Adaptive Multi-Stage Feature-View Fusion via Deep Graph Representation for Clustering

Deep neural networks primarily aim to enhance the representation ability of high-level semantic features by deepening the network to reinforce critical features. However, this often leads to the under-utilization or discarding of certain low-level information, resulting in suboptimal performance. Due to differences in dimensionality and key point selection, feature views (FVs) at various depth levels reflect diverse types of semantic information. Simply using the last layer of embeddings as the representation for clustering may be irrational. Therefore, fully leveraging low-level representations is crucial, rather than merely increasing network depth. On the other hand, existing work has rarely focused on adaptively learning the internal structure and potential connections through different FVs. Current FV-weight strategies often eliminate relatively unimportant FVs via sparse weights. To address these issues, we propose a multi-stage encoder structure incorporating a unified contrastive learning module to derive hierarchical FVs with diverse semantic information. Specifically, we devise a negative inverse technique with a rigorously proven adaptive non-parametric procedure, aiming to reinforce pivotal FVs while maintaining others for efficient FV fusion. Additionally, the multi-FV structure is adaptively constructed using selected neighbors for deep graph representation for clustering. Experimental results on seven datasets demonstrate the effectiveness of the proposed method for clustering, with improvements of up to 7.13% in accuracy and 2.99% in normalized mutual information over sixteen state-of-the-art (SOTA) baselines.