Weighted Deep Embedded Clustering for Robust Representation Learning in Noisy Data

Abstract

Deep clustering methods effectively learn meaningful representations from unlabeled data. Typically, existing approaches treat all samples equally during training, which can lead to unstable clustering in the presence of noise and outliers. In this research, we propose a Reliability-Based Deep Embedded Clustering (RDEC) approach that improves clustering reliability through a novel distance-aware weighting strategy. Specifically, rather than assigning equal importance to all samples, RDEC estimates and exploits the typicality of each instance according to its distance from the cluster centers in the latent space. Accordingly, samples that lie closer to cluster centers are assigned higher weights, while distant, atypical and ambiguous samples are gradually down-weighted using an inverse polynomial function. Moreover, the proposed weighting approach is coupled with a novel Kullback–Leibler divergence objective function to focus on the most representative data instances and guide the clustering process. Furthermore, the resulting latent distances and reliability weights demonstrate that RDEC effectively distinguishes reliable samples from noisy ones, resulting in more stable clustering behavior. RDEC was evaluated on three benchmark image datasets, namely MNIST, USPS, and Fashion-MNIST, under both clean conditions and controlled image corruption settings, including semantic outliers and synthetic perturbations, to assess its robustness during clustering. The experimental results demonstrated that RDEC achieved competitive clustering performance compared with representative clustering methods evaluated under the same experimental setting. In particular, RDEC yielded the best overall performance on the clean USPS dataset, attaining an ACC, NMI, and ARI of 0.8079, 0.7714, and 0.7392, respectively. Moreover, on the clean Fashion-MNIST dataset, RDEC achieved the highest NMI and ARI while maintaining competitive clustering accuracy, demonstrating its effectiveness on a more challenging clustering benchmark. Under these controlled image corruption scenarios, RDEC consistently achieved strong clustering performance across multiple corruption levels on MNIST, USPS, and Fashion-MNIST datasets, demonstrating the effectiveness of the proposed reliability-aware weighting mechanism under diverse image characteristics. The present study focuses on robustness under controlled image-based corruption scenarios, with the evaluation limited to benchmark image datasets.