MoM: Multi-to-Multi Identification of Decoupled Multifaceted Clustering with Dynamic Representation Fusion
Files
Date
Authors
Supervisor
Item type
Degree name
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
Multifaceted clustering is dedicated to excavating diverse and heterogeneous facets from complex multimedia data, which is a critical task for multimedia content understanding. Existing approaches primarily focus on decomposing a single dataset into multiple subspaces or sequentially generating alternative clusterings. However, these methods are severely constrained by the inherent limited information contained in the data itself, resulting in suboptimal clustering performance and indistinct facet mining results. To address the issue, we propose a novel paradigm shift from Single-to-Multi to Multi-to-Multi (MoM) clustering for multimedia data, which proactively integrates multiple informative views for decoupled multifaceted discovery, breaking the bottleneck of single-source data information scarcity. The proposed MoM framework consists of two core modules: the Dynamic Representation Fusion (DRF) module and the Decoupled Multifaceted Clustering (DMC) module. Specifically, the DRF module implements an adaptive fusion strategy that generates and dynamically fuses diverse representations guided by the anchors derived from a pretrained model, ensuring both fidelity and diversity. The DMC module leverages a structured self-expression model to jointly learn shared and facet-specific affinity matrices, producing well-separated and interpretable clusters. Unlike existing methods that passively decompose limited data into disjoint subspaces, our end-to-end framework actively employs a clustering-oriented representation fusion strategy to learn decoupled features tailored to each clustering facet. Extensive experiments show that MoM consistently outperforms state-of-the-art multifaceted clustering methods, validating the effectiveness of our MoM framework.

