Conceptualizing and defining digital innovation ecosystems: A systematic literature review

As noted in the previous section, we selected 43 articles from a total of 131 retrieved across three databases for our research. Figure 2 illustrates the distribution of the included literature by year of publication. The first mention of DIE dates back to 2011, but research on the topic has been growing since 2018, peaking in 2023. This pattern suggests that DIE is only beginning to gain major attention in scientific discourse, even though the studied literature shows in-depth research on this topic.

Figure 2

Distribution of the articles by years of publication.

Figure 3 shows the distribution of studies by country. As can be seen from the graph, China is the leader in DIE research, and the USA, Germany, Brazil, Indonesia, and the UK are following. Other EU countries represented in the scientific literature are Austria, Finland, France, Italy, Latvia, Poland, Slovenia, Spain, Sweden, and the Netherlands. In total, the countries of the EU contribute to roughly half of the studies.

Figure 3

Distribution of the articles by country.

The distribution of articles in Table 2 is heterogeneous; the articles are nearly evenly divided between sources and publication types – journal articles and conference proceedings. However, in 2022 and 2023, we observe a noticeable increase in topical studies within the same journals. The authors suggest that this trend may indicate the gradual integration of the DIE phenomenon into scientific discourse in recent years.

The literature offers several DIE definitions, as shown in Table 3. According to the evidence gathered, a DIE is frequently described as complex, indicating that it embraces numerous interconnected stakeholders (Chae, 2019; Cvar et al., 2020; Li et al., 2022; Liu et al., 2023a; Xu, 2020). The scholars emphasize the presence of the combination of social and technical aspects within the DIE, suggesting that DIEs do not just involve joint technology exploitation but also social interactions between stakeholders in the ecosystem (Kolloch & Dellermann, 2018; Wang, 2018, 2019). Some scholars stress that DIEs have loosely coupled architecture (Ji et al., 2023; Wang, 2021). The primary goal of a DIE is to foster innovation, leading to the development of novel products and services and value creation using digital technologies, as an integral part of the DIE (Suseno et al., 2018; Wang, 2019; Xu, 2020). The scholars stress the presence of technological (digital) and social/physical mutually interdependent and interconnected components (Beltagui et al., 2020; Chae, 2019; Kolloch & Dellermann, 2018; Xu, 2020); the components of the DIE constantly co-evolve, enhancing interaction capabilities (Chae, 2019; Cvar et al., 2020; Liu et al., 2023a). The dynamic nature of the DIE makes it resilient and adaptable to changing environments (Wang, 2019). Some definitions presented in the research stress the presence of competition in the DIEs (Beltagui et al., 2020; Liu et al., 2023a; Xu, 2020). Taking into consideration all the above, the authors propose the following DIE definition, based on the definitions presented in Table 3: DIE is an ecosystem that integrates technological aspects and unites various actors who continuously co-evolve, cooperate, compete, and leverage digital technologies to generate value.

In order to categorize the studies under investigation, the system-level approach by Vanhamaki et al. (2019) was applied; this research introduced the system-level approach to the circular economy and described the actors on macro, meso, and micro levels. The authors of the study brought this model to the research on DIE, dividing the studied literature by system levels to which the DIE was applied. Some of the research studied the universal nature of DIE applicable in all the levels, so we introduced the fourth group named “Meta level.” The results of the distribution are presented in Table 4.

Some authors reflect on the interrelation between DIEs and other innovation ecosystems. Beltagui et al. (2020), Kolloch and Dellermann (2018), and Pistorio et al. (2018) claim that DIEs are synonymous with innovation ecosystems. Wang (2018) argues that many innovation ecosystems are DIEs, while Baumane-Vītoliņa and Dudek (2020) state that DIEs are one of the types of innovation ecosystems. Since there is no unanimous consent on the nature of the DIE, we need to proceed with its integrative conceptualization. As a consequence of the in-depth study of the literature corpus, DIE definitions (Table 3), and DIE system levels (Table 4), we revealed several patterns that help to shape the DIEs’ conceptualization. Regardless of the system level at which the DIE has been considered, there is a clear tendency to mention technology as a meta-factor for the existence of the DIE. In addition, the general trend is the interconnection between the physical and social components of the ecosystem, as it was discussed above; the institutional landscape also has an important role in the functioning of the DIE. Figure 4 represents the DIE conceptual framework and the four main components of the DIEs. These components are as follows:

technology – this component comprises technological advancements in general (all the authors emphasize the importance of this component for the successful existence of the DIE), as well as digital solutions such as digital platforms (Baumane-Vītoliņa & Dudek, 2020; Misséri, 2013; Nepelski, 2019; Pistorio et al., 2018; Rao & Jimenez, 2011; Wang, 2018, 2019), that enable actors of the ecosystem to connect with each other, cooperate, sometimes compete, deliver value to others, exchange knowledge and experience.

Figure 4

DIE conceptual framework.

We believe all ecosystem components are equally important for the successful functioning of the ecosystem and exist autonomously, but at the same time, they are in interaction with each other. At the same time, the technology and institutional landscape components are meta elements as they serve as the foundation for the creation, existence, and development of the other two components (Physical environment and Social environment). Technologies and the institutional landscape create conditions for the interaction of the physical (Physical environment) and the social component (Social environment) and favor the emergence of other DIEs in the intersection of these fields.

To further validate the DIE conceptual framework and, therefore, to grant our results additional methodological robustness, we looked beyond the results of the qualitative investigation and coding and performed an LSA of the 43 retrieved papers. Essentially, here, we treat LSA as a form of quantitative triangulation to our qualitatively derived taxonomy and a way to further validate qualitatively derived results in a more neutral way that is less prone to the subjectivity of a researcher.

LSA is a well-established natural language processing technique based on the distributional hypothesis, assuming that words that are similar in meaning (i.e., have high semantic similarity) occur in semantically similar documents (Evangelopoulos, 2013). In essence, it treats words as variables, while the frequency of a specific word in a document constitutes a particular observation for a given variable. The resulting table is called the document-term matrix (DTM) and it is typically first pre-processed to account for different documents’ lengths. After that, a singular value decomposition (SVD) is applied to the DTM, identically to how a principal component analysis is performed to reduce the number of dimensions in a dataset.

Hence, the initial space of variables is reduced to a lower number of principal components (or dimensions/eigenvectors) that jointly summarize the textual corpus in fewer dimensions while retaining the initial variance of the DTM. Particularly, the importance of components is captured in descending order, starting from the most relevant dimension in terms of percentage of captured variance, and then followed by the series of orthogonal dimensions (therefore, uncorrelated) with diminishing contributions to the total variance. In this manner, a complex semantic space can be represented as a series of orthogonal dimensions in which terms that frequently co-occur are close to each other. Respectively, those documents that are overall similar, will tend to cluster.

The main reason we opted for LSA instead of more complex natural language processing techniques is that, unlike those, LSA is entirely deterministic and is not veiled behind the “black box” of weights and biases resulting from the machine learning processes. Hence, the process is replicable and leads to identical results each time SVD is applied. Moreover, LSA is a well-established and validated technique with a long “pedigree” in computational linguistics.

To begin with, we uploaded the documents to qualitative data analysis software (MAXQDA, release 22.6.1) and calculated the absolute frequencies of all words across the 43 papers. English language lemmatization was also performed to account for different spellings, plurals, and words with identical semantic roots (e.g., systems and systems are counted as one token). We have also applied a stop word list containing 412 auxiliary words used in English (e.g., a, above, about, and accordingly) to exclude terms that carry no particular semantic meaning from the initial frequency count.

Overall, 20,567 unique words were identified (type-token ratio = 0.0585). The contributing researchers have then separately assessed the respective absolute frequency table and selected all the nouns, verbs, and adjectives deemed directly or indirectly related to the topic of DIE. To increase the inter-rate reliability of the results, we compared the respective lists and collegially decided on the words to include in the final analysis of semantic relatedness. In total, 48 words were selected. According to LSA scholars (Landauer et al. 1998), while around 300 dimensions are optimal for moderate-size document collection (hundreds of thousands of documents), for a smaller-size textual corpus, like in our case, around 50 dimensions are sufficient.

Figure 5 shows all the words included in the analysis, with the size of words linearly proportional to their absolute frequency across papers.

Figure 5

The most frequent words related to the topic of DIE.

To account for the different lengths of the analyzed papers, we transformed all the absolute frequencies into relative ones by dividing (for each of the 43 papers) the absolute frequency of every keyword (Figure 6) in a paper by the sum of absolute frequencies of all the keywords selected for that paper. We then imported the obtained DTM into RStudio and performed a principal component analysis on the matrix.

Figure 6

Scree plot. The ten most contributing principal components.

Out of 48 initial variables, PCA extracted 40 uncorrelated components (dimensions). In order to understand which principal components are important for the interpretation, it is customary to rely on the scree plot (Figure 6) reporting the contribution of the individual components to the overall variance. In particular, Figure 6 shows the first ten components’ contributions. Given a relatively gradual decrease in variance explained and, hence, the difficulty in interpreting principal components beyond the second one, only the first two dimensions (accounting for roughly 20% of explained semantic variability of the analyzed papers) were analyzed in-depth.

Figure 7 shows the correlation of the variables (keywords) with the two most contributing PCA dimensions. As such, it can be understood as a summary representation of the semantic space of the 43 papers included in the review. Accordingly, terms that are close to each other (i.e., when the angle between vectors is small and the length of vectors is similar) are more likely to appear together in a document.

Figure 7

PCA. The correlation of active variables with the first two dimensions. Note: cos² denotes the quality of the representation of variables on the axes. For better presentational clarity, only variables with cos² > 0.1 are reported.

The horizontal dimension, which is also the most important in terms of the overall contribution to the construction of the PCA plane, seems to highlight the crucial conceptual distinction between papers primarily focused on technology-related themes and topics in the north-east and south-east parts of the plane (e.g., feature, algorithm, data, internet, game, feature, development, manufacture, and technology), and papers that are mainly dealing with topics related to business, management, and commercialization of the final products (e.g., product, business, firm, customer, market, organization). The vertical axis was more challenging to interpret. Still, the presence of the most significantly contributing keyword (market) in the northmost part of the graph suggests that the vertical axis captures the conceptual distinction between papers emphasizing the role of institutional macro and meso-level elements in the northern quadrants (e.g., market, university, environment, industry, economic, and government) and aspects related most abstract analyses of strategical importance (e.g., knowledge, platform, co-creation) in southern quadrants of the plane. The latter clearly indicates that the papers in this part of the graph are more centered on what we call meta-level analysis, treating themes and topics common to micro, meso, and macro levels of analyses. Moreover, when it comes to the DIE framework (Figure 5), the keywords denoting all of its elements are present in the PCA results. While the embedded institutional environment is clearly captured by the vertical axis and technology is captured by the horizontal axis, keywords denoting both physical and social environments are distributed more or less evenly on the plane (e.g., customer, actor, ecosystem, organization, and firm).

A valid follow-up to a visual exploration of PCA results is typically given by an agglomerative hierarchical clustering. The method follows a “bottom-up” approach, starting with single observations being considered as clusters on their own, and then proceeding through the aggregation of clusters until within-class variance is minimized and between-class variance is maximized. In the context of this work, the clustering was performed to further validate the interpretation of the PCA factorial plane (and, hence, the underlying semantic space it represents).

To perform the clustering, the Euclidean distance was chosen to measure the similarity between individual observations (papers), and Ward’s linkage method was used to ensure partitioning with small within-class variability and high between-class variability. Figure 8 shows the resulting cluster dendrogram projected onto the first two PCA dimensions (with dots denoting individual papers). The partitioning of papers into three macro-thematic clusters was deemed optimal by the R package used for the analyses (Factoshiny). Indeed, additional variability gains from considering the fourth cluster and onward were not substantial.

Figure 8

Cluster dendrogram projected on the first two PCA dimensions.

Table 5 shows the importance of individual keywords (and, hence, the related themes and topics) in the three clusters. Some keywords, such as “digital,” “system,” “network,” and others, present in the semantic space, are completely missing within the clusters’ definitions as they were not significantly contributing to clusters’ definitions. In essence, those were equally present not only within clusters, but also between them, and the related differences were not statistically significant. In this regard, Table 5 reports the ten most contributing keywords to the definition of each cluster (with p-value < 0.05).

Cluster 1 is the most numerous and is defined mainly in terms of keywords related to different organizational and business factors of micro and meso levels of analysis while also emphasizing physical and social environment components of the DIE framework (e.g., actor, business, organization). Cluster 2 is the most balanced, albeit least numerous. Indeed, the related papers mainly seem to belong to the meta-level of analysis, showcasing topics of more transversal relevance (e.g., co-creation, strategy, knowledge, and improve). The emphasis, however, is slightly in favor of managerial and organizational topics rather than blockchain technology itself (e.g., contract, distributor, retailer trade). Finally, the third cluster is focused on the macro level of analysis, emphasizing both institutional (e.g., industry, financial, and environment) and technology (e.g., smart, technology, and manufacture) factors.

In conclusion, clustering analysis mirrors the opposition and integration of micro-macro dynamics in the definition of DIE, as resulting from qualitative analysis and depicted in Figure 4. The opposition is captured mainly by the most contributing variables (highest t-scores in Table 5), which are more organization-centric in the first cluster, and more context-centric in the third, with the second cluster being more balanced. The integration dynamics are also present as certain synonyms have contributed to the creation of distinct clusters – i.e., firm, business, enterprise, and company. Hence, the identified categories and different levels of analysis of the conceptual framework in Figure 4 are all distinctly represented within the analyzed semantic structure of the analyzed textual corpus.