Abstract
Multi-view graph clustering (MVGC) is a critical technique for uncovering patterns in complex, interconnected data. However, real-world applications often suffer from significant attribute missingness, where nodes lack feature information across various views. We present AMMGC (Attribute-Missing Multi-view Graph Clustering), a novel end-to-end framework designed to handle extreme data scarcity. By integrating context-aware imputation, cross-view structural alignment, and high-confidence pseudo-labeling, AMMGC maintains high performance even when up to 90% of node attributes are missing.
Introduction
In modern data science, multi-view graphs—such as social networks with both friendship links and interaction frequencies—provide a rich basis for clustering. Traditional algorithms typically assume that attribute matrices for every node in every view are complete. In practice, factors like sensor failure, privacy restrictions, and data corruption lead to "attribute-missing" scenarios. Existing solutions often fail to account for the relational dependencies inherent in graphs, ignoring the fact that a node's missing features can be inferred from its neighbors. AMMGC addresses this gap by treating feature recovery as a graph-based optimization problem.
The AMMGC Framework
The AMMGC framework is composed of three synergistic modules designed to reconstruct missing data and learn robust latent representations.
- Context-Aware Imputation (CAI): Rather than employing simple mean or zero-filling, CAI leverages the graph's topology. It minimizes "Dirichlet energy" to iteratively propagate feature information from known nodes to missing ones, ensuring that the imputed attributes are consistent with the local network structure.
- Dual Structure Consistency (DSL): To process multiple views, the model utilizes a Laplacian filter to mitigate high-frequency noise. Features are then encoded into a latent space using Multi-Layer Perceptrons (MLPs) with shared parameters. An adaptive fusion mechanism weighs informative views more heavily, while cross-view similarity constraints ensure the learned embeddings align with the underlying graph structure.
- High-Confidence Guidance (HCG): To refine the clustering process, the model generates pseudo-labels via K-means on the fused embeddings. To avoid the "garbage in, garbage out" pitfall of self-supervised learning, HCG filters these labels, retaining only the most confident samples (those closest to cluster centroids) to guide the network's soft-assignment loss.
Experimental Evaluation
The framework was rigorously tested on four benchmark datasets: CORA and CITESEER (citation networks), AMAC (Amazon marketing data), and WIKI (Wikipedia network). The researchers evaluated performance under varying levels of attribute missingness, ranging from 30% to a severe 90%.
Results and Discussion
Experimental results demonstrate that AMMGC significantly outperforms state-of-the-art baselines, particularly in high-missingness scenarios.
- Superiority at Scale: On the CORA dataset with a 90% missing rate, AMMGC achieved an Accuracy (ACC) of 68.94%, outperforming the next best method (CONVERT) by nearly 18.5%.
- Resilience: The model exhibits remarkable stability. On the AMAC dataset, increasing the missing rate from 0.3 to 0.9 resulted in a negligible accuracy drop of only 2.66%, whereas competing methods like SAT experienced performance decreases exceeding 26%.
- Metric Consistency: As shown in the comparative analysis, AMMGC (denoted as CA in the results) consistently leads across Accuracy (ACC), Normalized Mutual Information (NMI), Adjusted Rand Index (ARI), and F1-score.
Conclusion
AMMGC proves that the graph structure itself is a powerful tool for data recovery. By treating feature imputation as an energy minimization problem and focusing on structural alignment over raw feature matching, the framework provides a robust solution for real-world, incomplete graph data. Future work may explore more advanced graph filtering techniques to further enhance performance on noisy or edge-deficient networks.