The Sustainable Innovation of AI: Text Mining the Core Capabilities of Researchers in the Digital Age of Industry 4.0

Abstract

Sustainable innovation in the field of artificial intelligence (AI) is essential for the development of Industry 4.0. Recognizing the innovation abilities of researchers is fundamental to achieving sustainable innovation within organizations. This study proposes a method for identifying the core innovative competency field of researchers through text mining, which involves the extraction of core competency tags, topic clustering, and calculating the relevance between researchers and topics. Using AI as a case study, the research identifies the core innovative competency field of researchers, uncovers opportunities for sustainable innovation, and highlights key innovators. This approach offers deeper insights for AI R&D activities, providing effective support for promoting sustainable innovation. Compared to traditional expertise identification methods, this approach provides a more in-depth and detailed portrayal of researchers’ expertise, particularly highlighting potential innovation domains with finer granularity. It is less influenced by subjective factors and can be conveniently applied to identify the core innovative competency field of researchers in any other research field, making it especially suitable for interdisciplinary areas. By offering a precise and comprehensive understanding of researchers’ capability fields, this method enhances the strategic planning and execution of innovative projects, ensuring that organizations can effectively leverage the expertise of their researchers to drive forward sustainable innovation.

1. Introduction

Under the impetus of Industry 4.0, innovations in artificial intelligence (AI) technology have become a key driving force for economic and social development [1,2]. In an ever-changing market and technological environment, possessing sustainable innovation capabilities in the AI field is important for enterprises and countries to achieve a sustainable competitive advantage [3,4]. However, maintaining the sustainability of innovation requires careful planning and execution in areas such as resource allocation, talent cultivation, and knowledge management [5]. This study focuses on identifying and evaluating the core innovative capabilities of researchers to improve the efficiency of talent allocation and the integration of knowledge resources.

The characterization of researchers’ expertise forms a fundamental basis for expert recommendations in innovation activities. Current studies on this topic focus on three dimensions: competency models [6], information organization [7], and expertise mining [8]. Within these dimensions, models constructed from the perspective of information organization, such as the FOAF (Friend of a Friend) model proposed by W3C and the CERIF (Common European Research Information Format) model developed by the European Union, provide a structured approach to categorizing expertise [9]. Expertise mining, on the other hand, leverages big data to offer a more objective characterization. This method has gained prominence due to its timeliness and accuracy, becoming the mainstream approach in the field [10]. The integration of vast datasets allows for a comprehensive and precise assessment of researchers’ skills and capabilities, which is beneficial for facilitating effective innovation and collaboration.

Despite these advancements, existing studies have not yet examined deeply the fine-grained knowledge backgrounds of researchers. The granularity of expertise characterization remains relatively weak, leaving gaps in understanding the specific nuances of researchers’ knowledge and skills [11]. A more detailed exploration could enhance the accuracy and relevance of expert recommendations, particularly in complex and rapidly evolving fields [12,13]. As the demand for precise expertise identification grows, future research could develop methods that capture the intricate details of researchers’ knowledge backgrounds [14]. This will provide a more comprehensive understanding of their capabilities and improve the quality and effectiveness of expert recommendations.

Researchers’ expertise is embodied in their skills and knowledge, which form a type of tacit knowledge providing a sustained competitive advantage [15,16]. Although researchers’ expertise can be characterized through thematic mining of the text content in their innovation results, existing studies primarily focus on identifying their research interests [17]. This approach is relatively superficial, as it does not capture the deeper, more stable aspects of their expertise. Research interests can fluctuate with shifts in attention, and a broad array of interests can obscure the core knowledge that researchers rely on for innovation [18,19]. This core knowledge is essential for understanding the foundational capabilities that drive their innovative activities.

An advanced approach is needed to accurately characterize researchers’ expertise, one that goes beyond surface-level interests and the fundamental knowledge underpinning their work. This deeper level of understanding is for maintaining a sustained competitive advantage in innovation activities. By focusing on the core innovative capabilities of researchers, future studies can provide a clearer picture of the skills and knowledge that truly enable researchers to excel and lead in their fields. This, in turn, will facilitate more effective expert recommendations and foster greater innovation.

This study proposes an automated text-mining method to extract and evaluate the core innovative capabilities from researchers’ outputs. Focusing on AI scientists, it identifies researchers’ core capability field through expertise tags, thematic clustering, and relevance analysis. Additionally, it determines the relative positions of researchers within the innovation community using an expertise evaluation model based on professional and comprehensive ability characteristics. By identifying and assessing the core capabilities of AI researchers based on their knowledge backgrounds, this study provides direct insights into the direction of sustainable innovation and the organization of innovation personnel, thereby supporting sustainable innovation in the AI field.

The remainder of this study is structured as follows: Section 2 provides an overview of related research. Section 3 presents the research approach and framework. Section 4 describes the experimental procedures. Section 5 analyzes the experimental results and interprets the advantages of this method compared to traditional methods. Section 6 discusses the research findings, and Section 7 concludes the study.

2. The Related Research

2.1. Research on Expertise Domain Identification

Individuals with professional knowledge in a specific field are generally recognized as experts [20]. However, identifying the precise domain of an expert’s expertise poses a significant challenge. In expert recommendation systems, researchers’ personal profiles or individual documents are often matched with search content to identify leading experts [21,22,23,24,25]. This matching process employs various thematic modeling methods. Some scholars use thematic clustering to pinpoint experts’ domains [26,27], while others cluster texts based on identified LDA latent topics to define researchers’ subprofiles [28].

To optimize matching results, joint modeling could be adopted to identify common themes between submissions and experts’ profiles, thereby ascertaining expertise [29]. Moreover, experts’ Wikipedia knowledge models are established by matching document terms with Wikipedia knowledge entities [16]. These approaches aim to refine the identification process, ensuring that recognized experts align closely with the specific thematic content of inquiries and enhancing the effectiveness of expert recommendation systems.

The identification of expertise domains is used to organize agile innovation teams, particularly for more complex tasks. In this context, expert recommendations require a detailed analysis of experts’ backgrounds [30]. An expert typically possesses multiple domains of expertise, making it essential to understand these varied skill sets comprehensively. T-shaped experts, who possess deep knowledge in one domain and broader, shallow knowledge in others, are especially valuable for building collaborative teams [31,32]. These individuals can contribute both specialized expertise and a broad understanding of related fields, enhancing team dynamics and innovation potential.

To identify T-shaped experts, a method based on time analysis helps in discovering individuals with this unique skill set [33]. Similarly, the formation of agile innovation teams has been proposed by identifying T-shaped experts through thematic modeling [34]. By employing these advanced identification techniques, organizations can more effectively assemble teams with the right mix of deep and broad expertise, aiming at fostering a more innovative and collaborative environment.

Existing research has explored various thematic modeling methods to finely detail experts’ domains of expertise. Few studies have focused on the significance of thematic terms in representing the core professional competencies of experts, which is evidently to uncover the sustainable innovative potential of researchers and forming high-quality innovative teams.

2.2. Research on the Evaluation of Researchers’ Expertise

The evaluation of researchers’ expertise is conducted through ranking experts within their fields, reflecting their capability to engage in innovative activities. Current studies assessing researchers’ capabilities primarily consider factors such as contributions to innovation, impact on the field, and relevance of research to specific domains [9]. The contributions of a researcher are often judged by the order of authorship on publications, indicating their role and involvement in the work [35].

In addition to authorship, a researcher’s impact can be measured using various metrics, such as citation counts and the h-index. Citation counts provide a quantitative measure of how frequently a researcher’s work is referenced by others, indicating its influence and reach within the field [36,37]. The h-index combines productivity and citation impact, offering a balanced metric to assess a researcher’s overall contribution [38,39]. These evaluations are relevant for understanding the relative standing of researchers in their expertise areas [40], guiding decisions in expert recommendations and team formation for innovation activities.

Innovative capacity is another dimension for evaluating researchers’ expertise. This approach stems from the development of a convenient metric known as the disruptiveness index [41]. The disruptiveness index (DI₁) is a quantitative instrument designed to measure the extent to which innovative outcomes impact existing innovation trajectories. Since its inception by Funk and Owen-Smith in 2017, the DI₁ has been extensively applied across various academic disciplines to identify the most disruptive publications [42]. By analyzing citation networks, the DI₁ differentiates between consolidating research, which often relies on established knowledge frameworks, and disruptive research that challenges conventions by introducing novel theories and methodologies [43]. Researchers’ innovative capacity in specialized fields could be measured using an optimized disruptiveness index [44]. For instance, a hybrid method that combines scientific impact with innovative capacity was designed to assess scientists [45]. The disruptiveness index provides a quantitative means to gauge researchers’ innovativeness.

A higher disruptiveness index indicates greater innovativeness of the outcomes, making it an effective tool for quantifying researchers’ innovative capabilities. While this metric offers valuable insights, existing studies have not strictly correlated the level of innovative capacity with the knowledge background that represents innovative capability. This gap suggests that current evaluations may overlook how a researcher’s foundational knowledge contributes to their ability to innovate. Future research could establish a more direct link between a researcher’s knowledge base and their innovative capacity to enhance the accuracy and relevance of such evaluations.

The extant research on the evaluation of researchers has progressively examined the dimension of innovative capacity. However, the assessment of innovative capacity has been based on the identification of areas of interest without tracing the true knowledge sources of innovative capability. This study will rank researchers’ innovative capacities within niche fields and innovation communities based on the identification of the core innovative competency field to find individuals with sustained innovation capacity in specific areas.

3. Research Approach and Framework

3.1. Theoretical Foundation

This study identifies core competency tags based on the theoretical assumption that different types of knowledge units in scientific papers exhibit varying abilities to reveal the author’s innovative capabilities. The innovative ‘spark’ knowledge units serve to characterize the author’s core competencies.

The role of knowledge units from different sources varies significantly in the innovation generation process. Knowledge units in scientific papers could be categorized into ‘heritage’ ones and ‘spark’ ones [46]. Heritage knowledge originates from references, while spark knowledge emerges solely from the innovator’s mind and does not appear in the references. Scientific innovation is essentially the fusion of spark knowledge with existing theories, frameworks, and methods from references, thereby creating new knowledge value. The core of innovation, in this context, originates from the author’s unique, unencoded tacit knowledge [47]. Spark knowledge units, being closer to this tacit knowledge, serve as the explicit manifestation of these unspoken insights to better reflect the core innovative capabilities of the researchers.

Further evidence of the importance of spark knowledge units has demonstrated that the innovative impact of scientific papers directly stems from spark knowledge entities. A disruptive innovation index algorithm based on the propagation features of knowledge entities was developed [48]. The results indicated that the proportion of spark knowledge entities in the focal literature, the total number of spark knowledge entities inherited from the focal literature, and the new entities generated in citing documents positively affect the innovation degree of the focal literature. This suggests that spark knowledge entities are essential in driving the innovative impact of scientific work, emphasizing the need to consider these unique contributions when assessing the innovative capabilities of researchers.

Spark knowledge units reveal the essence of researchers’ core innovative capabilities. To minimize the noise from heritage knowledge units and accurately reflect the domain-specific characteristics of these capabilities, this study will detect researchers’ core competency through spark knowledge units.

3.2. Algorithms and Models

3.2.1. Extraction of Core Competency Tags

Meme is a cultural entity such as a concept, melody, or recipe that spreads through imitation, involving replication and mutation. This process can be considered an analog of ‘genes’ in the cultural sphere [49]. A method is proposed for identifying knowledge memes in the scientific field based on the knowledge inheritance patterns in scientific paper citations [46]. Knowledge memes are identified as tiny knowledge units that can replicate through citations; the greater the likelihood that a knowledge unit in the cited literature is decomposed, altered, or non-existent in the citing literature, the less likely it is to be considered a knowledge meme. Through the computational method, important concepts or entities within any research field can be identified, providing a systematic way to track the propagation of knowledge [50,51].

Given that not all fields have a clear, standardized knowledge entity database, this study aims to identify significant spark knowledge in innovation based on the propagation characteristics of important knowledge memes within the paper citation network. These knowledge memes encompass both key concepts and entities and are more adept at revealing semantic features than knowledge entities alone. By considering the impact value on the academic field, the spark knowledge that can be further propagated will be used as characteristic words to identify the authors’ core innovative competency domains. This approach will help derive the authors’ expertise tags based on the authors’ core innovative competency domains, thereby offering a refined method for evaluating and categorizing researchers’ innovative capabilities. The specific steps and algorithms are as follows.

①

Identification of Knowledge Memes

This study employs the identification algorithm [46] to obtain knowledge memes within a specific field. Initially, nouns or noun phrases of 1–3 words in length are extracted from the titles and abstracts of papers as candidates. Subsequently, these candidates are scored according to Formulas (1) and (2).

$${\mathrm{P}}_{\mathrm{m}}={\displaystyle \frac{{\mathrm{d}}_{\mathrm{m}\to \mathrm{m}}}{{\mathrm{d}}_{\to \mathrm{m}}+\mathsf{\delta }}}/{\displaystyle \frac{{\mathrm{d}}_{\mathrm{m}\to \sout{\mathrm{m}}}+\mathsf{\delta }}{{\mathrm{d}}_{\to \sout{\mathrm{m}}}+\mathsf{\delta }}}$$

(1)

$${\mathrm{M}}_{\mathrm{m}}={\mathrm{f}}_{\mathrm{m}}·{\mathrm{P}}_{\mathrm{m}}$$

(2)

Specifically, ${\mathrm{f}}_{\mathrm{m}}$ denotes the frequency of a term m across all documents, ${\mathrm{d}}_{\mathrm{m}\to \mathrm{m}}$ represents the number of documents that carry m and cite other documents carrying m, ${\mathrm{d}}_{\mathrm{m}\to \sout{\mathrm{m}}}$ is the number of documents that carry m but do not cite documents carrying m, ${\mathrm{d}}_{\to \mathrm{m}}$ is the total number of documents citing those carrying m, and ${\mathrm{d}}_{\to \sout{\mathrm{m}}}$ is the total number of documents not citing those carrying m. A constant of $\mathsf{\delta }$(=3) is added to avoid division by zero and to prevent terms with extremely low frequency from obtaining high scores. Terms and phrases with scores of ${\mathrm{M}}_{\mathrm{m}}$ greater than 0 are identified as knowledge memes. This score measures the interestingness and importance of a knowledge meme, allowing for the identification of significant knowledge memes within a certain score threshold as candidates of expertise tags.

②

Extraction of Tags

Based on the propagation characteristics of candidate knowledge memes, spark knowledge memes with direct influence on the citing paper are extracted from the focal literature to serve as the author’s expertise tags. Figure 1 illustrates the process of extracting these tags. Initially, the focal paper’s references and citations within the field are identified, followed by the recognition of knowledge memes that propagate through citation relationships. The co-occurrence of knowledge memes in both the citing and cited literature indicates their propagation via citations. By comparing the set of knowledge memes propagated from references with those propagated from the focal literature, those present only in the latter set are identified as tag words. In Figure 1, the focal literature contains knowledge memes a and b inherited from references, as well as emergent knowledge memes d, e, and f; the citing literature includes knowledge memes a, b, d, and e propagated from the focal literature. Knowledge memes d and e are identified as spark knowledge memes with direct influence on the citing papers. Consequently, feature words d and e in the focal paper are assigned as the authors’ core competency tags.

3.2.2. Identification of Expertise Domains

The identification of expertise domains is achieved by clustering tags into themes and calculating the author–topic relevance, resulting in the creation of an author–topic relevance matrix. This matrix forms the basis for determining the thematic areas related to an author’s expertise. The specific steps and algorithms are as follows:

①

Topic Clustering

The LDA topic model, an unsupervised machine learning algorithm widely used in text mining and information retrieval [52,53], is employed to cluster the authors’ expertise tags. Topic clustering facilitates the depiction of researchers’ core competency field from a semantic content perspective.

②

Author–Topic Relevance Matrix

The author–topic relevance matrix reveals the author’s embedding characteristics within the thematic space. By selecting the top 30 topic words for each theme based on probability, the similarity between the author’s expertise tags and the topic words, as well as the relevance between the author and the topic, are calculated. The formula is as follows:

$${\mathrm{S}\mathrm{i}\mathrm{m}}_{\mathrm{i},\mathrm{j}}=\left\{\begin{array}{c}1,{\mathrm{w}\mathrm{o}\mathrm{r}\mathrm{d}}_{\mathrm{i}}={\mathrm{w}\mathrm{o}\mathrm{r}\mathrm{d}}_{\mathrm{j}}\\ 0,{\mathrm{w}\mathrm{o}\mathrm{r}\mathrm{d}}_{\mathrm{i}}\ne {\mathrm{w}\mathrm{o}\mathrm{r}\mathrm{d}}_{\mathrm{j}}\end{array}\right.,$$

(3)

$$\mathrm{Relevance} \mathrm{Score}=1/30\times {\sum }_{\mathrm{i},\mathrm{j}}{\mathrm{s}\mathrm{i}\mathrm{m}}_{\mathrm{i},\mathrm{j}},$$

(4)

where ${\mathrm{S}\mathrm{i}\mathrm{m}}_{\mathrm{i},\mathrm{j}}$ represents the similarity between any tag word i and any topic word $\mathrm{j}$. An author-topic relevance matrix is constructed based on scores of relevance between authors and topics.

3.2.3. Expertise Evaluation Model

The expertise evaluation model is constructed with two dimensions: the level of expertise in representative niche domains and the level of comprehensive expertise. The specific content and algorithms are as follows:

①

Evaluation of Expertise in Representative Niche Domains

The relevance between an author’s expertise tags and themes reflects the depth of a scientist’s mastery of professional knowledge in that field, i.e., their professional degree. Then, the relative ranking of an individual with their professional degree within the academic community can be identified.

②

Evaluation of Expertise Based on Comprehensive Expertise

A weighted author–topic bipartite network is constructed based on the author–topic relevance matrix, with relevance serving as the weight of the edges. The division of the community network structure can reveal innovative communities of researchers based on similar capability backgrounds. The weighted degree centrality of network nodes can assess the significance of researchers’ expertise within their community, considering both the depth of their professional degree (topic relevance) and its topic breadth (the number of related topics). This approach allows for the determination of the relative ranking of researchers based on their comprehensive expertise. The calculation formula for the weighted degree centrality is as follows:

$$\mathrm{W}\mathrm{e}\mathrm{i}\mathrm{g}\mathrm{h}\mathrm{t}\mathrm{e}\mathrm{d} \mathrm{D}\mathrm{e}\mathrm{g}\mathrm{r}\mathrm{e}\mathrm{e} \mathrm{C}\mathrm{e}\mathrm{n}\mathrm{t}\mathrm{r}\mathrm{a}\mathrm{l}\mathrm{i}\mathrm{t}\mathrm{y}={\sum }_{\mathrm{i}=1}^{\mathrm{n}}{\mathrm{w}}_{\mathrm{i}},$$

(5)

In this context, ${\mathrm{w}}_{\mathrm{i}}$ represents the weight of edge i of the author node, and $\mathrm{n}$ denotes the number of edges the author node has.

4. Experimental Process

4.1. Data Acquisition and Preprocessing

A search was conducted in the Web of Science for papers in the field of artificial intelligence from 1982 to 2020, using the search string TA: (“artif* intelli*” OR “comput* intelli*” OR “deep learn*” … “image* recogn*” OR “voice translat*”) AND IPC: (“G06N3/00” OR “G06N3/02” … OR “G06N7/08” OR “G06N99/00”) [54], yielding 484,653 papers. The bibliographic information of the papers was downloaded, including authors, affiliations, titles, abstracts, references, etc.

Using Python programming, the bibliographic information was processed to select prolific first-author scientists (with 5 or more publications) from 2016 to 2020, totaling 1439 individuals, as the target subjects for expertise identification. The first author’s name and affiliation were used together to mark the author’s identity. Papers with citation counts above the average level in the field were considered the target author’s significant outcomes, resulting in 5697 focal papers. This process formed a one-to-many relationship between the first-author scientists and the identified papers for expertise domain identification.

4.2. Expertise Identification

Python programming was used to tokenize, remove stop words, lemmatize, and process N-grams in the paper titles and abstracts. Based on Formulas (1) and (2), knowledge memes in the field of artificial intelligence were identified. Considering the long-tail distribution characteristic of the meme scores, this study selected 9530 knowledge memes with scores above 0.1 as candidates for expertise tag identification. The previously described method was then used to identify the expertise tags, representing the authors’ core innovative capability backgrounds.

Retaining the frequency of tag words in the text, the Gensim package in Python was used to perform LDA topic clustering on all expertise tags. Topic coherence was calculated, and the optimal topic model was determined by plotting a coherence curve.

Using Python programs and Formulas (3) and (4), the relevance between each author’s tags and topics was calculated to construct the author–topic relevance matrix, identifying the core competency areas of scientists. A threshold of 0.1 was used to determine whether a topic is significantly related to an author’s core innovative capabilities.

4.3. Expertise Evaluation

Based on the author–topic relevance matrix, the representative expertise domains of authors were identified. Expertise evaluation within niche domains was conducted based on the authors’ professional degree, determining the relative ranking of researchers.

Using the author–topic relevance matrix and a relevance threshold of 0.1, an author–topic bipartite network was established. The Gephi 0.9.2 software was used for visual analysis of the ‘author-expertise theme’ knowledge graph and statistical analysis of network topological features, including community structure, the weighted degree centrality of each author (Formula (5)), as well as the authors’ closeness centrality, betweenness centrality, and eigenvector centrality. Based on the weighted degree centrality ranking of author nodes within specific communities, top experts with comprehensive expertise were identified.

5. Results and Interpretations

5.1. Expertise Tag Identification Results

A preliminary analysis based on the method of sampling analysis revealed that the extracted tags effectively isolate key information from papers’ representative topic terms generated by Song et al.’s method [9], reflecting the authors’ core competency backgrounds. For instance, consider the author ‘Abate, Andrea F’. from the University of Salerno, Italy.

Table 1. The tags of expertise associated with “Abate, Andrea F.; Univ Salerno-Italy”.

Author	Title and Abstract	Tag Words
Abate, Andrea F.; Univ Salerno-Italy	Paper1/I-Am: Implicitly Authenticate Me-Person Authentication on Mobile Devices Through Ear Shape and Arm Gesture Today, identity verification is required in many common activities, and it is arguably true that most people would like to be authenticated in the easiest and most transparent way, without having to remember a personal identification number. To this regard, this paper presents a multibiometric system based on the observation that the instinctive gesture of responding to a phone call can be used to capture two different biometrics, namely ear and arm gesture, which are complementary due to their, respectively, physical and behavioral nature. We conducted a comprehensive set of experiments aimed at assessing the contribution of each of the two biometrics as well as the advantage in their fusion to the system’s overall performance. Experiments also provide objective measurement of both saliency and correlation of data captured by each sensor involved (accelerometer, gyroscope, and camera) according to various features extraction, features matching, and data-fusion techniques. The reports provide evidences about the potential of the proposed system and method for user authentication “in-the-wild”, whilst its eventual usage for person identification is also investigated. All of the experiments have been carried out on a specifically built, publicly available ear-arm database, including multibiometric captures of more than 100 subjects performed during different sessions, that represents an additional contribution of this paper.	[device, experiment, saliency, number, identification]
	Paper2/Near Real-Time Three Axis Head Pose Estimation Without Training Head pose estimation methods evaluate the amount of head rotation according to two or three axes, aiming at optimizing the face acquisition process, or extracting neutral-pose frames from a video sequence. Most approaches to pose estimation exploits machine-learning techniques requiring a training phase on a large number of positive and negative examples. In this paper, a novel pose estimation method that exploits a quad-tree-based representation of facial features is described. The locations of a set of landmarks detected over the face image guide its subdivision into smaller and smaller quadrants based on the presence or lack of landmarks within each quadrant. The proposed pose descriptor is both effective and efficient, providing accurate yaw, pitch and roll axis estimates almost in real-time, without need for any training or previous knowledge about the subject. The experiments conducted on both the BIWI Kinect Head Pose Database and the challenging automated facial landmarks in the wild dataset, highlight a pose estimate precision exceeding the state-of-the-art with regard to methods not involving training and machine learning approaches.	[training, machine, example, facial feature, process, learning, video, subject, machine learning, feature, paper, method]

The bolded words represent the tag words, while the underlined words are the representative thematic words of Paper1 or Paper2.

showcases his paper titles, abstract content, and extracted tags. In the table, bolded words correspond to the tag concepts in the text, while underlined words represent the representative thematic words of the outcomes. A comparison reveals that the core competency feature words predominantly originate from the document’s thematic words. Moreover, they effectively filter out general thematic words such as ‘arm’, ‘gesture’, and ‘biometric system’, instead highlighting the key research methods and problem areas discussed in the papers.

For example, in Paper 1, the feature words ‘experiment’ and ‘significance’ indicate the key research methods involved, while ‘device’ and ‘identification number’ point to the key problem areas addressed. In Paper 2, terms like ‘machine learning’, ‘training’, ‘facial features’, and ‘video’ suggest that the key methods and problem areas involve video data analysis using feature representation and machine learning. These core competency tags clearly reflect the essential skills and knowledge background that the authors relied upon in their innovative activities, providing an understanding of their core innovative competencies.

5.2. Topic Clustering Results

The optimal LDA topic model was determined based on the coherence score curve. As shown in Figure 2, the coherence score initially increases and then decreases with the number of topics. A higher coherence score indicates more semantically coherent words within the topic, reflecting better interpretability and quality. Consequently, the experiment identified 27 core competency themes based on the highest point of the coherence curve. The content of the first five themes is shown in

Table 2. Top 5 expertise topics.

Topics	Topic Words	Content Summary
Topic0	data, cloud, service, system, energy, user, resource, …	Cloud Computing and Data Management: Involves data storage, cloud services, resource allocation, user access, energy efficiency, security algorithms, and performance optimization.
Topic1	network, temperature, neural network, model, artificial neural network, result	Building Energy Efficiency and Material Studies: Examines the correlation between indoor temperature, artificial neural network models, carbon emissions, thermal conductivity, and fluid optimization parameters.
Topic2	fault, diagnosis, method, fault diagnosis, signal, feature	Fault Diagnosis and Machine Health Monitoring: Utilizes signal processing, feature learning, neural networks, and deep learning techniques for fault detection, vibration analysis, and equipment health monitoring.
Topic3	traffic, vehicle, manufacturing, system, road, stream	Traffic Flow and Intelligent Transportation Systems: Investigates traffic flow dynamics, vehicle behavior, road network predictions, and data analysis in multimodal intelligent transport systems.
Topic4	model, process, network, problem, analysis, design	Network Modeling and Performance Analysis: Encompasses network design, problem-solving, parameter calibration, model accuracy, and performance evaluation.

. The identified themes are closely related to the core content of the AI field and include multiple areas such as healthcare, architecture, and transportation, reflecting the diverse core competency backgrounds of researchers and the multidisciplinary integration characteristic of AI research.

Compared to the optimal topic model identified using papers’ representative topic terms, which only includes 15 research themes, our topic model offers clear advantages in granularity and content frontiers. For example, the ‘Fault Diagnosis and Signal Processing’ theme identified by traditional methods focuses on the basic principles and techniques of fault diagnosis, while the ‘Fault Diagnosis and Mechanical Health Monitoring’ theme identified by our method emphasizes the application of modern machine learning and deep learning technologies for more advanced fault detection and health monitoring. The advantages of these cutting-edge research topics are supposed to be more important, due to the extracted label words representing the core knowledge to promote innovation.

5.3. Expertise Domain Identification Results

The author–topic relevance matrix reveals the expertise domain of each researcher and semantically validates the accuracy of the topics in representing the researchers’ core competency fields. A portion of the author–topic relevance matrix results are shown in

Table 3. The author–topic relevance matrix.

Author	Topic0	Topic1	Topic2	Topic3	Topic4	Topic5	Topic6	Topic7	Topic8
Abate, Andrea F.; Univ Salerno-Italy	0.067	0.033	0.167	0.067	0.067	0.067	0.100	0.100	0.133
Abdel-Nasser, Mohamed; Univ Rovira and Virgili-Egypt	0.033	0.000	0.000	0.033	0.033	0.000	0.033	0.000	0.000
Abdulridha, Jaafar; Univ Florida-USA	0.033	0.000	0.100	0.067	0.033	0.067	0.033	0.033	0.067
Abdurahman, Abdujelil; Xinjiang Univ-China	0.000	0.000	0.000	0.000	0.000	0.000	0.000	0.000	0.000
Abed, Saed; Kuwait Univ-Kuwait	0.000	0.067	0.067	0.033	0.067	0.033	0.033	0.100	0.033

. Using the method of sampling analysis, authors with distinct core competency field and significantly different topic relevance were selected, and their top 3 ranked topics were analyzed. This analysis found that the identified thematic domains accurately reflect the authors’ core innovative capability fields.

For example, the expertise domain ranking for the author ‘Abate, Andrea F./Univ Salerno/Italy’ is shown in Figure 3. The top three topics for this author are as follows: ‘Topic 26/Data Classification and Deep Learning’, which involves the application of data classification, feature learning, and deep learning networks in sample representation and performance evaluation; ‘Topic 2/Fault Diagnosis and Mechanical Health Monitoring’, which uses signal processing, feature learning, neural networks, and deep learning methods for fault detection, vibration analysis, and health monitoring; and ‘Topic 13/Machine Learning and Regression Analysis’, which employs machine learning methods for data prediction and modeling, including soft sensors, support vector regression, and uncertainty estimation. Through comparison with the author’s textual content (as shown in Table 1), it is observed that the content of these topics demonstrates a close relation to the author’s core knowledge and skills, with a particular emphasis on Topic 26.

Compared to expertise domains identified by traditional methods, those based on core innovative capabilities provide more finely grained and deeper reference information for peer review and the organization of research activities. For instance, the representative expertise domains for the author ‘Abate, Andrea F./Univ Salerno/Italy’, identified through the thematic words of papers and topic relevance calculations, include ‘Feature Recognition and Facial Recognition’, ‘Image Features and Network Learning’, and ‘Data Classification and Learning Algorithms’. Although these themes strongly correlate with the author’s core innovative capability domain, ‘Data Classification and Deep Learning’, identified by our method, the differentiation among the content of each topic is relatively low, offering a somewhat limited representation of the author’s ability profile. This suggests that while traditional methods can capture general areas of expertise, they might not fully reflect the depth and accuracy of a researcher’s knowledge capability.

The results of the core competence expertise identified by this study consist of ‘Data Classification and Deep Learning’ and uncover other potential research domains closely aligned with the author’s core innovative capacity, such as ‘Fault Diagnosis and Machinery Health Monitoring’ and ‘Machine Learning and Regression Analysis’. This indicates that the author has the potential for innovation within these domains through learning or collaborative efforts. Furthermore, given that innovative activities rely more on core competencies than on general knowledge, the potential innovation areas identified in this study are more precise. It becomes evident that research specialties rooted in core competencies not only enable a more refined and in-depth profiling of a researcher’s individual traits but more accurately reveal the potential innovation directions closest to the author’s core competence. This approach provides a more accurate and comprehensive understanding of a researcher’s potential, thus supporting more effective and targeted research development and collaboration.

5.4. Expertise Evaluation Results

5.4.1. Peer Evaluation in Niche Domains

For researchers with distinct topic relevance, the top 1–3 ranked topics are identified as their representative expertise domains. Within the same domain, research peers are ranked by their professional degree to identify leading experts. Figure 4 shows the ranking of peer authors in ‘Topic 2’, demonstrating that this method has a high degree of differentiation for top-ranked experts. This approach prevents experts with less domain core knowledge from evaluating peers with more, facilitating the recognition and support of innovative projects with certain risks.

5.4.2. Expertise Evaluation Based on Comprehensive Expertise

Considering the need for multi-skilled talent in complex research tasks, an accurate and comprehensive evaluation of researchers’ expertise is essential. Moreover, in an environment where talent is scarce, well-rounded researchers without a specific representative expertise also hold significant value. Hence, identifying the academic relationships of researchers within large innovation communities based on the author-expertise knowledge graph can complement the expertise evaluation results in niche domains.

The knowledge graph reveals that the field of AI can be divided into seven communities with relatively close relationships. Authors within the same innovation community have similar expertise, facilitating easier knowledge sharing and a higher potential for collaborative innovation. Authors with multiple expertise domains within a community, serving as hub nodes, can accelerate the fusion of different thematic knowledge for innovation, making them potentially more important than experts with a single expertise. For example, the innovation community composed of ‘Topic 6’ and ‘Topic 10’ is shown in Figure 5. Among them, the hub node ‘Rahmani, Mostafa; Univ Cent Florida-USA’ (red colour) with a weighted degree centrality of 0.2 has higher community importance than the community peer ‘Tsai, Ching-Chih;Natl Chung Hsing Univ-Taiwan’ (red colour). However, scientists with outstanding professional degrees in a niche domain (Topic 6), such as ‘Li, Xiaodi; Shandong Normal Univ-China’ (weighted degree centrality = 0.23, red colour), can still be identified as important experts within the community.

The top 10 core experts in the field of AI in the period of 2016–2020 based on comprehensive expertise and their relative ranking are shown in

Table 4. Top 10 domain experts with comprehensive expertise.

Author	Weighted Degree Centrality	Closeness Centrality	Eigenvector Centrality	Betweenness Centrality
Hossain, M. Shamim; King Saud Univ-Arabia	9.6300	0.5080	0.5158	750.2046
Chen, Yuantao; Changsha Univ Sci & Technol-China	9.2800	0.5080	0.5158	750.2046
Wang, Danshi; Beijing Univ Posts & Telecommun-China	8.8000	0.5080	0.5158	750.2046
Sun, Wei; North China Elect Power Univ-China	8.6200	0.5080	0.5158	750.2046
Lopez-Martin, Manuel; Univ Valladolid-Spain	7.7600	0.5080	0.5158	750.2046
Nasir, Vahid; Univ British Columbia-Canada	7.5900	0.5080	0.5158	750.2046
Liang, Liang; Georgia Inst Technol-USA	7.1600	0.5080	0.5158	750.2046
Wang, Tian; Huaqiao Univ-China	7.0200	0.5080	0.5158	750.2046
Zhang, Qingxue; Univ Texas Dallas-China	6.9200	0.5080	0.5158	750.2046
Yin, Yuyu; Hangzhou Dianzi Univ-China	6.7300	0.5032	0.5125	620.2647

. The ranking result based on the weighted degree centrality is more precise than node importance metrics such as closeness centrality, eigenvector centrality, and betweenness centrality. Additionally, the consistency of the various ranking results indicates that more important experts have higher knowledge influence and information control power within the entire peer relationship network. Moreover, in line with common sense, these core experts all come from leading positions in niche domains. The expertise evaluation metrics of this study can accurately identify the relative ranking of domain researchers based on their comprehensive expertise.

6. Discussion

Based on the proposed text mining algorithm, we identified tags representing the core innovative capabilities of researchers in the AI field, which reflect the key research methods and problem areas in innovative activities and reveal a deeper understanding of researchers’ expertise compared to traditional methods. The results of this analysis show that researchers in the AI domain possess a multidisciplinary core knowledge background, effectively uncovering the characteristics of the core competency areas of AI innovators. This comprehensive insight into the foundational expertise of AI researchers has significant theoretical value for innovation management and the formulation of science and technology policies in the field. By highlighting the areas of knowledge and skill that drive innovation, this approach not only enhances our understanding of individual and collective research strengths but informs strategic decisions that support the advancement of AI. Policymakers and research managers can leverage these findings to better allocate resources, foster collaboration, and design programs that nurture the most promising areas of AI research. Additionally, the ability to accurately identify and understand the core innovative capabilities of AI scientists provides a viable framework for anticipating future trends and developments in the field, ensuring that efforts are aligned with the cutting-edge advancements that define the evolving landscape of artificial intelligence.

The themes of AI scientists’ core innovative capabilities exhibit more cutting-edge content than those identified by thematic words from the text content, offering substantial reference value for revealing the characteristics of innovation evolution and predicting future directions of innovation. By focusing on the core innovative capabilities, we gain deeper insights into the forefront of AI research and development, which traditional thematic analysis might overlook. This approach allows for a more precise identification of the areas where AI researchers are pushing the boundaries of technology and knowledge. As a result, it becomes possible to trace the historical trajectory of innovation within the field and to forecast emerging trends and potential breakthroughs. The advanced themes highlighted by core capabilities serve as a beacon, guiding researchers and stakeholders towards the most promising and transformative areas of AI. Consequently, this method significantly enhances our understanding of the dynamic landscape of AI innovation, providing a robust framework for strategic planning, funding allocation, and fostering collaborative efforts that are aligned with the future direction of technological advancements.

The themes of AI researchers’ core innovative capabilities identified by the study are more granular than the research themes derived from thematic words in the text content, revealing a broader array of potential research areas for scientists. This granular approach uncovers more specific and varied research topics and highlights areas where researchers are more likely to produce innovative outcomes. By pinpointing these detailed and closely related fields, this study provides a valuable theoretical reference for organizing innovation teams and selecting innovative projects. This, in turn, enhances the sustainable innovation capacity of organizations, institutions, and nations within the AI field. By focusing on the precise areas where AI researchers exhibit core innovative capabilities, it becomes possible to strategically align resources and collaborative efforts to strengthen greater creativity and breakthrough developments. The insights gained from this granular analysis inform better decision-making and policy formulation, ensuring that innovation initiatives are closely matched with the inherent strengths and potential of the research community. Consequently, this approach supports the development of a more robust and dynamic innovation ecosystem, capable of driving sustained progress and maintaining a competitive edge in the rapidly evolving landscape of artificial intelligence.

The top experts in the AI field, identified based on their core competency backgrounds, can offer deeper insights for innovation teams seeking specialized knowledge. These experts, possessing analogous core innovative capabilities, are more likely to engage in knowledge sharing and exchange, thereby fostering high-quality innovation. Their deep-seated expertise enhances the collaborative potential within innovation teams and ensures that the knowledge shared is both relevant and cutting-edge. Additionally, experts who bridge different communities may serve as discoverers of new and significant innovative problems, leveraging their unique perspectives to identify gaps and opportunities that might otherwise go unnoticed. This cross-pollination of ideas between diverse expert communities can lead to groundbreaking solutions and novel approaches, driving the AI field forward. By incorporating these top experts into innovation teams, organizations can cultivate an environment conducive to continuous learning and creativity, ultimately enhancing their capacity for sustainable innovation. This strategic alignment of expertise and collaborative effort is important to maintain a competitive edge and advance the frontiers of artificial intelligence.

7. Conclusions

Under the backdrop of Industry 4.0, the convergence and integration of artificial intelligence technology with other technical domains is closely intertwined with the technological advancements across various industries. Consequently, bolstering the sustainable innovation capabilities of AI technology is of paramount importance to a diverse array of market entities. Precisely identifying the expertise of researchers serves as the cornerstone for effectively marshaling innovative talent, which is pivotal for enhancing sustainable innovation capabilities. However, the intricate knowledge backgrounds of AI researchers underscore the current methods’ limitations in accurately identifying their capability backgrounds. In light of this, our study, grounded in theoretical analysis, introduces a method for deeply mining the core innovative capability fields of researchers from their textual content and applies empirical research within the AI field. Building upon the results of tag word identification, we employ topic clustering and author–topic relevance analysis to discern specialized sub-domains and the relative standing of experts with different professional degree. And we validate the objectivity, accuracy, and effectiveness of our methodology. The evaluation of experts encompasses two dimensions: representative expertise capabilities and comprehensive expertise capabilities. The experimental outcomes indicate that our proposed method can effectively pinpoint the core innovation areas of researchers and differentiate their relative standings based on the professional degree. Theoretically, this approach is equally applicable to the in-depth identification and differentiation of the capability backgrounds of researchers in technological fields beyond AI.

In contrast to conventional methods, this novel approach offers significant advantages for the identification and assessment of research expertise capabilities. The themes identified are more nuanced and avant-garde, offering a more valuable portrayal of researchers’ expertise background and unveiling potentially innovative domains. It facilitates the precise and agile recommendation of suitable experts for intricate research organizational endeavors. By concentrating on the thematic domains most pertinent to researchers’ core innovative capabilities, this study establishes a foundation for future improvements. The model can be further enhanced by incorporating factors such as the quantity of outcomes, influence, and innovation indices, thereby crafting a more holistic framework for expertise identification and evaluation.