SPORT: Structure-Aware Prototype Disentanglement for Incomplete Multi-View Clustering

Technical Summary: SPORT – Structure-Aware Prototype Disentanglement for Incomplete Multi-View Clustering

Problem Statement
Incomplete Multi-View Clustering (IMVC) aims to partition unlabeled data where samples may be missing one or more views due to sensor limitations or hardware failures. While recent prototype-based IMVC methods have improved robustness by using prototypes as semantic anchors for missing-view imputation, the authors identify three critical limitations in existing approaches:

1. Prototype Over-alignment Problem (POP): Existing methods enforce strict cross-view prototype consistency, ignoring view-specific information. This suppresses complementary information unique to specific views, limiting the expressiveness of multi-view representations.
2. Lack of Cluster-Level Structural Preservation: Most methods rely on instance-level contrastive learning that aligns only paired samples across views. They fail to preserve relational structures among semantically similar instances (samples within the same cluster), potentially pushing them apart in the feature space.
3. Inaccurate Missing-View Imputation: Current imputation strategies typically rely solely on global prototypes. This coarse-grained approach overlooks local geometric neighborhood structures, leading to over-smoothed reconstructions that fail to preserve fine-grained instance-level consistency.

Methodology: The SPORT Framework
The authors propose SPORT (Structure-aware PrOtotype disentanglement foR incomplete multi-view clusTering), a framework designed to jointly exploit cross-view consistency, view-specific semantics, and structural relationships. The architecture consists of four key components:

1. Feature Representation Learning:
   The model utilizes view-specific autoencoders to project observed features from each view into a shared latent subspace. A reconstruction loss ($L_{REC}$) ensures mapping fidelity and preserves inherent structural characteristics.

2. Structure-Aware Relational Contrastive Learning:
   To address the neglect of cluster-level structures, SPORT introduces a weighted contrastive learning mechanism. Unlike hard contrastive learning that treats only same-sample pairs as positives, this module uses pairwise similarity weights ($w_{ij}$) derived from cosine similarity. This allows semantically similar samples (even if not the same instance) to be softly treated as positive pairs, enforcing cross-view consistency while preserving the underlying geometric structure of the data.

3. Prototype Partial Alignment (Disentanglement):
   To solve the Prototype Over-alignment Problem, SPORT explicitly decomposes prototypes into two orthogonal components:

   • Shared Components ($s_i$): Aligned across views to capture view-invariant consensus semantics.
   • View-Specific Components ($c_i$): Decorrelated across views to preserve complementary, view-dependent information.
     This decomposition is achieved using a learnable orthogonal projection matrix (via Cayley transform re-parameterization). The objective includes a shared alignment loss ($L_{SHARED}$) and a view-specific de-correlation loss ($L_{DE}$).

4. Hybrid Prototype Imputation Strategy:
   For missing data recovery, SPORT employs a hybrid strategy that integrates global and local information. Missing features are reconstructed as a weighted sum of:

   • The most relevant prototype (global semantic anchor).
   • The most similar neighboring sample (local geometric context).
     This approach balances cluster-level semantics with local structural consistency, improving reconstruction quality under high missing rates.

The overall objective function combines reconstruction loss, contrastive loss, and prototype alignment loss, optimized in a two-phase training process (pretraining followed by fine-tuning).

Key Contributions
The paper claims the following contributions:

• Unified Prototype Alignment: A novel framework that decomposes prototypes into shared and view-specific components, aligning the former while decorrelating the latter to effectively integrate shared semantics and view-specific information.
• Structure-Aware Feature Alignment: A mechanism that introduces instance-wise similarity weights to softly encode inter-sample relationships, enforcing cross-view consistency while preserving cluster-level relational structures.
• Hybrid Imputation Strategy: A missing-data imputation method that jointly exploits matched prototypes and neighboring samples, leveraging both global cluster semantics and local structural information.
• Empirical Superiority: Extensive experiments demonstrating that SPORT outperforms state-of-the-art methods across various missing rates.

Experimental Results
The authors evaluated SPORT on six benchmark datasets (ALOI_100, Animal1, Digit4k, 100Leaves, Reuters_21578, and VGGFace2_50) against 14 state-of-the-art IMVC methods.

• Performance: SPORT achieved superior performance in Accuracy (ACC), Adjusted Rand Index (ARI), and F-Score across most datasets and missing rates (0.1 to 0.7). For instance, on the Animal dataset with a 0.5 missing rate, SPORT improved ACC by 17.7% over the second-best method.
• Robustness: The method demonstrated remarkable robustness as missing rates increased. On the 100Leaves dataset, SPORT showed significantly smaller performance fluctuations compared to other methods when the missing rate rose from 0.1 to 0.5.
• Ablation Studies: Removing any of the three core modules (representation learning, structure-aware contrastive learning, or prototype partial alignment) resulted in significant performance degradation, confirming the necessity and complementarity of each component.
• Imputation Analysis: The hybrid imputation strategy (balancing prototype and neighbor contributions) consistently outperformed degenerate variants that relied solely on either prototypes or neighbors.

Significance and Claims
The paper posits that SPORT addresses fundamental bottlenecks in prototype-based IMVC by moving beyond strict cross-view alignment and instance-level contrastive learning. By explicitly disentangling shared and view-specific prototype components and incorporating local neighborhood structures into the imputation process, the framework achieves more discriminative and robust clustering representations. The authors conclude that SPORT effectively learns discriminative representations under incomplete multi-view settings, offering a reliable solution for real-world scenarios where data incompleteness is ubiquitous.