PEC-CDC: A prediction error-based calibration framework for robust unsupervised deep clustering

Abstract

While deep clustering has made significant strides, the inherent overconfidence of Softmax-based calibration remains a critical bottleneck. Even dedicated frameworks like Calibrated Deep Clustering (CDC) fail to fully eliminate this issue, as they continue to rely on probability-based metrics. These relative metrics are constrained by inter-class normalization, which often forces high confidence scores onto ambiguous samples where discriminative features are underdeveloped. To address this limitation, we propose PEC-CDC, a novel framework that reconfigures the CDC architecture by replacing its probability-based calibration head with a regression-based Prediction Error-based Classification (PEC) head. Our core motivation is that prediction error serves as an absolute measure of distributional support independent of other classes, making it inherently resistant to overconfidence. Specifically, we utilize a frozen Teacher-Student module to shift the model’s evaluation from relative confidence probability to absolute sample–cluster distribution consistency. Qualitative analysis via t-SNE and error distribution visualizations confirms that our framework establishes superior semantic boundaries and compact cluster manifolds. Extensive experiments demonstrate that PEC-CDC achieves state-of-the-art performance; notably, on the challenging CIFAR-100 dataset, it attains an Adjusted Rand Index (ARI) of 45.61%, effectively quadrupling the performance of leading baselines such as CueCo.