Personalized Clustering via Targeted Representation Learning

TL;DR

The paper tackles personalized clustering by allowing user-driven clustering orientation through targeted representation learning and active querying of pairwise constraints. It introduces PCL, a framework that uses two data augmentations, cross-instance attention, and a constrained contrastive loss to learn representations aligned with a specified cluster direction, guided by an active query strategy that balances uncertainty and hard negatives. Theoretical results bound clustering risk and show that actively querying informative pairs reduces the generalization gap, while experiments on CIFAR-10/100 and ImageNet-10 demonstrate strong performance especially in personalized settings with minimal constraint budgets. This approach enables efficient, user-centered clustering adaptable to diverse applications and data orientations.

Abstract

Clustering traditionally aims to reveal a natural grouping structure within unlabeled data. However, this structure may not always align with users' preferences. In this paper, we propose a personalized clustering method that explicitly performs targeted representation learning by interacting with users via modicum task information (e.g., $\textit{must-link}$ or $\textit{cannot-link}$ pairs) to guide the clustering direction. We query users with the most informative pairs, i.e., those pairs most hard to cluster and those most easy to miscluster, to facilitate the representation learning in terms of the clustering preference. Moreover, by exploiting attention mechanism, the targeted representation is learned and augmented. By leveraging the targeted representation and constrained contrastive loss as well, personalized clustering is obtained. Theoretically, we verify that the risk of personalized clustering is tightly bounded, guaranteeing that active queries to users do mitigate the clustering risk. Experimentally, extensive results show that our method performs well across different clustering tasks and datasets, even when only a limited number of queries are available.

Figures (7)

• Figure 1: The diversity of cluster orientation. Different tasks have different orientations for feature learning and image clustering.
• Figure 2: The framework of PCL. A deep neural network creates representations from two random augmentations of the data. By assessing the position of images in the feature space, PCL selects the most informative sample pairs to guide the training of the model. The model is then retrained by querying whether these pairs are must-link or cannot-link. This approach allows the final model to concentrate on features relevant to the desired cluster orientation, resulting in accurate clustering outcomes.
• Figure 3: (Left) The illustration shows a sample pair with uncertain similarity, highlighted as red nodes, which are prioritized for querying. (Right) Some samples, due to incorrect feature extraction, are mistakenly grouped together despite belonging to different clusters, as shown by the red and yellow nodes. Identifying these can be achieved by assessing the similarity between their positive samples.
• Figure 4: Clustering accuracy and NMI on train sets and test sets for different number of pairwise constraints.
• Figure 5: The evolution of features across the training process under two cluster orientations: (a) Initial distribution of samples, (b) Sample distribution after clustering along the default orientation and (c) Sample distribution after clustering along a given personalized orientation. The color of the dots denotes their original class labels of CIFAR-10.