Beyond One-Layer Decision: Hierarchical Multi-view Clustering with Progressive Cross-Layer Fusion

Kaixuan Zhou†, Pengyuan Li†, Jiahui Zhang, Dongxia Chang*, Yiming Wang, Yao Zhao

Paper | Code

• We propose a novel Hierarchical-Aware Multi-view Clustering framework. Unlike conventional approaches restricted to single-layer semantics, this framework establishes a holistic learning paradigm that transcends the representation bottleneck by simultaneously exploiting intra-view hierarchy and inter-view consensus.

Disentangled Contrastive Multi-view Clustering via Semantic Relevance Invariance

Pengyuan Li, Dongxia Chang*, Yiming Wang, Zisen Kong, Linhua Kong, Yao Zhao

Paper | Code

• We propose a disentangled contrastive multi-view clustering via semantic relevance invariance, which achieves intra-view and inter-view disentanglement and thus a more discriminative representation. The method not only makes disentangled representations containing different underlying information but also ensures their semantic relevance consistency.

Deep Multi-view Clustering with Intra-view Similarity and Cross-view Correlation Learning

Pengyuan Li, Dongxia Chang*, Yiming Wang, Man Liu, Zisen Kong, Linhua Kong, Yao Zhao

Paper | Code

• We present a novel deep learning framework that mitigates view bias through joint optimization of intra-view similarity and cross-view correlation. The proposed method enhances fine-grained structures within each view and adaptively balances diverse information across views, ultimately improving clustering performance.

AEMVC: Mitigate Imbalanced Embedding Space in Multi-view Clustering

Pengyuan Li†, Man Liu†, Dongxia Chang*, Yiming Wang, Zisen Kong, Yao Zhao

Paper | Code

• We found that the embedding space learned using the encoder-decoder architecture cannot embrace the efficacy of different feature directions. Therefore, we propose a novel Activate-Then-Eliminate Strategy for Multi-View Clustering to adjust the contribution strength of different feature directions dynamically.

DCMVC: Dual Contrastive Multi-view Clustering

Pengyuan Li, Dongxia Chang*, Zisen Kong, Yiming Wang, Yao Zhao

Paper | Code

• We propose a novel deep contrastive multi-view clustering method termed DCMVC. The dual contrastive mechanism can alleviate the constraints of a single positive sample on contrastive learning by incorporating category information to regularize the feature structure and fully explore the consistency of similar samples.

Deep Learning on Histology Images for Differentiating of Fibro-Osseous

Aobo Zhang†, Pengyuan Li†, Jiang Xue†, Jianyun Zhang, Zhu You, Shaohua Ge, Zhixiu Xu, Zhipeng Sun, Dongxia Chang*, Lisha Sun, Tiejun Li

Paper | Code

• Our results demonstrate that integrating multi-slide and weakly supervised strategies significantly enhances diagnostic performance for fibro-osseous lesions. Compared to human pathologists, the multi-slide models achieved higher accuracy, whereas weakly supervised models consistently outperformed fully supervised models.

Multi-Level Decoupled Trend Learning for GNN-Based Multivariate Time Series Prediction

Shaohan Li, Zhenfeng Zhu, Youru Li, Yeyu Yan, Shuai Zheng, Pengyuan Li, Yan Zhuang, Yao Zhao

Paper | Code

• We propose a multi-level decoupled trend learning (MDTL) framework for MTS prediction, which decouples the complex MTS signals at trend and spatial-temporal dependency levels and then fuses them in a flexible way.

Revisiting Radar Camera Alignment by Contrastive Learning for 3D Object Detection

Linhua Kong, Dongxia Chang, Lian Liu, Zisen Kong, Pengyuan Li, Yao Zhao

Paper | Code

• We propose a novel radar camera alignment model called RCAlign based on the sparse BEV alignment methods for 3D object detection.

ATMCA: Augmented Tensorized Consensus Learning for Multi-view Clustering with Anchor-Aligned

Zisen Kong, Pengyuan Li, Dongxia Chang, Yiming Wang, Yao Zhao

Paper | Code

• We provide an intuitive solution to the Anchor-Unaligned Problem. The method introduces the reordering alignment mechanism and augmented tensorized consensus learning into the joint optimization framework.

Beyond Forced Modality Balance: Intrinsic Information Budgets for Multimodal Learning、

Zechang Xiong, Da Li*, Kexin Tang, Pengyuan Li, Wenkang Kong, Yulan Hu

Paper | Code

• We argue that modality balance should be defined by an intrinsic equilibrium determined by the information capacity of each modality, rather than heuristic equalization. Therefore, we propose IIBalance, a multimodal learning framework for balanced learning under capacity-aware guidance.

Tensorial Multi-view Clustering via Alternative Rank Minimization and Inter-view Alignment

Zisen Kong, Dongxia Chang*, Yiming Wang, Pengyuan Li, Yao Zhao

Paper | Code

• We propose a novel rank minimization strategy for tighter rank function approximation. The strategy can effectively utilize the low-rank structure and higher-order correlations embedded in different views, which helps to generate a discriminative consensus representation.

Bipartite Contrastive Multi-view Clustering with Singular Value Modulation

Teng Zhang, Pengyuan Li, Zisen Kong, Dongxia Chang*, Yao Zhao

Paper | Code

• We reformulate contrastive learning as a binary classification problem, avoiding the limitation in previous contrastive methods that heavily depend on naturally paired data. By capturing sample-level and category-level consistency relationships among multiple views, the learned representations are further refined.

Local Geometry-Enhanced Anchor Learning for Multi-view Clustering

Zisen Kong, Zhiqiang Fu, Dongxia Chang*, Yiming Wang, Pengyuan Li, Yao Zhao

Paper | Code

• We introduce a coarse-grained anchor learning mechanism that maps each view anchor to the consensus space, effectively improving the expressiveness and learning of the framework.