You Can Trust Your Clustering Model: A Parameter-free Self-Boosting Plug-in for Deep Clustering

TL;DR

DCBoost addresses the gap where deep clustering methods often exhibit weak global structure despite strong local neighborhoods. It is a parameter-free plug-in that uses adaptive $k$-NN consistency within mini-batches to select high-confidence samples as anchors, then optimizes a pseudo-label augmented discriminative loss $L=L_{pos}+L_{neg}+L_{ins}$ to tighten intra-class clustering and enhance inter-class separation. Across five benchmarks and six baselines, DCBoost delivers consistent gains, markedly improving silhouette scores and ACC while maintaining efficiency, and even extending benefits to CLIP-based models. By strategically leveraging reliable local cues to refine the global feature space, DCBoost offers a practical, scalable boost for deep clustering with broad applicability.

Abstract

Recent deep clustering models have produced impressive clustering performance. However, a common issue with existing methods is the disparity between global and local feature structures. While local structures typically show strong consistency and compactness within class samples, global features often present intertwined boundaries and poorly separated clusters. Motivated by this observation, we propose DCBoost, a parameter-free plug-in designed to enhance the global feature structures of current deep clustering models. By harnessing reliable local structural cues, our method aims to elevate clustering performance effectively. Specifically, we first identify high-confidence samples through adaptive $k$-nearest neighbors-based consistency filtering, aiming to select a sufficient number of samples with high label reliability to serve as trustworthy anchors for self-supervision. Subsequently, these samples are utilized to compute a discriminative loss, which promotes both intra-class compactness and inter-class separability, to guide network optimization. Extensive experiments across various benchmark datasets showcase that our DCBoost significantly improves the clustering performance of diverse existing deep clustering models. Notably, our method improves the performance of current state-of-the-art baselines (e.g., ProPos) by more than 3% and amplifies the silhouette coefficient by over $7\times$. Code is available at <https://github.com/l-h-y168/DCBoost>.

Figures (12)

• Figure 1: (a) Quantitative comparisons of the global structure in terms of intra-class similarity, inter-class similarity, and silhouette coefficient, and the local structure in terms of $k$-NN accuracy on CIFAR-10 and CIFAR-20. The original model ProPos and CC show weak global structure but strong local structure. Our method maintains their high local accuracy while significantly improving the global structure and overall clustering performance. (b) T-SNE visualization of clustering without (upper) and with (bottom) applying our method on CIFAR-20. Initially, the existing model exhibits severe overlap between different classes, leading to poor separability. After boosting, class boundaries become significantly clearer and more distinguishable.
• Figure 2: Illustration of our DCBoost framework. During training, the gradient backpropagation of the target network is detached, and the parameters of $f_t(\cdot)$ are updated using exponential moving average (EMA) from those of $f_o(\cdot)$. The overall discriminative loss contains three terms: positive loss $L_{pos}$, negative loss $L_{neg}$, and instance consistency loss $L_{ins}$. The inference branch outputs the pseudo labels of all samples at the end of each epoch.
• Figure 3: Impact of the value of $k$ on (a) the true accuracy of high-confidence samples, and (b) the proportion of high-confidence samples, i.e., the ratio of selected high-confidence samples to the total number of samples in a batch under given $k$. A smaller $k$ may reduce the reliability of the selection, whereas a larger $k$ generally improves sample quality but selects fewer samples. (c) Comparison of imbalance ratios between high-confidence samples and all samples across classes. The number of high-confidence samples selected per class can be imbalanced, without proper constraints, classes with larger sample counts may dominate training and negatively affect performance.
• Figure 4: T-SNE visualization of ProPos (left) and ProPos+Ours (right) on four datasets. The visualization of CIFAR-20 has been presented on Fig. \ref{['CIFAR20']}(b).
• Figure 5: Influence on silhouette coefficient (global structure) and $k$-NN accuracy (local structure).