Profiling and Prioritizing Export Success in a Technopark with LLM-Based Feature Extraction, Explainable Machine Learning, and Fuzzy MCDM

Abstract

Science and technology parks function as primary instruments of regional innovation policy designed to accelerate economic development. However, the evaluation literature predominantly estimates average historical park effects rather than identifying specific tenant firms requiring targeted public support, leaving park administrators without objective tools for allocating scarce resources. The purpose of this research is to address this resource allocation problem by developing a data-driven decision support framework that systematically profiles and prioritizes tenant firms based on their empirical export potential. The methodology applies a large language model to extract technological complexity metrics from 557 unstructured administrative project narratives. It then utilizes explainable machine learning classifiers to predict firm-level export success and translates the resulting algorithmic feature attributions into objective criterion weights for a fuzzy multi-criteria prioritization model. The principal results indicate that park-specific institutional tenure and realized research revenue are the dominant predictors of commercial export success. The regularized predictive model achieves strong out-of-sample classification performance, while the extracted textual innovation scores provide additional predictive information beyond traditional headcount metrics. The results demonstrate that integrating natural language processing with explainable artificial intelligence provides a data-driven alternative to subjective expert judgments in resource allocation. The resulting prioritization framework provides policymakers with an auditable and reproducible tool for effectively targeting internationalization grants and optimizing the distribution of public innovation incentives.