Abstract
In this paper, we present an illustration of the history of Artificial Intelligence (AI) through a statistical analysis of publications since 1940. We collected and mined the IEEE publication database to analyze the geographic and chronological variance of AI research activity. The connections between different institutes are mapped, revealing that the leading communities of AI research are primarily situated in the USA, China, Europe, and Japan. Key institutes, authors, and research hotspots are identified. Our findings indicate that research communities in fields like Data Mining, Computer Vision, and Pattern Recognition are highly consistent, implying strong cross-field interaction. Additionally, Electronic Engineering and industrial applications are particularly active in California, while Japan remains a leader in robotics. While limited by the scope of the IEEE database, this study improves upon simple publication counts by applying network keynode analysis to the research community.
Introduction
Most histories of Artificial Intelligence are narrative-driven, focusing on landmark events like the Dartmouth Conference or the "AI Winters." While compelling, these accounts often lack quantitative backing. This study provides a data-driven "biography" of AI by performing a network analysis on over 600,000 IEEE publications spanning 1940 to 2013. By using bibliometrics and network science, we identify the leading players, map the geographic shifts in research power, and reveal the interconnectedness of various sub-fields.
Methods and Data
The analysis relies on the IEEE Xplore database, comprising 610,051 articles published between 1940 and October 2013. The methodology involved:
- Data Retrieval & Normalization: Metadata extraction via the IEEE Xplore XML API, followed by text mining in R to unify synonyms and resolve affiliations.
- Geocoding: Mapping affiliations to coordinates using the Google Maps Geocode API to visualize global hubs.
- Network Construction: Building undirected graphs for co-authorship (nodes as authors, edges as shared papers) and keyword co-occurrence.
- Community Detection: Utilizing a random walk algorithm to cluster 1,000 popular terms into eight distinct sub-fields, such as Robotics and Machine Learning.
Key Insights and Results
The Scale-Free Nature of AI Research
The co-author network exhibits a power-law distribution, characteristic of "scale-free" complex networks. This implies the existence of "super-connectors"—highly prolific authors who anchor the network—while the majority of researchers have few connections. Interestingly, the data reveals a "Rule of Three": an abnormal decrease in authors with a degree of 2 to 4 suggests that AI research is predominantly conducted in groups of three or more rather than solo or in pairs.
Global Research Landscape and Specialization
AI publications have grown exponentially over the last seven decades. Geographically, the USA and China have emerged as the primary partners in global research cooperation. While California (specifically the University of California system) hosts the largest concentration of productive researchers, Tsinghua University stands out as the most productive single institution in the dataset.
Regional specializations are also evident:
- California: A powerhouse for Electronic Engineering and commercial applications.
- Japan: A leader in Robotics, maintaining strong academic ties with the US and Europe despite geographic proximity to China.
- Interconnected Fields: Communities in Data Mining, Computer Vision, and Pattern Recognition show high consistency and heavy interaction, whereas fields like Robotics remain more specialized within specific geographic hubs.
Limitations and Future Work
The study's reliance on the IEEE database may under-represent fields that favor other venues like ACM or arXiv. Furthermore, the current model treats all publications equally, regardless of impact. Future research should incorporate citation maps and algorithms like PageRank to weight papers by their actual influence on the field rather than simple volume.
Summary updated