Technical Summary: DAS-PMVC

Problem Statement

Multi-view clustering typically assumes a complete correspondence between samples across different views. However, in practical scenarios involving hardware failures, noise, or independent data collection and storage, this assumption often fails, leading to the Partial View Alignment Problem (PVAP). In PVAP, multi-view data contains both aligned samples (where views correspond) and misaligned samples (where views do not correspond).

Existing solutions often treat misaligned data as incomplete multi-view data by discarding unaligned samples, which leads to significant data loss and increased processing complexity. Other methods rely on shallow models or perform view alignment and downstream clustering as separate, sequential steps. These approaches frequently fail to effectively leverage structural information between views, overlook structural consistency, and suffer from substantial alignment errors due to single-step alignment processes.

Methodology

The paper proposes DAS-PMVC (Dual Alignment and Structure Enhancement for Partial Multi-View Clustering), a framework designed to recover correspondences between views while enhancing feature discriminative power. The methodology consists of three core components:

1. Anchor Graph Structure Alignment

To address spatial heterogeneity and achieve initial alignment, the framework constructs anchor graphs for each view.

• Anchor Selection: Aligned samples within a view serve as anchor points.
• Bipartite Graph Construction: A bipartite graph is built to measure the similarity between misaligned samples and the selected anchors using squared cosine distance.
• Structural Filtering: A filtering process sets edge weights to zero for dissimilar samples, reducing the impact of irrelevant data.
• Initial Alignment: By mapping misaligned samples to a common latent space via anchor relationships, the framework constructs a similarity matrix. The Hungarian algorithm is then applied to find the optimal matching between views, generating a permutation matrix to reorder misaligned samples for preliminary alignment.

2. Structure-Enhanced Feature Learning

The framework employs a two-stage training process to extract deep latent features and refine representations:

• Pre-training Phase: The model learns view structure information using a reconstruction loss ($L_1$) and a structural alignment loss ($L_2$). This phase ensures that the encoder-decoder architecture captures semantic features and maintains structural consistency between aligned samples.
• Training Phase: A multi-view Graph Convolutional Network (GCN) is utilized. The aligned data is used to construct a similarity graph ($G^{(v)}$). The network learns to reconstruct both the node attributes and the graph structure (via structural reconstruction loss $L_3$).
• Contrastive Learning: To further optimize alignment and feature discriminability, a contrastive learning loss ($L_4$) is introduced. Aligned nodes form positive pairs, while misaligned nodes form negative pairs. The loss function minimizes the distance between positive pairs and maximizes the distance between negative pairs, guided by a margin parameter $m$.

3. Dual Alignment Strategy

DAS-PMVC implements a two-step alignment process to minimize errors:

1. First Alignment (Pre-training): Achieved through the anchor graph structure, establishing a joint embedding representation and initial sample correspondence.
2. Second Alignment (Training): After the GCN extracts latent features ($H^{(v)}$), the Hungarian algorithm is applied again in the feature space to the bipartite graph constructed from these features. This generates a new alignment matrix ($Q_{w,u}$) to reorder the latent features, correcting errors introduced in the initial alignment.

• Final Clustering: Spectral clustering is performed on the doubly aligned latent features to produce the final clustering results.

Key Contributions

The authors summarize their contributions as follows:

1. Anchor Graph Structure Alignment: Utilizing anchor point relationships to derive joint embedding representations with a consistent latent space. Structural filtering is applied to reduce the impact of irrelevant samples, achieving robust initial alignment.
2. Structure-Enhanced Feature Learning: Integrating pre-training with multi-view GCNs to extract deep latent features from aligned graph structures. Dual structural guidance is implemented to enhance the discriminative power of the learned representations.
3. Dual Alignment Strategy: A two-phase approach where initial alignment is performed via anchor graphs, followed by a secondary optimization of latent feature alignment using contrastive learning loss and the Hungarian algorithm during the training phase.

Experimental Results

The framework was evaluated on six multi-view datasets: Caltech20, BDGP, Scene-15, Aloi, 3Sources, and BBCsports. The alignment rate was set to 50% (the remaining 50% randomly shuffled). Performance was measured using Accuracy (ACC), Normalized Mutual Information (NMI), and Adjusted Rand Index (ARI).

• Overall Performance: DAS-PMVC outperformed eight state-of-the-art methods (including PVC, MvCLN, EGPVC, and TCLPVC) on most datasets.
  • On the 3Sources dataset, DAS-PMVC improved ACC by 5.79% over the second-best method.
  • On Caltech20, it achieved a 5.97% improvement in ACC.
  • It achieved the best ACC on five of the six datasets (excluding BDGP).
• Ablation Studies:
  • Alignment Strategy: Comparing anchor-based alignment against traditional attribute-based alignment showed that DAS-PMVC significantly outperformed the latter, with improvements exceeding 50 percentage points in alignment performance on certain datasets.
  • Module Effectiveness: Removing the anchor graph alignment module (creating a comparative model called PMGCN) resulted in lower clustering accuracy, validating the necessity of the initial anchor-based alignment.
• Limitations Observed:
  • On the BDGP dataset, performance was suboptimal compared to some competitors. The authors attribute this to the unsuitability of GCNs for this specific data, which lacks strong local structural relationships (derived from Drosophila embryo images).
  • On the Aloi dataset, while ACC remained high, NMI and ARI scores were lower than some competitors, potentially due to the large number of categories (100) causing clustering distribution imbalances.

Significance and Claims

The paper claims that DAS-PMVC effectively addresses the Partial View Alignment Problem by moving beyond the limitations of shallow models and single-step alignment. By jointly performing view alignment and downstream clustering tasks, the framework leverages the representational power of neural networks.

The significance of the work lies in its dual alignment strategy, which reduces errors inherent in single-step alignment, and its structure enhancement mechanism, which mitigates the influence of irrelevant samples through structural filtering and graph-based feature learning. The authors assert that the framework demonstrates superior effectiveness and robustness in handling misaligned multi-view data, particularly in scenarios where structural consistency and semantic relevance are critical for accurate clustering.