Navigating the AI technology landscape from GitHub data

Abstract

As artificial intelligence (AI) is considered a pivotal technology determining competitiveness, understanding the current and future state of AI technology has become crucial. Conventional approaches to mapping the technology landscape have relied heavily on patent data, but patents cannot adequately capture the state of the art in rapidly changing technologies like AI, due to significant time lags from development to registration. Given that much of the AI technology is developed through open source projects on GitHub, the largest and most popular code host and social coding platform, GitHub emerges as a promising data source for navigating the AI technology landscape. This study aims to explore and predict the AI landscape based on GitHub data. We propose a new bibliometric-like measure, called library coupling, which leverages the unique aspect of code reuse in open source software development to capture the relationships between GitHub repositories. A total of 2879 AI-related repositories with Python-based libraries were collected from GitHub. An AI repository network is constructed based on library coupling relationships among these repositories. Using the attributed graph clustering technique, the AI repositories within the network are grouped into 20 AI technology clusters. Subsequently, we employ graph convolutional network-based link prediction to predict the changes in the AI technology landscape. The proposed GitHub-based technology landscaping approach can be effectively utilized to grasp the current state of rapidly evolving AI technologies and predict their future trends, thereby supporting informed decision making in national AI policy formulation and corporate AI strategy.