Unpaired Multi-view Clustering via Reliable View Guidance

TL;DR

Unpaired Multi-view Clustering (UMC) tackles clustering when no cross-view sample pairings exist and labels are unavailable. The paper introduces reliable-view guidance with two models, RG-UMC (one reliable view) and RGs-UMC (multiple reliable views), which employ adaptive KL-divergence alignment, an orthogonal latent representation constraint, and a compactness term to enforce consistent cross-view cluster structures. Empirical results on five datasets show substantial improvements over state-of-the-art methods in NMI, ACC, and F-score (RG-UMC: +24.14% NMI, +37.49% ACC, +33.05% F-score; RGs-UMC: +29.42% NMI, +41.36% ACC, +35.47% F-score). The approach demonstrates that reliable-view guidance can robustly align unpaired views and improve joint clustering without supervision, with strong performance in both two-view and multi-view settings and potential applicability to real-world multi-modal data.

Abstract

This paper focuses on unpaired multi-view clustering (UMC), a challenging problem where paired observed samples are unavailable across multiple views. The goal is to perform effective joint clustering using the unpaired observed samples in all views. In incomplete multi-view clustering, existing methods typically rely on sample pairing between views to capture their complementary. However, that is not applicable in the case of UMC. Hence, we aim to extract the consistent cluster structure across views. In UMC, two challenging issues arise: uncertain cluster structure due to lack of label and uncertain pairing relationship due to absence of paired samples. We assume that the view with a good cluster structure is the reliable view, which acts as a supervisor to guide the clustering of the other views. With the guidance of reliable views, a more certain cluster structure of these views is obtained while achieving alignment between reliable views and other views. Then we propose Reliable view Guidance with one reliable view (RG-UMC) and multiple reliable views (RGs-UMC) for UMC. Specifically, we design alignment modules with one reliable view and multiple reliable views, respectively, to adaptively guide the optimization process. Also, we utilize the compactness module to enhance the relationship of samples within the same cluster. Meanwhile, an orthogonal constraint is applied to latent representation to obtain discriminate features. Extensive experiments show that both RG-UMC and RGs-UMC outperform the best state-of-the-art method by an average of 24.14\% and 29.42\% in NMI, respectively.

Figures (9)

• Figure 1: The strategies of reliable view guidance with one view and multiple views. Assuming the reliable ranking of three views is as follows: view 1 $>$ view 2 $>$ view 3. In (a), view 1 is designated as the reliable view, guiding both its cluster structure learning and that of other views (view 1 and view 2). In (b), views with stronger cluster structures serve as directors to guide views with weaker cluster structures. e.g., view 1 and view 2 both guide the learning of view 3. The strategies fully leverage all reliable views to effectively align cluster structures across different views.
• Figure 2: The RG-UMC framework illustrates the process for a batch dataset with three views. The model is designed to achieve a clear and consistent cluster structure across multiple views. It consists of three essential components: an orthogonal constrain, an alignment module with one reliable view, and a compactness module. In the alignment module, the most reliable view dynamically changes with the arrival of different batches of data.
• Figure 3: The alignment module of RGs-UMC involves three views. The reliable view dynamically changes based on the silhouette coefficient in each batch during round $T$ optimization. The strategy fully leverages all reliable views to effectively align cluster structures across different views.
• Figure 4: Taking three views as an example, the original weights are processed using three strategies: Equal, Normalization, and Normalization after sigmoid.
• Figure 5: Training loss in RG-UMC and RGs-UMC.