Cognition-enhanced geospatial decision framework integrating fuzzy FCA, surprisingly popular method, and a large language model

Abstract

This study introduces a cognition-enhanced framework for geospatial decision-making by integrating Fuzzy Formal Concept Analysis (FCA), the Surprisingly Popular (SP) method, and a Large Language Model (GPT-4o). Our approach captures cognitive influences that are often overlooked in traditional geospatial analyses. Fuzzy FCA is used to extract interpretable concept hierarchies from spatial data, while the GPT-4o model estimates SP scores, identifying choices that reflect underlying cognitive biases. These cognitively informed features are combined within machine learning models, improving both prediction accuracy and interpretability. Experiments on real-world urban mobility and environmental risk scenarios demonstrate significant performance gains, with models like XGBoost achieving an accuracy of 0.8412. We also introduce a novel method for evaluating the cognitive validity of LLM-generated model explanations, which involves assessing how well these explanations align with human intuition and reasoning. Our results show that incorporating cognitive elements into geospatial models not only improves outcomes but also bridges the gap between data-driven predictions and human decision-making. This framework offers broad potential for applications in GIS, urban planning, and environmental management.

Keywords: A large language model; Cognitive bias; Formal concept analysis; Geospatial decision-making; Surprisingly popular method.

Fig. 1

An illustration of logistics company’s challenge in planning optimal delivery routes in a complex urban environment, balancing efficiency, risk mitigation, and access to critical points of interest while avoiding hazardous zones. Map generated by the authors using uMap 3.0.6 (https://umap.openstreetmap.fr/en/)

Fig. 2

Visual comparison of model performance metrics across the Urban Mobility and Environmental Risk scenarios.

Fig. 3

User study process evaluating LLM-generated explanations.

Algorithm 1

FCA_Prompt_CoT_SC: Extracting Concept Labels with GPT-4o and Chain-of-Thought

Algorithm 2

SP_Score_CoT_SC: Estimating Surprisingly Popular Scores via LLM Reasoning

Fig. 4

Integrated framework combining geospatial data processing, fuzzy FCA, SP method, LLM-based labeling, and machine learning models.

Fig. 5

FCA analysis example for 10 routes.