Scholarlink

<p>Predicting where public complaints are likely to occur is essential for proactive urban management, yet remains challenging due to the heterogeneous structured nature of urban data. Although large language models (LLMs) have demonstrated strong performance in complaint analytics, existing approaches rely on transforming structured data into text, limiting their ability to capture spatial relationships and cross-feature interactions. To address this limitation, this paper proposed an image-space embedding framework that transformed structured data into multi-channel, image-like representations and integrated them with customized LLMs for complaint location prediction. 48,103 complaints were collected from Seberang Perai, Malaysia with complaint information, points of interest, weather conditions, population density, and socioeconomic indicator. These features are embedded into a unified image-space representation and the existing LLMs are customized by replacing their token embedding layers with convolutional neural network architecture for direct operation in image space. The framework is compared against existing text-based LLM pipelines under both partial and full fine-tuning. Experimental results demonstrated that combining LaBSE embeddings with customized CerebrasGPT-590M, achieved the highest accuracy (74.2%), Cohen’s kappa (71.8%), and F1-score (65.6%) while requiring significantly less training time (19.531 minutes) than fully fine-tuned baseline LLMs. Furthermore, strong performance gains obtained with customized vision-based models confirmed the effectiveness of the proposed image-space representations. Overall, the proposed framework provides an efficient solution for complaint location prediction and advances the integration of structured data with LLMs in image space.</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

Enhancing complaint locations prediction with image-space embedding representations and customized large language models

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年6月15日

查看全文
  • Theng-Jia Law,
  • Choo-Yee Ting,
  • Hu Ng,
  • Hui-Ngo Goh

摘要

Predicting where public complaints are likely to occur is essential for proactive urban management, yet remains challenging due to the heterogeneous structured nature of urban data. Although large language models (LLMs) have demonstrated strong performance in complaint analytics, existing approaches rely on transforming structured data into text, limiting their ability to capture spatial relationships and cross-feature interactions. To address this limitation, this paper proposed an image-space embedding framework that transformed structured data into multi-channel, image-like representations and integrated them with customized LLMs for complaint location prediction. 48,103 complaints were collected from Seberang Perai, Malaysia with complaint information, points of interest, weather conditions, population density, and socioeconomic indicator. These features are embedded into a unified image-space representation and the existing LLMs are customized by replacing their token embedding layers with convolutional neural network architecture for direct operation in image space. The framework is compared against existing text-based LLM pipelines under both partial and full fine-tuning. Experimental results demonstrated that combining LaBSE embeddings with customized CerebrasGPT-590M, achieved the highest accuracy (74.2%), Cohen’s kappa (71.8%), and F1-score (65.6%) while requiring significantly less training time (19.531 minutes) than fully fine-tuned baseline LLMs. Furthermore, strong performance gains obtained with customized vision-based models confirmed the effectiveness of the proposed image-space representations. Overall, the proposed framework provides an efficient solution for complaint location prediction and advances the integration of structured data with LLMs in image space.