AI Research Trends Mapped Through Decade of Conference Data
In the rapidly evolving landscape of artificial intelligence, staying ahead of technological shifts is no longer a luxury—it’s a necessity. As global powers invest heavily in AI for both civilian and military applications, understanding the trajectory of research has become critical for policymakers, industry leaders, and academic institutions alike. A recent study led by Zhao Yucheng, a graduate researcher at Shenyang Aerospace University, in collaboration with Associate Professor Chen Jianjun from Shenyang Northern Software College of Information Technology, has taken a comprehensive approach to mapping the intellectual contours of AI research over the past decade. By constructing a detailed knowledge graph from thousands of scholarly papers, the team has unveiled patterns, trends, and key influencers shaping the future of artificial intelligence.
The study, published in Computer & Digital Engineering, leverages data from two of the most prestigious conferences in the field: the Association for the Advancement of Artificial Intelligence (AAAI) and the International Joint Conference on Artificial Intelligence (IJCAI). These forums have long served as incubators for groundbreaking ideas, from early expert systems to modern deep learning architectures. By systematically analyzing over 10,000 papers presented between 2000 and 2018, the researchers have created a structured, interconnected map of AI research that goes far beyond traditional bibliometric analysis.
At the heart of this work is the construction of a domain-specific knowledge graph—a semantic network that organizes entities such as researchers, publications, keywords, and research topics into a machine-readable format. Unlike general-purpose knowledge graphs like Google’s or Wikidata, which aim for broad coverage, this project focuses narrowly on the AI domain, allowing for deeper insights and more precise trend detection. The resulting graph contains half a million semantic triples—structured statements in the form of subject-predicate-object—that capture relationships such as “Zhi-Hua Zhou researches deep learning” or “reinforcement learning is related to game theory.”
The methodology employed in this study blends top-down and bottom-up approaches to knowledge graph construction. The top-down component involves defining an ontology that outlines key entity types and their relationships. In this case, the core classes include “Domain Literature,” “Domain Researchers,” and “Research Hotspots.” Each paper is treated as a node with attributes such as title, abstract, authors, keywords, publication year, and conference source. Relationships are then established between these nodes: authors are linked to papers through authorship, papers are connected to research topics via keywords, and researchers are tied together through co-authorship networks.
On the data side, the process is largely bottom-up. The team developed a web crawler to systematically extract metadata from the official proceedings of AAAI and IJCAI. While some papers included structured keywords, many did not—posing a significant challenge for topic modeling. To address this, the researchers implemented a hybrid keyword extraction strategy using two natural language processing algorithms: TextRank and RAKE (Rapid Automatic Keyword Extraction). Both methods analyze text to identify salient terms based on statistical and syntactic features, but they operate differently. TextRank uses a graph-based ranking model similar to Google’s PageRank, where words are nodes and edges represent co-occurrence within a sliding window. RAKE, on the other hand, relies on stopword delimiters to identify candidate phrases and scores them based on word frequency and degree of co-occurrence.
One of the most significant findings of the study was that relying solely on abstracts for keyword extraction led to poor recall. A staggering 91% of manually assigned keywords in the dataset did not appear in the abstracts, a phenomenon the authors term “unregistered keywords.” However, these missing terms were often present in the paper titles. Recognizing that titles encapsulate the core contribution of a study, the team combined titles and abstracts as input for keyword extraction. This simple yet effective adjustment significantly improved the accuracy of the extracted terms. Comparative experiments showed that using both title and abstract consistently outperformed using abstract alone, regardless of the algorithm used. For instance, in 2015, TextRank extracted 1,153 correct keywords when using both title and abstract, compared to only 1,043 when using abstract alone. Similarly, RAKE’s performance increased from 1,067 to 1,179 correct extractions in the same year.
This finding has broader implications for automated literature analysis. It suggests that in scientific domains where brevity and precision are valued, important semantic content may be concentrated in titles rather than abstracts. Therefore, future NLP pipelines for academic mining should treat titles not as mere metadata but as rich sources of conceptual information.
Once the keywords were extracted and validated, the researchers integrated them into the knowledge graph, establishing connections between research topics, authors, and time periods. This allowed for dynamic querying of the dataset—for example, identifying the most active researchers in a given year or tracking the evolution of a specific subfield over time.
One of the most revealing analyses involved identifying the most prolific authors in AI research. The top ten contributors each published more than 50 papers between 2000 and 2018. Leading the list was Tuomas Sandholm from Carnegie Mellon University, known for his work in game theory and automated negotiation. Other frequent contributors included Feiping Nie, Qiang Yang, and Zhi-Hua Zhou—all prominent figures in machine learning and data mining. The presence of these individuals at the top of the productivity rankings underscores the importance of sustained, long-term engagement in high-impact research.
Beyond mere publication counts, the knowledge graph enabled deeper social network analysis. By mapping co-authorship relationships, the researchers could visualize collaboration patterns within the AI community. For instance, they identified clusters of researchers working in reinforcement learning, multi-agent systems, and computer vision. More importantly, the graph revealed potential collaboration opportunities—researchers who share common interests but have never co-authored a paper. These “latent collaborations” represent untapped synergies that could accelerate innovation if properly facilitated.
The temporal dimension of the graph also provided valuable insights into the evolution of research interests. By aggregating keyword frequencies across years, the team identified both persistent and emerging themes. Central nodes in the keyword network—such as “machine learning,” “reinforcement learning,” “game theory,” “planning,” and “crowdsourcing”—have remained consistently prominent, indicating stable pillars of AI research. These areas represent the foundational technologies that continue to underpin advances across the field.
In contrast, peripheral nodes reflect more transient or rapidly growing areas. Notably, terms like “deep learning,” “neural networks,” and “classification” gained significant traction in the mid-2010s, coinciding with the deep learning revolution sparked by breakthroughs in image and speech recognition. The surge in these topics illustrates how technological milestones can rapidly shift the focus of an entire discipline. What began as a niche area of study became a dominant paradigm within a few years, absorbing talent, funding, and attention from adjacent fields.
The rise of deep learning also highlights a broader trend: the increasing interdisciplinarity of AI research. Many of the top authors in the study have backgrounds that span computer science, statistics, cognitive science, and engineering. Their work often bridges theoretical foundations with practical applications, from autonomous systems to healthcare analytics. This convergence is reflected in the structure of the knowledge graph, where topics like “transfer learning,” “federated learning,” and “explainable AI” sit at the intersection of multiple subfields.
Another observation is the growing emphasis on human-AI interaction. Keywords such as “crowdsourcing,” “human-in-the-loop,” and “preference learning” suggest a shift toward systems that are not only intelligent but also collaborative and interpretable. This reflects a maturation of the field—from building algorithms that perform well in isolation to designing systems that function effectively in real-world contexts with human users.
The geographical distribution of authors, while not explicitly analyzed in the paper, can be inferred from institutional affiliations. A significant portion of the top contributors are affiliated with North American and Chinese universities, reflecting the global competition in AI research. Institutions like CMU, Tsinghua University, and the Chinese University of Hong Kong appear frequently, indicating strong research ecosystems in both the U.S. and China. This aligns with national strategies—such as the U.S. National AI Research and Development Strategic Plan and China’s New Generation Artificial Intelligence Development Plan—that prioritize AI as a strategic technology.
From a methodological standpoint, the success of this study demonstrates the power of knowledge graphs in synthesizing complex, unstructured data into actionable insights. Traditional literature reviews are limited by human cognitive capacity and publication bias. In contrast, a machine-readable knowledge graph can scale to tens of thousands of documents, uncover hidden patterns, and support interactive exploration. It transforms static archives into dynamic knowledge bases that evolve with new inputs.
Moreover, the open-ended nature of the graph allows for multiple types of queries. Policymakers could use it to identify national strengths and gaps in AI research. Funding agencies might leverage it to detect emerging fields worthy of investment. Educators could design curricula based on the most relevant and enduring topics. And individual researchers could discover collaborators or explore adjacent domains for inspiration.
The implications extend beyond academia. In industry, knowledge graphs are already used for competitive intelligence, talent acquisition, and R&D planning. A domain-specific graph like the one built by Zhao and Chen could help tech companies anticipate technological shifts, scout for experts, or identify potential acquisition targets. For example, a sudden increase in publications on “neural architecture search” or “AI ethics” might signal a strategic opportunity or risk.
Despite its strengths, the study has limitations. The reliance on conference papers means that journal articles, preprints, and non-English publications are underrepresented. AAAI and IJCAI, while prestigious, do not capture the full spectrum of AI research—especially in applied domains like robotics or natural language processing, where other venues may be more dominant. Additionally, keyword extraction, even with title-abstract fusion, is imperfect. Some nuanced concepts may be missed, and synonymy (e.g., “neural net” vs. “neural network”) can lead to fragmentation in the graph.
Future work could address these issues by incorporating data from additional sources such as arXiv, IEEE Xplore, and SpringerLink. Advanced entity linking techniques could normalize variations in terminology, and citation analysis could enrich the graph with influence metrics. Integrating temporal embeddings or dynamic graph neural networks might also allow for more sophisticated forecasting of research trends.
Nonetheless, the current study represents a significant step forward in the systematic analysis of scientific progress. It exemplifies how artificial intelligence can be used to understand itself—a meta-application of the technology. By turning the lens of data science on the scientific process, researchers can gain a clearer picture of where the field has been and where it might be going.
As AI continues to transform society, the ability to navigate its intellectual landscape will only grow in importance. Tools like knowledge graphs offer a way to cut through the noise of information overload and focus on what truly matters. They enable a more strategic, evidence-based approach to research and development—one that is essential for maintaining competitiveness in the 21st century.
The work of Zhao Yucheng and Chen Jianjun serves as both a technical achievement and a model for future scholarship. It shows that with the right data, methods, and vision, we can build not just intelligent machines, but intelligent maps of human knowledge.
Zhao Yucheng, Chen Jianjun, Computer & Digital Engineering, DOI: 10.3969/j.issn.1672-9722.2021.03.019