A Neural Network-Based Decision Framework Utilizing Large Language Models and Neutrosophic Sets for Sustainable Supplier Selection

Abstract

Sustainable supplier selection is a high-impact decision problem in which organizations must jointly evaluate economic performance, environmental impact, and social responsibility under heterogeneous stakeholder preferences, asymmetric decision authority, and predominantly linguistic assessments. Conventional multiattribute group decision-making (MAGDM) methods, based on crisp scores, fixed aggregation rules, and implicit consensus are poorly suited to contexts where expert opinions are expressed in natural language, indeterminacy is substantial, and strategic priorities may override majority preferences. To address these limitations, this study proposes a neural-based decision framework that integrates large language models (LLMs), neutrosophic sets, an autocratic aggregation strategy, and a neural network for supplier selection. Criterion-specific textual judgments provided by experts are processed by a pretrained model to extract sentiment polarity and confidence, which are then mapped to neutrosophic triplets. An autocratic mechanism empowers a designated decision leader to choose context-appropriate scoring and aggregation functions, avoiding forced consensus and reducing the need for extensive pairwise comparisons. A neural aggregation layer further learns how to combine multiple neutrosophic score functions, refining ranking behavior and lowering computational complexity. The framework is validated through a case study on sustainable supplier selection. The results demonstrate that the proposed approach produces stable, interpretable rankings with reduced computational burden and closer alignment with strategic sustainability goals than benchmark MAGDM methods.