An Empirical Study of Artifacts in AI Research Ecosystem

Abstract

The increasing complexity of artificial intelligence and its widespread applications have prompted urgent needs for empirical studies of the AI research ecosystem. This dissertation explores the complex structure of the AI research ecosystem through empirical network analysis in three different dimensions. In the first work, we construct an AI research space to capture the intellectual proximity between research topics, revealing how country-specific patterns of specialization influence high-impact scientific research outcomes. Our findings suggest that countries develop expertise in topics that are closely related to their existing strengths, and that papers aligned with a country's specialization topics are more likely to have high impact.

Second, we analyze the technological impact of code availability in AI research by examining patent citations in academic papers. Using a comprehensive dataset of 250,000 AI papers and USPTO patents, we find that papers with accessible code implementations receive more patent citations, exhibit broader technical impact across domains, have a shorter time lag to first citation, and maintain a more persistent impact in industrial applications.

Finally, we conducted a large-scale analysis of 170,000 machine learning repositories to reveal the hidden structure of successful implementation practices. By constructing a network of package co-occurrences, we identified a set of core base packages and measured the novelty of the repositories by a combination of atypical packages. Our results show that repositories with moderate dependence on core packages and high novelty of package combinations are the most popular.

Date of Award	12 Nov 2025
Original language	English
Awarding Institution	City University of Hong Kong
Supervisor	Qing KE (Supervisor), Jian Hua Jonathan ZHU (Co-supervisor) & Dingxuan ZHOU (Supervisor)