A Human-Centered Explainability Framework for Latent Clustering Based on NMF, Fuzzy Logic, and LLMs

Explaining unsupervised models in a human-centered manner remains challenging, particularly when relying on latent representations. While Non-negative Matrix Factorization (NMF) provides structured latent factors, their semantic interpretation is often limited. Conversely, Large Language Models (LLMs) can generate fluent explanations but require semantically grounded inputs to ensure faithfulness. In this work, we propose a modular framework integrating NMF, fuzzy semantic modeling, and LLM-based explanation generation to support human-centered explanations of clustering results. Fuzzy logic maps latent factors and cluster representatives into linguistically meaningful concepts, which guide LLMs through structured prompting. We evaluate six prompting strategies with increasing semantic enrichment across eleven open-source LLMs. The framework is validated on an acoustic analysis task related to bipolar disorder, where interpretability is critical for expert understanding. Results show that combining concept-based representations with structured prompting improves the accessibility and reliability of LLM-generated explanations in domain-sensitive contexts.