Scholarlink

<p>Deep clustering, owing to its capabilities in feature learning and label learning, has been widely applied across various aspects of machine learning. Among these approaches, deep clustering based on large language models stands out as a prominent method. However, traditional large language model-based deep clustering primarily focuses on textual features while neglecting visual characteristics. To address this limitation, this paper proposes a novel CLIP-based clustering method assisted by diffusion models. Specifically, to harness the knowledge embedded in these large-scale models for improved clustering, we first adapt the CLIP model to function as a pre-clustering model by adding a remap layer and a contrastive module. Building on the pre-clustering results, we employ a diffusion model to further extract sample features and optimize the pre-clustering by the relationships between samples and guidance through diffusion model sampling. The method proposed in this paper is evaluated on a diverse set of datasets, and the experiments demonstrate that our model is competitive with state-of-the-art methods in recent years.</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

Improving clustering with score-based diffusion models

  • 期刊论文

  • 发表日期：

    2026年5月28日

查看全文
  • Jiale Chen,
  • Xulun Ye

摘要

Deep clustering, owing to its capabilities in feature learning and label learning, has been widely applied across various aspects of machine learning. Among these approaches, deep clustering based on large language models stands out as a prominent method. However, traditional large language model-based deep clustering primarily focuses on textual features while neglecting visual characteristics. To address this limitation, this paper proposes a novel CLIP-based clustering method assisted by diffusion models. Specifically, to harness the knowledge embedded in these large-scale models for improved clustering, we first adapt the CLIP model to function as a pre-clustering model by adding a remap layer and a contrastive module. Building on the pre-clustering results, we employ a diffusion model to further extract sample features and optimize the pre-clustering by the relationships between samples and guidance through diffusion model sampling. The method proposed in this paper is evaluated on a diverse set of datasets, and the experiments demonstrate that our model is competitive with state-of-the-art methods in recent years.