Interactive Personalized Clustering: Navigating beyond Rigid Criteria and Sample Ambiguity.

Clustering methods have made significant strides in recent years, yet most of them remain an inherently static paradigm. Owing to the fixed optimization objective and the exclusive reliance on internal data, this static clustering paradigm faces two fundamental challenges: i) Rigid Criteria: prioritizing semantics under the dominant criterion while overlooking others, and ii) Sample Ambiguity: struggling to discriminate hard samples at cluster boundaries. Recognizing that these two challenges arise from the intrinsic data limitation, we propose incorporating external user interaction to guide the clustering process. To be specific, we present an Interactive Personalized Clustering paradigm (IPC) that consists of two stages, i.e., Criterion Alignment and Grouping Refinement, where each stage is dedicated to resolving one challenge. In the Criterion Alignment stage, IPC derives the representation aligned with the user specified criterion via text basis transformation. Subsequently, in the Grouping Refinement stage, IPC identifies high-value sample pairs for friendly user feedback, utilizing three carefully designed losses to refine cluster assignments. Extensive experiments on 6 datasets across 16 criteria demonstrate that our proposed IPC substantially outperforms state-of-the-art baselines, achieving a 23.6% average performance improvement. Furthermore, IPC exhibits robustness and generalizability across various LLMs and VLMs. The code will be released.