS considering the role of institutions to mention a few. However,
when the contributions found in the literature are studied, and their content is approached
from the application of bibliometric methods, many of these a priori positions cannot be
sustained vis-à-vis our results. In other words, these models do not have a unique flavour
but are rather similar in their taste. As a consequence, we in fact know rather little about
the way in which these models differ both in terms of theorizations and integration into our
thinking of regional industrial agglomerations. Likewise, even if one may accept that these
models are indeed different, the way in which they are theorised and the features that dif-
ferentiate them from one to another remain vague and fuzzy.
The schools of thought (and the scholars) behind every TIM are quite clear on what they
mean by such concepts as ‘innovation, ‘regional development’, ‘culture’, ‘governance’, ‘pol-
icy’ and ‘institutions’, and have their own motivations, rationales and interests for undertaking
research on TIMs. But at a wider scale, dialogue between researchers across schools of thought,
TIMs and even geographical locations is often stymied, as it is often difficult to acknowledge
the limitations of one’s own (TIM) approach or to recognise that other, equally valid (TIM)
approaches, exist and may come to complement the conclusions of a particular discipline.
The evidences shown in the article could, at least, pave the way for developing a richer theory
of TIMs. The paper provides the first comprehensive attempt to make sense of TIMs in terms of
scientific research. Based on the above, the paper has demonstrated that there is quite little that is
truly new in the different TIMs in terms of theory-building. This raises a concern about the fact
that, sadly to said, the scientific and research content of the TIMs is superficial and unorganised.
Is this result something temporary or could we also fall in the same traps in the future?
In recent times we have witnessed how the term ‘ecosystems’ has been increasingly incor-
porated into the discourse and the analyses of the members of the innovation studies com-
munity (Mazzucato and Robinson 2018; Horváth and Rabetino 2019). Can we expect it to
provide something new to the community, or are we going to continue making the same
(old) mistakes by introducing a new concept that brings no added value? Similar arguments

13
Scientometrics

could also be posed as regards the literature on innovation management for example, with
the emergence of the social innovation concept, the adoption of open innovation models
(Enkel et al. 2009) or in the case of the literature on innovation policy with the emergence
of transformative innovation policies (Schot and Steinmueller 2018).
The paper also presents some limitations. Firstly, the data was gathered from the Thom-
son Reuters Web of Science, which includes a good coverage of the peer-reviewed jour-
nals that target the literature on TIMs. Including other publication forms (e.g. conference
proceedings, books, book chapters, working papers, etc.) and platforms (e.g. Scopus, PRO-
QUEST, Google Scholar) could provide a more complete and nuanced picture of the stat-
of-the art. Despite we believe that considering other publication forms may lead to similar
conclusions, the choice of the Web of Science implies certain limitations that should be
considered. This database is strongly over-representing articles published in journals led
by English speaking countries and excluding some disciplines (Archambault et  al. 2006;
Chavarro et al. 2018). The use of new translation techniques allows for the potential con-
sideration of works published in other languages. As regards the methodology followed, we
have two human-dependent processes that may influence the results. First, as to date, it is not
possible to remove all (subjective) human participation, since we are the ones outlining the
articles to be considered in the research process and those to be excluded, due to our famili-
arity with the content and the breadth of the conversations taking place in the discipline.
Second, the fact that the actual technology used in the paper implies the programming of the
code to apply natural language processing means that the way keywords and meta-keywords
have been obtained and aggregated is also subject to the actual manually-programmed code.
The analysis pursued in the paper has also helped us to identify potential knowledge gaps
that could be addressed in further research endeavours. First, the results evidence that the
different streams of the literature on TIMs have a clear lack of theoretical origin and contri-
bution, as well as a clear limitation in terms of the methodologies applied in them. This calls
for a critical redefinition of the field of TIMs, demanding more sound theoretical contribu-
tions and the utilization of more robust and sophisticated methodologies that move beyond
the qualitative and quantitative (mainly parametric) methods traditionally applied. Second,
we need to better explain the reasons why there is little divergence among TIMs. This calls
for a more thorough investigation of the key questions guiding the research in the differ-
ent TIMs by examining more comprehensively the geographical and social-political context
from which each TIM emanates. A debate could be opened as to the extent to which TIMs
are context-dependent or otherwise. Finally, we would like to foresee the future trends to be
developed in each model. We believe that the approach followed in this paper constitutes a
good starting point tackle such question. The use of conceptual evolution maps could help
identify the most central and embryonic themes in each model, and hence, identify if com-
mon or divergent patterns also emerge in this contemporaneous research.

Funding  The funding was provided by Eusko Jaurlaritza (Grand No. IT885-16) and H2020 Societal Chal-
lenges (Grand No. H2020-700367).

Appendix

See Table 3.11

11
  The meta-keywords highlighted in bold below represent those identified as key distinctive features of
TIMs by Moulaert and Sekia (2003). The remaining are the meta-keywords we have added to complement
their analysis.

13
Scientometrics

Table 3  Relative weight of the meta-keywords in each of the TIMs considered. Source: own elaboration
Clusters Regional inno- Learning region Industrial districts Milieux innovateur New industrial spaces Local
vation systems production
systems

Core of innovation dynamics 594 (17.4%) 257 (21.7%) 101 (14.9%) 92 (17.9%) 37 (21.7%) 9 (8.8%) 12 (18.7%)
Culture 87 (2.5%) 17 (1.4%) 23 (3.4%) 22 (4.3%) 5 (2.9%) 4 (3.9%) 1 (1.6%)
Regional development 702 (20.6%) 138 (11.7%) 137 (20.2%) 74 (14.4%) 25 (14.7%) 30 (29.4%) 8 (12.5%)
Role of institutions 539 (15.8%) 239 (20.2%) 118 (17.4%) 87 (16.9%) 35 (20.2%) 13 (12.7%) 10 (15.6%)
Type of relations among agents 514 (15.1%) 164 (13.9%) 81 (11.9%) 95 (18.5%) 26 (15%) 20 (19.6%) 14 (21.9%)
Type of relations with the environment 192 (5.6%) 49 (4.1%) 29 (4.3%) 16 (3.1%) 3 (1./%) 6 (6%) 1 (1.6%)
Governance—politics 310 (9.1%) 192 (16.2%) 88 (12.9%) 46 (8.9%) 7 (4%) 6 (5.9%) 4 (6.2%)
Methodologies 188 (5.5%) 55 (4.6%) 33 (4.8%) 25 (4.9%) 8 (4.6%) 2 (1.9%) 7 (10.9%)
Theories 27 (0.8%) 6 (0.5%) 33 (4.8%) 12 (2.3%) 4 (2.3%) 7 (6.9%) –
TIM 257 (7.5%) 65 (5.5) 36 (5.3%) 45 (8.8%) 23 (13.3%) 5 (4.9%) 7 (10.9%)
Total 3410 (100%) 1182 (100%) 679 (100%) 514 (100%) 173 (100%) 102 (100%) 64 (100%)

13
Scientometrics

References
Agrawal, R., Imieliński, T., & Swami, A. (1993). Mining association rules between sets of items in large
databases. In ACM sigmod record (Vol. 22, No. 2, pp. 207–216). ACM, New York.
Amara, N., & Landry, R. (2012). Counting citations in the field of business and management: Why use
Google Scholar rather than the Web of Science. Scientometrics, 93(3), 553–581.
Archambault, É., Campbell, D., Gingras, Y., & Larivière, V. (2009). Comparing bibliometric statistics
obtained from the Web of Science and Scopus. Journal of the American Society for Information Sci-
ence and Technology, 60(7), 1320–1326.
Archambault, É., Vignola-Gagné, É., Côté, G., Larivire, V., & Gingrasb, Y. (2006). Benchmarking scientific
output in the social sciences and humanities: The limits of existing databases. Scientometrics, 68(3),
329–342.
Asheim, B. T., Boschma, R., & Cooke, P. (2011). Constructing regional advantage: Platform policies based
on related variety and differentiated knowledge bases. Regional Studies, 45(7), 893–904.
Asheim, B. T., & Gertler, M. S. (2005). The geography of innovation: Regional innovation systems. In
J. Fagerberg, D. Mowery, & R. Nelson (Eds.), The Oxford handbook of innovation (pp. 291–317).
Oxford: Oxford University Press.
Asheim, B. T., Grillitsch, M., & Trippl, M. (2016). Regional innovation systems: Past–present–future. In R.
Shearmur, C. Carrincazeaux, & D. Doloreux (Eds.), Handbook on the geographies of innovations (pp.
45–62). Cheltenham: Edward Elgar.
Barrutia, J., & Zabala-Iturriagagoitia, J. M. (2018). Towards an epigenetics understanding of evolutionary
economic geography. Evolutionary and Institutional Economics Review, 15(2), 213–241.
Becattini, G. (1979). Dal ‘settore’ industriale al ‘distretto’ industriale. Alcune considerazioni sull’unità
d’indagine dell’economia industriale. Rivista di Economia e Politica Industriale, 5(1), 7–21.
Becattini, G., Bellandi, M., & De Propis, L. (2009). A handbook of industrial districts. Cheltenham: Edward
Elgar.
Bergman, E. M. L. (2012). Finding citations to social work literature: The relative benefits of using Web of
Science, Scopus, or Google Scholar. The Journal of Academic Librarianship, 38(6), 370–379.
Borgelt, C. (2005). An implementation of the FP-growth algorithm. In Proceedings of the 1st international
workshop on open source data mining: Frequent pattern mining implementations (pp. 1–5). Chicago,
Illinois, August 21, 2005, ACM (Association for Computing Machinery).
Bornmann, L., Haunschild, R., & Hug, S. E. (2018). Visualizing the context of citations referencing papers
published by Eugene Garfield: A new type of keyword co-occurrence analysis. Scientometrics, 114,
427–437.
Boschma, R. A., & Ter Wal, A. L. (2007). Knowledge networks and innovative performance in an industrial
district: The case of a footwear district in the South of Italy. Industry and Innovation, 14(2), 177–199.
Cainelli, G. (2008). Spatial agglomeration, technological innovations, and firm productivity: Evidence from
Italian industrial districts. Growth and Change, 39(3), 414–435.
Chavarro, D., Rafols, I., & Tang, P. (2018). To what extent is inclusion in the Web of Science an indicator of
journal ‘quality’? Research Evaluation, 27(2), 106–118.
Chen, B., Tsutsui, S., Ding, Y., & Ma, F. (2017). Understanding the topic evolution in a scientific domain:
An exploratory study for the field of information retrieval. Journal of Informetrics, 11(4), 1175–1189.
Cobo, M. J., López-Herrera, A. G., Herrera-Viedma, E., & Herrera, F. (2012). SciMAT: A new science
mapping analysis software tool. Journal of the American Society for Information Science and Technol-
ogy, 63(8), 1609–1630.
Cooke, P., Uranga, M. G., & Etxebarria, G. (1997). Regional innovation systems: Institutional and organisa-
tional dimensions. Research Policy, 26(4–5), 475–491.
Courlet, C., & Soulage, B. (1994). Dynamiques industrielles et territoire. In G. Garofoli & A. Vázquez-
Barquero (Eds.), Organization of production and territory: Local models of development (pp. 45–91).
Pavia: Gianni Luculano.
Crescenzi, R., & Rodríguez-Pose, A. (2011). Reconciling top-down and bottom-up development policies.
Environment and Planning A, 43(4), 773–780.
Crevoisier, O. (2004). The innovative milieus approach: Toward a territorialized understanding of the econ-
omy? Economic Geography, 80(4), 367–379.
Cruz, S. C., & Teixeira, A. A. (2010). The evolution of the cluster literature: Shedding light on the regional
studies–regional science debate. Regional Studies, 44(9), 1263–1288.
De Marchi, V., & Grandinetti, R. (2014). Industrial districts and the collapse of the Marshallian model:
Looking at the Italian experience. Competition & Change, 18(1), 70–87.

13
Scientometrics

Doloreux, D., & Porto Gómez, I. (2017). A review of (almost) 20  years of regional innovation systems
research. European Planning Studies, 25(3), 371–387.
Dotti, N. F. (2014). Literature review on territorial innovation models, geography of research and policy
innovations. GREATPI working paper no 1. Cosmopolis, centre for urban research, Department of
Geography, Vrije Universiteit, Brussel.
Elgaml, E. M., Ibrahim, D. M., & Sallam, E. A. (2015). Improved FP-growth algorithm with multiple mini-
mum supports using maximum constraints. International Journal of Computer, Electrical, Automa-
tion, Control and Information Engineering, 9(5), 1087–1094.
Enkel, E., Gassmann, O., & Chesbrough, H. (2009). Open R&D and open innovation: Exploring the phe-
nomenon. R&D Management, 39(4), 311–316.
Fang, L. (2015). Do clusters encourage innovation? A meta-analysis. Journal of Planning Literature, 30(3),
239–260.
Florida, R. (1995). Towards the learning region. Futures, 27(5), 527–536.
García-Lillo, F., Claver-Cortés, E., Marco-Lajara, B., Úbeda-García, M., & Seva-Larrosa, P. (2017). On
clusters and industrial districts: A literature review using bibliometrics methods, 2000–2015. Papers in
Regional Science, 97(4), 835–861.
Hahsler, M., & Chelluboina, S. (2011). Visualizing association rules: Introduction to the R-extension pack-
age arulesViz. R Project Module, 6, 223–238.
Hart, C. (2014). Doing a literature review. Releasing the social science research imagination (8th ed.). Lon-
don: SAGE Publications Ltd.
Hassink, R. (2005). How to unlock regional economies from path dependency? From learning region to
learning cluster. European Planning Studies, 13(4), 521–535.
Henry, N., Pinch, S., & Russell, S. (1996). In pole position? Untraded interdependencies, new industrial
spaces and the British motorsport industry. Area, 28(1), 25–36.
Hervas-Oliver, J. L., Gonzalez, G., Caja, P., & Sempere-Ripoll, F. (2015). Clusters and industrial districts:
Where is the literature going? Identifying emerging sub-fields of research. European Planning Studies,
23(9), 1827–1872.
Hong, S., Huaxuan, Z., Shiping, C., & Chunyan, H. (2013). The study of improved FP-growth algorithm in
MapReduce. In Proceedings of the the 1st international workshop on cloud computing and information
security (pp. 250–253) Nov 9–11, 2013, Shangai, China.
Horváth, K., & Rabetino, R. (2019). Knowledge-intensive territorial servitization: Regional driving forces
and the role of the entrepreneurial ecosystem. Regional Studies, 53(3), 330–340.
Isaksen, A. (1996). Towards increased regional specialisation? The quantitative importance of new
industrial spaces in Norway, 1970–1990. Norsk Geografisk Tidsskrift, 50, 113–123.
Isaksen, A. (2009). Innovation dynamics of global competitive regional clusters: The case of the Norwe-
gian centres of expertise. Regional Studies, 43(9), 1155–1166.
Isaksen, A., & Trippl, M. (2017). Exogenously led and policy-supported new path development in
peripheral regions: Analytical and synthetic routes. Economic Geography, 93(5), 436–457.
Jesson, J. K., Matheson, L., & Lacey, F. M. (2011). Doing your literature review. Traditional and sys-
tematic techniques. London: SAGE Publications Ltd.
Lee, P. C., & Su, H. N. (2010). Investigating the structure of regional innovation system research through
keyword co-occurrence and social network analysis. Innovation, 12(1), 26–40.
Liu, Z., Yin, Y., Liu, W., & Dunford, M. (2015). Visualizing the intellectual structure and evolution of
innovation systems research: A bibliometric analysis. Scientometrics, 103(1), 135–158.
Lounsbury, J. W., Roisum, K. G., Pokorny, L., Sills, A., & Meissen, G. J. (1979). An analysis of topic
areas and topic trends in theCommunity Mental Health Journal from 1965 through 1977. Commu-
nity Mental Health Journal, 15(4), 267–276.
Lundvall, B.-A. (1988). Innovation as an interactive process: From user–producer interaction to the
national system of innovation. In G. Dosi, C. Freeman, R. Nelson, G. Silverberg, & L. Soete (Eds.),
Technical change and economic theory (pp. 349–369). London: Pinter.
Maillat, D. (1998). From the industrial district to the innovative milieu: Contribution to an analysis of
territorialised productive organisations. Recherches Économiques de Louvain/Louvain Economic
Review, 64(1), 111–129.
Markusen, A. (2003). Fuzzy concepts, scanty evidence, policy distance: The case for rigour and policy
relevance in critical regional studies. Regional Studies, 37, 701–717.
Marshall, A. (1890). Principles of economics. London: MacMillan.
Marshall, A. (1919). Industry and trade. London: Macmillan.
Mazzucato, M., & Robinson, D. K. R. (2018). Co-creating and directing Innovation Ecosystems?
NASA’s changing approach to public-private partnerships in low-earth orbit. Technological Fore-
casting and Social Change, 136, 166–177.

13
Scientometrics

Miettinen, R. (2002). National innovation system: Scientific concept or political rhetoric. Helsinki: Edita
Prima Ltd.
Morgan, K. (1997). The learning region: Institutions, innovation and regional renewal. Regional Studies,
31, 491–503.
Moulaert, F., & Nussbaumer, J. (2005). The social region: Beyond the territorial dynamics of the learn-
ing economy. European Urban and Regional Studies, 12(1), 45–64.
Moulaert, F., & Sekia, F. (2003). Territorial innovation models: A critical survey. Regional Studies, 37,
289–302.
Muscio, A. (2006). From regional innovation systems to local innovation systems: Evidence from Italian
industrial districts. European Planning Studies, 14(6), 773–789.
O’Gorman, C., & Kautonen, M. (2004). Policies to promote new knowledge-intensive industrial agglom-
erations. Entrepreneurship & Regional Development, 16(6), 459–479.
Olsen, L. S. (2012). Territorial knowledge dynamics: Making a difference to territorial innovation mod-
els and public policy? European Planning Studies, 20(11), 1785–1801.
Onan, A., Korukoğlu, S., & Bulut, H. (2016). Ensemble of keyword extraction methods and classifiers in
text classification. Expert Systems with Applications, 57, 232–247.
Ortega-Colomer, F. J., Molina-Morales, F. X., & Fernández de Lucio, I. (2016). Discussing the concepts
of cluster and industrial district. Journal of Technology Management & Innovation, 11(2), 139–147.
Piore, M., & Sabel, C. (1984). The second industrial divide: Possibilities for prosperity. New York:
Basic Books.
Porter, M. E. (1990). The competitive advantage of nations. New York: Free Press.
Porter, M. E. (2000). Location, competition, and economic development: Local clusters in a global econ-
omy. Economic Development Quarterly, 14(1), 15–20.
Porter, M. E. (2008). Clusters, innovation, and competitiveness. In EU conference on innovation and
clusters, Stockholm.
Porto Gómez, I. (2014). ROSIS: Un Sistema Regional Sectorial y Abierto para la innovación en la
comarca del Duranguesado. PhD dissertation. Bilbao: University of the Basque Country.
Porto Gómez, I., Otegi Olaso, J. R., & Zabala-Iturriagagoitia, J. M. (2016). ROSA, ROSAE, ROSIS:
Modelling a regional open sectoral innovation system. Entrepreneurship & Regional Development,
28(1–2), 26–50.
Rose, S., Engel, D., Cramer, N., & Cowley, W. (2010). Automatic keyword extraction from individual
documents. Text Mining: Applications and Theory, 1, 1–20.
Santos, I., de-la-Peña-Sordo, J., Pastor-López, I., Galán-García, P., & Bringas, P. G. (2012). Automatic cat-
egorisation of comments in social news websites. Expert Systems with Applications, 39, 13417–13425.
Scaringella, L., & Chanaron, J. J. (2016). Grenoble–GIANT territorial innovation models: Are investments
in research infrastructures worthwhile? Technological Forecasting and Social Change, 112, 92–101.
Scaringella, L., & Radziwon, A. (2018). Innovation, entrepreneurial, knowledge, and business ecosystems:
Old wine in new bottles? Technological Forecasting and Social Change, 136, 59–87.
Schot, J., & Steinmueller, W. E. (2018). Three frames for innovation policy: R&D, systems of innovation
and transformative change. Research Policy, 47(9), 1554–1567.
Scott, A. J. (1988). Flexible production systems and regional development: The rise of new industrial spaces
in North America and Western Europe. International Journal of Urban and Regional Research, 12(2),
171–186.
Storper, M., & Scott, A. J. (1988). The geographical foundations and social regulation of flexible produc-
tion complexes. In J. Wolch & M. Dear (Eds.), The power of geography (pp. 21–40). London: Allen &
Unwin.
Sun, Y., & Grimes, S. (2016). The emerging dynamic structure of national innovation studies: A bibliomet-
ric analysis. Scientometrics, 106(1), 17–40.
Thushara, M. G., Krishnapriya, M. S., & Nair, S. S. (2017). A model for auto-tagging of research papers
based on keyphrase extraction methods. In 2017 International conference on advances in computing,
communications and informatics (ICACCI) (pp. 1695–1700), IEEE. Sept 13–16, Karnataka, India.
Tödtling, F., & Trippl, M. (2005). One size fits all?: Towards a differentiated regional innovation policy
approach. Research Policy, 34(8), 1203–1219.
Toivanen, H., & Ponomariov, B. (2011). African regional innovation systems: Bibliometric analysis of
research collaboration patterns 2005–2009. Scientometrics, 88(2), 471–493.
Tranfield, D., Denyer, D., & Smart, P. (2003). Toward a methodology for developing evidence-informed
management knowledge by means of systematic review. British Journal of Management, 14, 207–222.
Tremblay, D. G., & Rousseau, S. (2005). Le secteur du multimédia à Montréal peut-il être considéré comme
un milieu innovateur? Géographie, Économie, Société, 7(1), 37–56.

13
Scientometrics

Uriona-Maldonado, M., dos Santos, R. N., & Varvakis, G. (2012). State of the art on the systems of innova-
tion research: A bibliometrics study up to 2009. Scientometrics, 91(3), 977–996.
Uyarra, E., & Ramlogan, R. (2016). The impact of cluster policy on innovation. In J. Edler, P. Cunning-
ham, A. Gök, & P. Shapira (Eds.), Handbook of innovation policy impact (pp. 196–238). Cheltenham:
Edward Elgar Publishing.
Wolfe, D. A., & Gertler, M. S. (2004). Clusters from the inside and out: Local dynamics and global link-
ages. Urban Studies, 41(5–6), 1071–1093.

13

You might also like