An initial overview of the dataset, as summarised in Figure 2, reveals a collection of 1,094 documents published between 2010 and 2024. These publications originate from a substantial pool of 4,203 distinct authors affiliated with 451 different sources, encompassing journals and conference proceedings. Collectively, these works cite a vast body of literature, totalling 53,615 references. The average citation count per document stands at a notable 21.46, suggesting a domain where research findings are gaining scholarly recognition and impact. With an average document age of merely 3.43 years, the dataset prominently features contemporary research, indicative of a field experiencing vibrant and recent activity.

A particularly striking statistic is the field’s high annual growth rate, calculated at an impressive 37.52%, which unequivocally signals an exponential increase in scholarly output at the confluence of AI, knowledge discovery, and innovation research, especially evident in the latter part of the analysed period. Furthermore, the data paints a picture of a highly collaborative environment, evidenced by an average of 5.22 co-authors per document and an international co-authorship rate reaching 28.43%. Conversely, documents authored by a single individual constitute a marginal proportion (only twenty-nine documents), collectively underscoring the inherently interdisciplinary and globally interconnected nature of research within this specific area.

Figure 3 visually reinforces the extraordinary growth trajectory highlighted by the descriptive statistics. The annual scientific production curve exhibits a relatively gradual increase from 2010 through to approximately 2018. However, mirroring the field’s rapid expansion, the number of publications per year displays a dramatically steeper incline from 2019 onwards, culminating in a significant peak in 2024. This pronounced acceleration in scholarly output during the more recent years strongly attests to the increasing strategic importance and practical adoption of AI-driven approaches for knowledge discovery tasks within innovation research contexts. It reflects not only advancements in AI methodologies, but also the growing availability of relevant data and the escalating interest from the research community in leveraging these capabilities to gain deeper insights into innovation processes and outcomes.

Offering a more nuanced perspective on the field’s scholarly impact, figure 4 tracks the average citations per document, per publication year, within the dataset. While the overall average citation rate suggests a field with considerable influence, the year-by-year averages reveal noticeable fluctuations, with peaks observed in earlier periods, specifically in 2013 and again in 2018. The subsequent decline in average citations for publications from the most recent years (post-2020) is a commonly observed pattern in fast-growing research domains, primarily because newer articles have had less time to accrue citations compared to their older counterparts (Waltman, 2016). Nevertheless, the presence of earlier citation peaks could signify periods during which particularly foundational or highly impactful studies were published, papers that continue to heavily influence the field and warrant closer examination in the analysis of highly cited documents later in this study.

Valuable insights into the core concepts and dynamically evolving themes that define this study domain are gleaned from an analysis of author-supplied keywords. A review of the top keywords by frequency, presented in table 1, clearly identifies terms central to the field’s discourse. High-frequency keywords such as classification (n=72), model (n=44), system (n=43), network (n=39), and methodological terms like neural-networks (n=30) and algorithm (n=27) prominently feature the AI techniques being applied. Concurrently, the frequent appearance of keywords denoting the application context, such as management (n=20) and the implicit presence of innovation, technology, and forecasting from the search strategy, underscores the embedding of these AI methods within innovation research. This consistent co-occurrence of computational techniques with innovation-related terminology compellingly affirms the interdisciplinary character of the field, where AI’s analytical power is being directed towards understanding and analysing various facets of innovation processes and complex systems.

The temporal trend analysis in figure 5 offers a dynamic view of the field’s evolving focus. In the earlier years (roughly 2014-2018), foundational and more specific analytical techniques, such as productivity and principal component analysis, show initial activity. Entering the more recent period (approximately 2021 onwards), a noticeable concentration emerges relating to terms central to AI methodologies and their application frameworks. However, the most current hotspots, characterised by large circles positioned around 2022-2024, are dominated by core application-oriented terms like system, network, and science. This cluster strongly suggests that contemporary research is heavily invested in applying predictive modelling, analysing complex systems, developing robust models, and situating these efforts within a broader scientific context.

Identifying the primary publication outlets is a critical step for comprehending the core communication channels and influential sources that disseminate research within this field. Table 2, which presents the top journals ranked according to various impact metrics (H-index, G-index, M-index, Total Citations-TC, Number of Publications-NP), reveals a varied portfolio of journals, mirroring the interdisciplinary nature of research on AI for knowledge discovery in innovation research. Leading the list is Pattern Recognition, with standout metrics (H-index of 27, 2410 total citations, and 86 publications). Other key venues include journals specialising in applied AI (Expert Systems with Applications), alongside prominent titles in specific engineering and application areas relevant to innovation dynamics (Transportation Research Part C: Emerging Technologies). Of particular significance is the inclusion of Scientometrics journal among these top-ranked sources. As a cornerstone publication in the SoS field, its presence strongly indicates that scholarly work applying AI to analyse scientific, technological, and innovation phenomena is not only published, but also holds significant recognition within the very community dedicated to studying science itself.

The cumulative publication trends over time, shown in figure 6, illustrate the shifting dominance and growth trajectories of key journals. In recent years, several journals have shown rapid increases in their publication volume within this domain. The most notable growth is seen in Pattern Recognition (blue line), which exhibits an exceptionally steep upward curve, solidifying its position as the primary venue. Following this trend, the International Journal of Pattern Recognition and Artificial Intelligence (green line) and IEEE Access (red line) also demonstrate significant and sustained growth, establishing themselves as major contributors. Furthermore, IEEE Transactions on Geoscience and Remote Sensing (brown line) and Transportation Research Part C: Emerging Technologies (pink line) show considerable acceleration in their publication rates, indicating their increasing importance as outlets for this research. This visualisation effectively highlights a competitive and expanding publication landscape, with a clear recent surge across several leading journals.

Table 3 provides crucial insights into the foundational works and highly impactful studies that have significantly shaped the intellectual contours of research on AI for knowledge discovery in innovation research. These publications, representing key reference points that have garnered substantial scholarly attention, are led by Wei et al. (2018) in Renewable and Sustainable Energy Reviews (534 total citations), followed by influential contributions from Mao et al. (2018) in IEEE Communications Surveys and Tutorials and Chavarriaga et al. (2013) in Pattern Recognition Letters. The diverse range of publication venues vividly underscores the field’s interdisciplinary nature, with impactful papers originating from core AI and pattern recognition journals (e.g., Pattern Recognition Letters, Swarm And Evolutionary Computation, Pattern Recognition) alongside journals focusing on specific application domains highly relevant to innovation systems, such as energy, communications, transportation (Transportation Research Part C: Emerging Technologies), urban studies (Cities), and engineering (Structural Health Monitoring). Spanning publication years from 2013 to 2023, these highly cited works encompass both established foundational research and more recent breakthroughs. Metrics, like TC per year and normalised TC, further highlight the rapid and significant uptake of recent works, such as Malekloo et al. (2021) and Yuan et al. (2023). Collectively, these highly cited documents illuminate the key theoretical frameworks, methodological approaches, and application successes defining the intellectual core of this rapidly evolving field.

Analysis of author affiliations provides a geographical and institutional perspective, identifying the leading organisations that are most actively contributing to research in this domain. Table 4, presenting the top affiliations by publication output, indicates a strong global presence in this study area, with a particularly notable concentration of highly productive institutions from China among the top ranks within this specific dataset. Universities and research bodies such as the Chinese Academy of Sciences (104 articles), Shenzhen University (sixty articles), Xidian University (forty articles), and Xi’an Jiaotong University (thirty-six articles), among others, demonstrate substantial scholarly output. This pattern points towards significant institutional focus, strategic investment, and concentrated expertise, dedicated to exploring the complex interplay between AI, knowledge discovery methodologies, and their application within the context of innovation research across these organisations. Understanding the geographical distribution and institutional strengths is thus vital for identifying potential research hubs and comprehending the global dynamics of knowledge production in this rapidly advancing interdisciplinary domain.

Complementing the analysis of affiliations, table 5 maps the top countries, based on their total publication output, offering a broader geographical perspective on the global distribution of research productivity in this field. Consistent with the institutional-level data, China emerges as the leading country by a considerable margin in terms of publication frequency (2,172), followed by the USA (214) and South Korea (ninety-one). Other nations like India (seventy-eight), the UK (seventy-eight), Australia (sixty-eight), and Canada (fifty-four) also appear prominently in the rankings, confirming that research at this intersection is indeed a widely pursued international endeavour. It should be noted that the frequency metric represents the total number of appearances of a country in the affiliation lists of all documents; therefore, a single internationally co-authored paper will contribute to the frequency count of each participating country. The observed dominance of certain countries underscores the strategic importance assigned to leveraging AI and data-driven approaches for understanding and fostering innovation within these nations’ research and development ecosystems. An exploration of the international collaboration patterns among these leading countries, a task reserved for the subsequent Network analysis section, will further illuminate the cross-border dynamics and collaborative structures that are actively shaping this evolving research landscape on a global scale.

The keyword co-occurrence network, visualised in figure 7, provides a structural map of the research themes animating the field of AI for knowledge discovery in innovation research. Generated using VOSviewer, this network elucidates the relationships between key concepts, based on their co-occurrence in the literature, with node size indicating frequency and colours representing distinct thematic clusters. Four primary clusters emerge, revealing the intellectual organisation of the domain. The largest cluster (red), situated on the left, appears to represent the core conceptual foundations, application contexts, and meta-analytical perspectives. It prominently features overarching terms like artificial intelligence and machine learning alongside innovation-centred keywords such as innovation, technology, emerging technologies, patent analysis, and trends. Crucially, this cluster also encompasses terms related to evaluation (performance, impact) and the methodologies used for analysing the scientific landscape itself, including bibliometrics, scientometrics, text mining, and natural language processing, directly linking the research to science of science practices.

Adjacent to this, the blue cluster focuses on classical machine learning tasks and algorithms, featuring keywords like classification, prediction, pattern recognition, feature selection, support vector machine, algorithm, and diagnosis, suggesting a focus on predictive modelling and established recognition techniques. The green cluster, dominating the right side, clearly pertains to advanced AI techniques, data processing, and applications often involving unstructured data. Central nodes like data mining and deep learning are surrounded by terms such as feature extraction, convolutional neural networks, transformers, image segmentation, remote sensing, and training, highlighting the adoption of sophisticated deep learning architectures. Finally, a smaller, less distinct yellow cluster is discernible at the bottom, centred upon sensors and potentially linked to adjacent terms like predictive models and graph neural networks, hinting at research focused on specific data sources, or emerging modelling approaches related to sensor data or complex network structures. The significant interconnections between these clusters underscore the field’s interdisciplinary nature, bridging core AI methods, advanced techniques, innovation studies, and self-reflective scientific analysis.

The structure of this network is highly revealing of the field’s interdisciplinary identity and its dual nature. The large red cluster’s role as a central hub connecting core AI methods, innovation contexts, and meta-analytical techniques is particularly significant. It demonstrates that researchers are engaged in a dual mission: (1) applying AI as a practical tool to analyse innovation (evidenced by terms like patent analysis and trends), and simultaneously (2) using the methodologies of the Science of Science (bibliometrics, scientometrics) to study and map the application of AI within this very context. This reflexivity, using the science of science to understand its own evolving methods, is a core tenet of the AI for science of science agenda and explains why these seemingly disparate concepts co-occur so frequently. The distinct blue and green clusters, representing classical and advanced AI techniques respectively, illustrate a methodological continuum, and their strong links to the red cluster show how these different computational tools are operationalised to address the core innovation challenges. Ultimately, the network map suggests a field that is actively integrating AI into a self-aware analytical framework for understanding scientific and technological progress.

Delving into the social structure underpinning knowledge creation in this field, figure 8 illuminates the landscape of scholarly collaboration through an author co-authorship network. This visualisation reveals several distinct, relatively dense clusters of researchers who collaborate frequently amongst themselves, connected by sparser links indicating inter-group collaboration. Each cluster appears to represent a community focused on specific methodological or application areas within the broader theme. For instance, the green cluster, prominently featuring Wei Wang (Chinese Academy of Sciences), suggests a focus on deep learning security, adversarial attacks, and data protection techniques. Adjacent to this, the yellow cluster led by Zhihui Lai (Shenzhen University) seems dedicated to advanced feature extraction methodologies, particularly manifold learning and adaptive graph embedding for pattern recognition. The blue cluster, centred on Licheng Jiao (Xidian University), indicates research into autoencoder architectures and feature learning, with specific applications in remote sensing image analysis like SAR change detection.

Bridging towards the right, the purple cluster represented by Yang Liu (Dongguan Polytechnic) points towards applications in educational data mining and learning analytics, using deep learning for student risk prediction. Further right, the red cluster, with Lu Bai (Central University of Finance and Economics) as a key figure, delves into novel graph representation techniques, potentially drawing from quantum-inspired methods and entropy for graph analysis. Connected to this, the orange cluster involving Xinbo Gao (Chongqing University of Posts and Telecommunications) appears focused on multi-view clustering and graph learning, potentially utilising tensor methods. Lastly, the distinct indigo cluster led by Lu Leng (Nanchang Hangkong University) concentrates on biometric recognition, specifically palmprint analysis using specialised network architectures. The overall network structure, characterised by these specialised yet interconnected communities, highlights both focused expertise within groups and the essential cross-pollination of ideas across different research teams and sub-disciplines.

The formation of these distinct, yet interconnected, research communities is likely driven by a combination of institutional, thematic, and social factors. The high concentration of researchers from leading Chinese institutions (e.g., Chinese Academy of Sciences, Shenzhen University, Xidian University) within specific clusters suggests that geographical proximity, shared institutional resources, and established academic networks, play a significant role in fostering these collaborations. Furthermore, the highly specialised nature of each cluster, such as the yellow cluster’s focus on manifold learning, or the blue cluster’s dedication to remote sensing, points to collaborations coalescing upon specific, complex research problems that require sustained, in-depth exploration. These specialised groups may represent core research labs, or teams funded by specific national grants. This structure reveals a knowledge production model where deep expertise is cultivated within focused groups, while broader field progress is achieved through the cross-pollination of ideas and methods between these specialised hubs.

The document co-citation network, presented in figure 9, illuminates the intellectual foundations upon which the research field of AI for knowledge discovery in innovation research is built. This map visualises which publications are frequently cited together within the analysed literature, revealing clusters of influential works that represent key knowledge pillars. Four distinct clusters emerge, signifying different, yet interconnected, streams of foundational influence. The large red cluster is anchored by seminal works in deep learning for computer vision, prominently featuring He et al. (2016) on Residual Networks (ResNet), alongside foundational contributions like LeCun et al.’s (1998) early convolutional neural networks (CNN) work and Deng et al.’s (2009) ImageNet dataset. Closely associated, the yellow cluster highlights other highly impactful vision architectures, potentially including works like Ronneberger et al. (2015) on U-Nets for segmentation or Hu et al. (2018) on attention mechanisms within convolutional neural networks, building upon the core CNN foundations.

Transitioning towards sequence-based models, the green cluster is centred on Vaswani et al. (2017), which introduced the revolutionary transformer architecture (Attention Is All You Need), fundamentally impacting natural language processing and beyond. This cluster also incorporates related works on sequence modelling and visualisation techniques like t-SNE (van der Maaten, 2008). Distinctly separate, yet connected through bridging citations, the blue cluster represents the domain of probabilistic topic modelling and bibliometric analysis tools. It is anchored by Blei et al. (2003) on Latent Dirichlet Allocation (LDA) and includes the pivotal work by van Eck and Waltman (2010) detailing the VOSviewer software, a tool fundamental to creating the very maps used in this study. The structure reveals a strong reliance on foundational deep learning literature (especially vision), alongside significant influence from breakthrough sequence models (Transformers) and key methods for text analysis (LDA) and science mapping (VOSviewer), reflecting the diverse methodological toolkit employed in the field.

Mapping the current research fronts through shared intellectual foundations, figure 10 presents the document bibliometric coupling network. This visualisation connects documents within the dataset that cite similar references, thus highlighting clusters of contemporary research focused on related themes, or drawing upon a common knowledge base. Distinct clusters emerge, signifying active research areas citing similar foundational literature. For instance, a prominent grouping (red and blue) showcases recent applications of advanced deep learning, exemplified by Mao et al.’s (2018) work applying deep learning (DL) to wireless networks and Yuan et al.’s (2023) investigation into hybrid CNN-Transformer architectures for medical image segmentation. Adjacent research, represented by Wei et al.’s (2018) highly coupled work (indigo), delves into data-driven building energy analysis and prediction.

Another significant constellation (yellow, green, orange) appears focused on leveraging real-world sensor or social data, addressing topics like human activity recognition benchmarking (Chavarriaga et al., 2013), traffic incident detection from social media (Gu et al., 2016), and array-based chemical sensing (Geng et al., 2019). More methodologically oriented research streams are also visible, such as the purple cluster surrounding Nguyen et al.’s (2020) work on swarm intelligence for feature selection, and the somewhat peripheral pink cluster linked to early work on traffic data imputation (Li et al., 2013). Furthermore, the distinct brown cluster (Wu et al., 2022) stands out, investigating the specialised area of transfer learning within brain-computer interfaces. The connections linking these varied clusters signify the shared foundational knowledge, likely common AI techniques or influential datasets, underpinning these diverse contemporary research directions. In short, this coupling map reveals a landscape characterised by distinct, but interconnected, research fronts, primarily driven by the application of sophisticated AI methodologies to diverse innovation-related challenges, spanning specific domain problems and methodological refinements.

Visualising the global footprint of research partnerships, Figure 11 maps the international collaboration network, based on co-authorship between countries within the dataset. This geographical representation highlights the key national players and the intensity of their collaborative ties in the field of AI for knowledge discovery in innovation research. Dominating the landscape are China and the USA, depicted with darker shading and serving as major hubs with numerous connections radiating outwards. A particularly strong and dense set of links connects China and the USA, signifying a highly significant bilateral collaboration axis in this domain. Beyond this central dyad, China also demonstrates substantial collaborative activity with other prominent research nations, including the United Kingdom, Australia, Singapore, and Canada. Likewise, the USA maintains robust collaborative links not only with China, but also with various European countries and Canada. While research productivity is globally distributed, as shown in the descriptive analysis, this collaboration map underscores the central role played by China and the USA in driving international partnerships, forming the backbone of a globally interconnected network that facilitates knowledge exchange across continents.