This paper introduces Attention-Weighted Clustering (AWC), a novel framework leveraging self-attention mechanisms to dynamically weight and fuse heterogeneous data modalities for unsupervised clustering. The approach includes innovations such as a Modality-Aware Attention Module, a Heterogeneous Feature Fusion Network, and an Adaptive Cluster Assignment Mechanism. Extensive experiments demonstrate AWC's superior performance over state-of-the-art deep clustering methods.
Key findings
Proposes Attention-Weighted Clustering (AWC) for clustering heterogeneous unstructured data.
Introduces Modality-Aware Attention Module to learn cross-modal correlations without explicit alignment.
Develops a Heterogeneous Feature Fusion Network to project diverse data into a unified latent space.
Presents an Adaptive Cluster Assignment Mechanism for joint optimization of representation learning and clustering.
Theoretical analysis shows convergence properties and equivalence to kernelized clustering under mild assumptions.
Limitations & open questions
The framework's performance in real-world scenarios with highly incomplete or asynchronous data is not fully explored.
The complexity of the model may pose challenges for extremely large-scale datasets.