Combining LLM and DIDEX method to predict Internal Migrations in Serbia

Abstract

In this study, we explore the integration of vast knowledge from large language models to enhance the data-induced decision expert (DEX) model’s ability to understand and forecast internal migrations in Serbia. We combine LLMs with Decision Support Systems (DSS), specifically focusing on data-induced decision expert (DIDEX) methodology, to significantly improve attribute selection, and model interpretation, which is vital for making informed decisions. The fundamental idea of this paper involves utilizing LLMs to define a hierarchy of attributes and using DIDEX on data utilizing those attributes to generate the necessary decision rules for the DEX model. The proposed DSS is enabling policymakers in evaluating the impact of various potential municipal interventions for sustainable internal migrations across Serbian local self governments. A comparative analysis of traditional machine learning models and DIDEX was performed, utilizing both GPT 3.5 Turbo and GPT 4 Turbo. The findings indicate that LLMs achieve results comparable to machine learning models while inheriting the advantages of DEX models. More specifically, classification accuracy is around 64%, while the DIDEX model achieves 65%. While the time needed to create a DIDEX model reduced significantly by using LLMs the interpretation of the obtained DEX model was highly increased.

Sandro Radovanović

Assistant Professor at University of Belgrade

My research interests include machine learning, development and design of decision support systems, decision theory, and fairness and justice concepts in algorithmic decision making.