Scholarlink

<p>Clustering analysis (CA) plays a key role in many industrial machine learning (IML) tasks by discovering and interpreting hidden patterns among data. The growing IML data may yield overlapped data—“similar in characteristics but different in category”. This impacts the CA performance and demands the need to redefine features, distance matrices, centroids, and other hyperparameters at every overlap appearance. These issues raise doubt on the reliability of the recent CA solutions in terms of interpretation and feasibility unsupervisedly. This study showcases an in-depth investigation by exploring the complex relationships among the given settings and issues, aiming to highlight the main challenges associated with CAs that impact prediction, partitioning, and false confidence in outcomes of CA models. We propose a novel experimental framework that has been able to produce over 271 million possible settings of the Iris dataset based on various models, equations, and procedures. Findings indicate that achieving reliable solutions is questionable and doubt the use of supervised information upfront, which is against the CA concept. Thereby, we designed an open issue that challenges the CA models. Experiment code, results, and open issues are released at <a href="https://github.com/MustafaRaadKadhim/CaseStudy">https://github.com/MustafaRaadKadhim/CaseStudy</a>.</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

Exploring the reliability of clustering models: a case study on interpretation and overlapping issues

  • 期刊论文

  • 发表日期：

    2025年7月8日

查看全文
  • Mustafa Kadhim,
  • Peilin Li,
  • Guangxi Lu,
  • Jianbo Wang,
  • Ruyi Tang,
  • Songlin Ye

摘要

Clustering analysis (CA) plays a key role in many industrial machine learning (IML) tasks by discovering and interpreting hidden patterns among data. The growing IML data may yield overlapped data—“similar in characteristics but different in category”. This impacts the CA performance and demands the need to redefine features, distance matrices, centroids, and other hyperparameters at every overlap appearance. These issues raise doubt on the reliability of the recent CA solutions in terms of interpretation and feasibility unsupervisedly. This study showcases an in-depth investigation by exploring the complex relationships among the given settings and issues, aiming to highlight the main challenges associated with CAs that impact prediction, partitioning, and false confidence in outcomes of CA models. We propose a novel experimental framework that has been able to produce over 271 million possible settings of the Iris dataset based on various models, equations, and procedures. Findings indicate that achieving reliable solutions is questionable and doubt the use of supervised information upfront, which is against the CA concept. Thereby, we designed an open issue that challenges the CA models. Experiment code, results, and open issues are released at https://github.com/MustafaRaadKadhim/CaseStudy.