Neural embeddings unveil simplicity in complex systems

The cornerstone of any complex systems analysis, whether it is the Internet, social networks, or biological organisms, is the effectiveness of their representation. While traditional methods utilize discrete representations such as networks, recent neural network advancements provide continuous representations, or embeddings. Neural embeddings map entities into a vector space, enabling new operationalization of abstract inquiries.

However, the ``black box'' nature of neural networks makes these embeddings difficult to interpret, posing a challenge for their use in generating rigorous scientific understanding. In this talk, we will address this issue by linking neural embeddings with interpretable, tangible physical measurements. We will delve into the mechanics of neural embeddings, using community detection tasks as a practical example. Then, we will focus on human mobility in science, showcasing a robust and implicit connection between embedding distance and human mobility flow, an interpretable, tangible quantity. Finally, we will illustrate how neural embedding can reveal new perspectives on the structure of complex systems by analyzing the citation dynamics in science, law, and patents.