📄 中文摘要
高维数据集在科学和工业领域日渐普遍，然而，由于距离度量的有效性降低以及聚类在降维投影时易塌陷或重叠，使得对其进行有效聚类变得困难。传统的降维技术生成静态的2D或3D嵌入，其可解释性有限，并且未能提供一种机制来利用用户领域知识。相反，交互式降维方法允许用户通过调整投影参数来探索数据，但它们通常无法直接控制聚类结果。此外，传统的半监督聚类算法虽然利用了少量标记数据，但其用户交互通常局限于初始标记阶段，无法在聚类过程后期进行动态调整或纠正。为了解决这些挑战，IPBC（Interactive Projection-Based Clustering）框架被提出，它将交互式降维与半监督聚类相结合，为用户提供了在聚类过程中持续介入的能力。IPBC的核心思想是允许用户通过调整数据投影来引导聚类过程，从而直观地影响聚类结果。该框架通过一个用户友好的界面实现，用户可以在其中实时观察投影变化对聚类结构的影响，并根据领域知识进行迭代修正。具体而言，IPBC采用了一种新颖的投影优化策略，它不仅考虑了数据自身的结构信息，还整合了用户提供的“必连”（must-link）和“勿连”（cannot-link）约束，以及通过交互式投影反馈获得的隐式指导。这种方法使得聚类不再是一个黑箱过程，而是成为一个由专家知识驱动的协同探索。IPBC通过在投影空间中优化聚类目标函数，同时确保投影能够有效区分用户指定的约束对，从而解决了高维数据聚类中的可解释性和准确性问题。实验结果表明，与现有方法相比，IPBC在高维数据集上显著提升了聚类性能和用户满意度，尤其是在需要人机协作来处理复杂数据模式的场景中表现出色。
📄 English Summary
IPBC: An Interactive Projection-Based Framework for Human-in-the-Loop Semi-Supervised Clustering of High-Dimensional Data
High-dimensional datasets are increasingly prevalent across scientific and industrial domains, yet effective clustering remains challenging due to the diminishing utility of distance metrics and the tendency of clusters to collapse or overlap when projected into lower dimensions. Traditional dimensionality reduction techniques generate static 2D or 3D embeddings that offer limited interpretability and lack mechanisms to leverage user domain knowledge. Conversely, interactive dimensionality reduction methods enable users to explore data by adjusting projection parameters, but they often do not directly control clustering outcomes. Furthermore, while traditional semi-supervised clustering algorithms utilize limited labeled data, user interaction is typically confined to the initial labeling phase, precluding dynamic adjustments or corrections later in the clustering process. To address these challenges, the Interactive Projection-Based Clustering (IPBC) framework is proposed, integrating interactive dimensionality reduction with semi-supervised clustering to provide users with continuous intervention capabilities throughout the clustering process. The core idea of IPBC is to allow users to guide the clustering process by adjusting data projections, thereby intuitively influencing clustering results. This framework is implemented through a user-friendly interface where users can observe the real-time impact of projection changes on cluster structures and make iterative corrections based on domain knowledge. Specifically, IPBC employs a novel projection optimization strategy that not only considers the intrinsic data structure but also integrates user-provided must-link and cannot-link constraints, along with implicit guidance obtained through interactive projection feedback. This approach transforms clustering from a black-box process into a collaborative exploration driven by expert knowledge. IPBC addresses interpretability and accuracy issues in high-dimensional data clustering by optimizing the clustering objective function within the projection space, while simultaneously ensuring that projections effectively differentiate user-specified constrained pairs. Experimental results demonstrate that IPBC significantly improves clustering performance and user satisfaction on high-dimensional datasets compared to existing methods, particularly excelling in scenarios requiring human-computer collaboration to handle complex data patterns.