Scholarlink

<p>To effectively support the sustainability of small businesses in dynamic and uncertain commercial environments, there is a growing need for AI systems that are both interpretable and actionable. Despite this need, few frameworks exist that effectively bridge predictive modeling and actionable strategy generation. In this study, we present an integrated framework that combines counterfactual explanation algorithms and large language models (LLMs) to support high-quality consulting strategies for business closure prevention. Using commercial district data from Seoul (2022–2023), we train a high-performing closure prediction model with key variables derived by the SHAP algorithm. We then apply the DiCE algorithm to generate counterfactual instances that balance interpretability and diversity. These instances are subsequently translated into actionable business strategies using prompt-engineered LLMs. Our results demonstrate that the integration of counterfactual reasoning with carefully designed LLM prompts enables scalable and transparent consulting support for small business decision-making. This research contributes a novel pipeline that bridges predictive modeling with policy-relevant strategy generation.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

From prediction to action: a framework for business closure analysis and LLM-based consulting strategy generation

  • 期刊论文

  • 发表日期：

    2025年12月11日

查看全文
  • Munil Yang,
  • Kwangho Kim,
  • Dongha Kim

摘要

To effectively support the sustainability of small businesses in dynamic and uncertain commercial environments, there is a growing need for AI systems that are both interpretable and actionable. Despite this need, few frameworks exist that effectively bridge predictive modeling and actionable strategy generation. In this study, we present an integrated framework that combines counterfactual explanation algorithms and large language models (LLMs) to support high-quality consulting strategies for business closure prevention. Using commercial district data from Seoul (2022–2023), we train a high-performing closure prediction model with key variables derived by the SHAP algorithm. We then apply the DiCE algorithm to generate counterfactual instances that balance interpretability and diversity. These instances are subsequently translated into actionable business strategies using prompt-engineered LLMs. Our results demonstrate that the integration of counterfactual reasoning with carefully designed LLM prompts enables scalable and transparent consulting support for small business decision-making. This research contributes a novel pipeline that bridges predictive modeling with policy-relevant strategy generation.