The open innovation kaleidoscope: navigating pathways and overcoming failures

To ensure a comprehensive examination of the open innovation literature, our primary data source was the Web of Science Core Collection (WoSCC) database. The WoSCC is widely recognized for its extensive coverage of high-quality research articles across various disciplines. WoS is one of the most comprehensive and widely used citation databases in the academic community. It covers more than 13,610 journals across all disciplines (Singh et al. 2021; Falagas et al. 2008). To find most relevant articles in the field, we used a search string based on Gao et al. (2020). Additionally, in recognition of the first article in this domain written by Chesbrough (2003a, b), we refined our search to encompass articles from 2003 to 2023 (Gao et al. 2020; Kovacs et al. 2015). The detailed search criteria and keywords employed are presented in Table 1.

The initial retrieval yielded a sample of 2,551 articles. Recognizing the importance of data quality in conducting rigorous research, we undertook a thorough data-cleaning process. This involved supplementing missing abstracts, removing duplicates, and discarding articles without abstracts. After this rigorous cleaning process, we were left with a final sample of 2,537 unique articles These formed the basis for our subsequent analyses.

After finalizing the collection of relevant articles, we explored topic modeling, a method essential for uncovering hidden thematic patterns in large volumes of text (Blei 2012). This technique offers an objective perspective on dominant trends and provides a detailed understanding of the subject matter. In natural language processing and machine learning, a topic model is a statistical method for determining the “topics” in a set of documents. Latent Dirichlet Allocation (LDA) can uncover hidden semantic structures and topics in a large body of unstructured textual data using natural language processing, machine learning, and statistical algorithms (Blei et al. 2012; Wang and Blei 2011).

There are several notable advantages to using topic models. They rest on mathematically robust principles, elucidating the intricate dynamics of document generation. Moreover, they operate without needing prior categorization or labeling of documents, enabling an autonomous and expert-independent analysis. This autonomy extends to their capability to systematically organize and summarize vast swathes of documents, making them invaluable in text mining applications (Lee and Kang 2018). These attributes have led to an increased interest in topic models, finding successful applications across diverse text mining activities (Yan 2014). Several researchers in management studies have employed topic modeling techniques. Specifically, these approaches have been explored in fields such as marketing (Mustak et al. 2021; Amado et al. 2018), technology and innovation management (Lee and Kang 2018), information systems (Jeyaraj and Zadeh 2020), crisis innovation (Brem et al. 2023), open innovation (Lu and Chesbrough 2022), and human resource management (Thakral et al. 2023).

Before conducting topic modeling to identify relevant topics, some pre-processing procedures were required. Both the title and abstract of the articles were amalgamated to serve as the model’s input. This decision was based on the rationale that titles encapsulate the most representative terms, and the abstract delineated the study’s context, objectives, methodologies, and conclusions. We employed several steps to create a corpus that was used for topic modeling. In the initial phase, the entire text was divided into sentences and the sentences into the tokens (tokenization). Punctuation and numerical characters were subsequently excluded, and all characters were converted to lowercase. The subsequent step entailed the removal of words with fewer than three characters, including the extension of this process to eliminate structural words commonly found in abstracts, such as “aim,” “purpose,” “study,” “framework,” and “effect” (stop words). Next, word bigrams were created to link words that co-occur frequently. For instance, the combination of “business” and “model” was treated as “business_model”. Using lemmatization algorithms, words were transformed into their root forms to reduce dimensionality without loss of generality. For example, the term “industry” could manifest as “industry” or “industries” as a noun, “industrial” as an adjective, “industrialize,” “industrializes,” or “industrialized” as a verb, and “ndustrially” as an adverb. This was followed by word-stemming procedures that streamlined these words to their base forms, exemplified by the stemming of various forms of the term “industry” to “industri”. In addition, we removed all terms that occurred fewer than five times across all documents or that appeared in more than 70% of records. The final step in preprocessing was to convert the documents into a bag-of-words format. In this model, each document was depicted as a vector consisting of an unsequenced set of words. All of these tasks were implemented using gensim v. 3.8.3 (Rehurek and Sojka 2010), NLTK v. 3.7, and spaCy v. 3.0.0. (Honnibal et al. 2020).

We employed Latent Dirichlet Allocation-LDA (Blei et al. 2003) for topic modeling and used the Machine Learning for Language Toolkit (MALLET) for implementation. MALLET is an open-source Java-based machine learning package known for its sophisticated tools for statistical natural processing, document classification, sequence tagging, numerical optimization, and topic modeling, among others (McCullum 2002). One of its primary advantages is its scalable and efficient implementation of Gibbs sampling. Additionally, it provides efficient methods for document-topic hyperparameter optimization and has built-in tools for inferring topics on unseen documents using trained models. The MALLET tool is multi-threaded and optimized for performance on a single machine. However, it is worth noting its limitations. It can be memory-intensive, and handling extremely large datasets might lead to frequent garbage collection. As a result, it might not be scalable for massive datasets and can be challenging to scale across multiple nodes of a cluster (Sukhija et al. 2016). Despite these limitations, MALLET's robust features and efficiency, combined with the size of our data, made it a suitable choice for our study’s dataset and requirements.

Determining the optimal number of topics is a significant challenge in topic modeling. Researchers often use several measures, such as coherence and perplexity, to pinpoint the optimal number of topics. Although the perplexity measure is commonly used for this purpose (Jeong et al. 2019), there are no standard packages for its computation in MALLET. To address this issue, we used the topic coherence score to determine the optimal number of topics. This metric measures the quality of a given topic model by computing the semantic similarity between its highest-scoring terms. We used two specific coherence measures: c_v and u_mass. The c_v measure, rooted in a sliding window approach, uses a one-set segmentation of top words and indirect cosine similarity for confirmation. In contrast, u_mass is based on document co-occurrence counts and a one-preceding segmentation, confirming using the measure of log conditional probability (Röder et al. 2015). For c_v, higher values indicate better topic coherence, while for u_mass, values closer to zero suggest peak coherence (Röder et al. 2015) Our analysis employed both the c_v and u_mass measures. While some studies adopt statistical approaches, others rely on subjective analysis, where experts evaluate the appropriateness of number of topics (Madzík et al. 2023). Subsequently, we also engaged expert opinions to validate and refine our topic selection based on the coherence scores. As shown in Fig. 2, the coherence scores, when plotted against the number of topics, provide insightful observations. The c_v scores consistently hover around their peak within the range of eight to 13 topics, reaching a peak of 0.3991 at 10 topics. This consistency suggests a stable and coherent representation of the data within this range. Meanwhile, the u_mass score indicates two topics as optimal, with scores closest to zero. However, such minimalistic categorization could potentially oversimplify our dataset, neglecting its inherent complexity and nuances. Within the range of eight to 13 topics, the u_mass scores are less negative, hinting at a semantic closeness and meaningful topic delineation. Given these findings, and for a more comprehensive perspective, we deemed it prudent to explore the topic range of eight to 13, striking a balance between coherence and detailed representation. A thorough evaluation of these topics subjectively led us to determine the optimal topic number at 10, marked by a c_v score of 0.3991 and an u_mass score of − 2.154.

Full size image

Upon building the LDA model with 10 topics, we employed various techniques to further explore these topics. We visualized topics using the word_cloud library v. 1.8.1, where the size of each word within a specific topic is proportionate to its frequency within that topic. Additionally, we used the PyLDAvis library v. 2.1.2 for topic interpretation, based on the 10 topics identified earlier (Sievert and Shirley 2014). PyLDAvis facilitates topic visualization and offers deeper insights through its unique inter-topic distance mapping. This tool computes topic centers using Jensen–Shannon divergence (JSD) and calculates inter-topic distances using multidimensional scaling. It effectively maps multi-dimensional topic distances onto a two-dimensional plane, providing a spatial representation of topic proximity.

Subsequently, we also normalized the weight of each topic per year to analyze the annual changes in open innovation research. We used Mann–Kendall (MK) test. The null hypothesis in this nonparametric test is that the sample data are independent and randomly distributed (Hamed and Rao 1998). We used the pyMannKendall python package Version 1.4.2, which is a pure Python implementation of non-parametric Mann–Kendall trend analysis (Hussain and Mahmud 2019). However, before using the Mann–Kendall test, we needed to confirm that autocorrelation was not present in our data, as it could bias the Mann–Kendall results. For this, we employed the Durbin–Watson statistic, which tests for serial correlation between errors (Neeti and Eastman 2011). The Durbin–Watson test produces values that range from zero to four. A value close to two suggests no autocorrelation, while a value near zero indicates positive autocorrelation. Values near four imply negative autocorrelation. Our results shows that all values are below two, indicating the presence of positive autocorrelation in the data. In this case we used Hamed and Rao’s modification test to identify trends in topics (Hamed and Rao 1998). Upon further investigation, we noticed fluctuations in the data before 2002. To mitigate the potential biases these fluctuations might introduce to our trend analysis, we excluded data prior to 2002. After this adjustment, we recalculated the Mann–Kendall test to discern trends in the topics. To classify the most pertinent documents for each topic, we extracted the dominant topic for each document in our corpus. Using our LDA model, we assigned each document to the topic that had the highest contribution in that document. The “Dominant Topic” in our dataset is determined by identifying the topic number with the highest percentage contribution for that document. This is recorded as “Contribution %,” which signifies the weight of the topic in that particular document. This method ensures that each document is associated with its most relevant topic. It streamlines the process of analyzing journals and other sources based on topic dominance.

Following the identification of topics, we embarked on an interpretive analysis of each one. Every article in our sample was categorized based on its predominant topic. Subsequently, to contextualize and synthesize the most pivotal contributions, we reviewed the articles with a high percentage of contribution to their associated topic. Our initial approach involved a thorough examination of the abstracts of articles with significant contributions to each topic. This facilitated the selection of works serving as exemplars. The representativeness of these chosen articles was cross-verified against the topic-related terms for accuracy. Upon this validation, we investigated these selected articles with other significant contributions within the same topic domain. As a culminating step in our analytical process, after an exhaustive review of articles within each topic, we assigned descriptive failure labels to every topic to encapsulate its essence. We adopted a failure lens for the selected articles in each topic, based on the rationale that these papers investigate OI cases representing constraints, limits, risks, or challenges for OI implementation. We then categorized the associating level as micro, meso, or macro, and included theoretical and conceptual papers in this analysis.

The use of topic modeling enabled us to uncover latent semantic structures within the unstructured textual data we collected. Through this technique, we identified a model consisting of 10 distinct pathways, which effectively encapsulated a comprehensive range of OI-related subjects. The word-cloud analysis of the topic modeling outcomes, presented in Fig. 4, visually portrays the word distribution within the eleven identified topics.

Full size image

Table 3 provides an overview of the key terms associated with each topic, along with the corresponding count of documents and the cumulative citation count. These findings collectively underscore the diverse array of topics that span across various academic disciplines.

In this comprehensive examination of ten pivotal topics in the field of open innovation research, we present our result in each pathway implementing a dual analytical framework. Initially, we describe the principal attributes inherent in each pathway, subsequently progressing to an abductive exploration that unveils the associated failure mechanisms embedded within the open innovation landscape. This multifaceted analytical approach extends across the micro (individual), meso (organizational), and macro (system) levels of analysis, providing a holistic view of OI failures in these pathways.

As previously noted, we employed the weight of each topic per year to analyze the annual shifts in open innovation research. Figure 5 visually captures the evolution of these topics from 2003 to 2023. At a glance, it appears that the publication proportion for all topics experienced fluctuations, especially between 2003 and 2005. For a more rigorous understanding of these observed trends, we applied the Mann–Kendall test. The results of this test, delineated in Table 5, give a statistically substantiated account of the topic trajectories.

Full size image

The Mann–Kendall Test results presented in Table 5 provide valuable insights into the annual trends of open innovation research topics over a specific time period. Each research topic has been examined for its trajectory, with particular attention paid to the direction and significance of trends, denoted by p-values and z-scores. Moreover, the tau statistic has been employed to gauge the strength and direction of these trends, offering a comprehensive perspective on the evolution of each research theme.

On the other hand, certain research areas have seen a surge in interest. “Quadruple Helix” (Topic 5) and “OI and Resilience” (Topic 6) display increasing trends, with highly significant p-values of 0.034 and 0.001, and positive z-scores of 2.116 and 3.177 respectively. These results signify a growing focus on collaborative endeavors, business models, and ecosystem dynamics within the academic community. Meanwhile, “Managing Knowledge Inflows and Outflows” (Topic 3), “The Fuzzy Front-End of Innovation” (Topic 4), “Platforms and Communities” (Topic 8), and “Crowdsourcing” (Topic 10) present no significant trends, as indicated by their p-values and z-scores. These research themes have maintained a relatively stable level of attention in recent years. Finally, “OI Outputs” (Topic 9) has emerged as an area of increasing significance, shown by a significant p-value and a positive z-score of 5.324. This points to a notable uptick in scholarly focus on factors affecting firm performance within the open innovation context.

The inter-topic distance map provides valuable insights into the relationships and proximities between the identified topics in our open innovation research (Fig. 6). The two-dimensional visualization provides an insightful representation of the relationships between the topics in our corpus. The x and y axes represent the Inter-topic Distance Map, where each bubble represents a topic, and the distance between the bubbles indicates how distinct or similar the topics are from each other.

Full size image

The map shows that Topic 10 (Crowdsourcing), Topic 5 (Quadruple Helix), and Topic 8 (Platforms and Communities) are very similar. This is likely because crowdsourcing is often enabled by platforms and communities. Topic 1 (Value Creation and Capture Process) and Topic 2 (OI Strategic Management) are also close together on the map, because the two topics are closely related. Open innovation is a process that companies use to create new value and develop new services by leveraging external resources and ideas.

“The Fuzzy Front-End of Innovation” (Topic 4) and “OI Outputs” (Topic 9) exhibit intertwined themes. The cross-linkage of terms such as firm, product, and strategy reflects the potential interplay between product development and the overarching performance of the firm. It signifies that the effectiveness of new products or service developments often translates to the overall performance metrics of firms.

However, what stands out is “Managing Knowledge Inflows and Outflows” (Topic 3). Accompanied by terms like knowledge, absorptive capacity, and transfer, this topic encompasses the mechanics of managing, assimilating, and exploiting knowledge from both internal and external sources. Topic 3 indicates that knowledge management is a broad theme that may intersect with other OI topics without being tied to particular settings. Additionally, the intertopic map reveals three distinct controversies in the open innovation scholarly research community, which we will discuss and use as guidelines for future research directions.

Our abductive analysis shows researchers have examined OI failure at the micro-level. Specifically, learning and culture were found to contribute to risk in the majority of topics. This aligns with Cricelli et al. (2023), who discussed elements of OI adoption resistance, “not invented here syndrome” (NIH), and similar micro-foundational issues. Our contribution to the micro-foundational elements includes the cognitive limitations of individuals and the limited attentional capacity of top management teams in managing and organizing for open innovation.

When discussing cognitive limitations, we refer to the inability of individuals to assimilate knowledge within organizations. This perspective goes beyond prior studies, which assume the success of collaboration relies on the interpretation and perception of members about themselves and others in the collaboration (e.g., Skippari et al. 2017). Another micro-level factor that could better inform the future of OI research is the attentional capacity of management teams toward OI activities within organizations. In a study by Sisodiya et al. (2013), it was found that managers gave different levels of attention to inbound versus outbound OI activities and their potential effects. This imbalance negatively affected OI outcomes. When examining micro-level risk factors for organizing OI activities, organizational behavior theories such as selective attention theory could enrich research on OI failure.

At the meso level, significant risks associated with OI failure include inadequate business model adjustments (e.g., Albats et al. 2023), IP-related failures (e.g., Grandstrand and Holgersson 2014), and challenges related to measuring OI performance at the organizational level (e.g., Brunswicker and Chesbrough 2018). Research on OI failure at the organizational level is the most developed, which is not surprising given that OI is mainly developed and practiced at this level. However, some factors remain relatively new to the field and call for further investigation.

One such factor is the excessive practice of OI, which refers to situations where an organization overinvests in building external relationships, ideas, methods, and innovations without considering the specific requirements of each investment (e.g., Greco et al. 2016). Research on OI failure at the meso level could benefit from exploring the question of what constitutes efficient and balanced OI activities.

At the macro level, our findings align with those of Bertello et al. (2024), indicating that this is the least developed research domain in both OI research and in addressing failure. The impact of external influences on OI has been discussed previously, particularly in the context of post-COVID-19 scenarios, where OI was used as a strategy for business continuity during crises (e.g., Liu et al. 2022; Bertello et al. 2022). These studies demonstrate that overcoming crises often necessitates open innovation practices, especially for SMEs, to mitigate the negative impacts of a large-scale crisis such as the pandemic (Markovic et al. 2021).

However, research has overlooked the influence of economic crises on ongoing OI activities of firms. Two research foci that we found particularly interesting and underexplored are the effects of stagnant economic growth and market turbulence on open innovation activities. It remains to be seen how a stagnant market or industry impacts the OI activities of companies and whether marginal economic growth in a sector challenges collaborative innovation. Similarly, research on OI has not sufficiently addressed the challenges posed by market turbulence, such as high variations in customer preferences and product demand (Jaworski and Kohli 1993), on existing OI processes.