Uncertainty-Aware Probabilistic Constrained Clustering from Entangled Pairwise Supervision

A recent paper on arXiv presents uncertainty-aware probabilistic constrained clustering (UPCC), which tackles the shortcomings of current deep constrained clustering (DCC) techniques that depend on rigid must-link and cannot-link labels. The researchers establish a framework where pairwise supervision is both real-valued and intertwined with inherent ambiguity, expert insights, and random corruption. They introduce a canonical aleatoric target through a varied observation process and examine its conditional identifiability. The method, named ProbPair, utilizes an angular pairwise objective for probabilistic relationships, while ECI-PP serves as an estimator-corrector-integrator framework that enhances flawed supervision through belief estimation, correction, and reliability-focused integration. Tests across various benchmarks highlight its effectiveness in complex probabilistic supervision scenarios. The paper can be found on arXiv under ID 2608.12027.