Beyond the noise: Early identification of emerging technologies through weak signal prediction in higher-order networks

The early identification of emerging technologies from weak signals is a critical strategic imperative. Existing network-based methods remain largely confined to lower-order representations and static, snapshot-based conceptualizations of weak signals, emphasizing retrospective detection over prospective prediction. Therefore, this study constructs a multi-source heterogeneous hypergraph with higher-order semantics, spanning the “Science-Technology-Industry” dimensions. By mapping the structural phase transitions of weak signals across an “Outlier-Mutation-Key Node” trajectory, we devise a quantifiable indicator system to shift the analytical paradigm from static retrospective detection to dynamic evolutionary tracking. Using a dataset of 145,316 papers, 79,050 patents, and 5941 project records from the New Energy Vehicle (NEV) fuel cell domain (2015–2024), we identify 226 candidate nodes exhibiting this trajectory. A two-round Delphi expert evaluation confirms 193 of these nodes as expert-confirmed weak signals, establishing Delphi-validated reference labels for subsequent model evaluation. We then employ Multi-Layer Perceptron (MLP), Dynamic Hypergraph Convolutional Network (DHCN), and Large Language Models (ZS-LLM and ICL-LLM) to prospectively predict weak signals with disruptive potential. The ICL-LLM achieves the best predictive performance among the tested models, with a Macro-F1 score of 61.25%, outperforming the MLP (49.76%), DHCN (56.84%), and ZS-LLM (46.35%). This synergy between higher-order topology and semantic intelligence effectively filters stochastic noise and enables meaningful predictive capability of emerging technologies, providing an exploratory decision-support framework for technology foresight.