Research map / Label-efficient and robust learning / Recognition and long-tail

Hierarchical Consensus Network for Multiview Feature Learning

AAAI 2025 · Label-efficient and robust learning › Recognition and long-tail · approach: Uncertainty-aware multi-view learning · 2 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes the Hierarchical Consensus Network (HCN), which combines classifying (CCA-style), coding (contrastive-style) and global consensus objectives with a view-specific autoencoder for unsupervised multiview clustering and feature learning.

Problem

Learning view-consistent discriminative features for multiview data in an unsupervised setting, where existing consistency and contrastive methods are costly or rely on extra components and sample selection.

Approach

Derives three consensus indices (classifying, coding, global) from inner-product decompositions of cross-view feature matrices, linking classifying consensus to CCA and coding consensus to contrastive learning, and enforces them jointly with drop-feature augmentation and a view-specific autoencoder, without extra projectors or negative-pair sampling.

Key results

• On two-view clustering, HCN reaches 32.81 ACC / 38.58 NMI / 17.86 ARI on LandUse-21, versus 30.02 / 34.75 / 15.93 for the best prior method (SEM).
• On Caltech101-20 (two views), HCN achieves 77.39 ACC and 88.70 ARI, above DCP's 70.18 ACC and 76.88 ARI.
• Ablation on the two-view setting shows removing classifying consensus drops LandUse-21 ACC from 32.81 to 24.57 and Noisy MNIST ACC from 98.07 to 25.62.

Measured effects reported in the paper

Surprising findings

• Removing the classifying consensus loss collapses Noisy MNIST accuracy to about 25.6%, the largest effect among the components.
• Mean-of-five-seeds results on Noisy MNIST are reported with the best-run selection, and DealMVC performs very poorly on several datasets (e.g., 10.41% ACC on LandUse-21).

Limitations

• Evaluation limited to four datasets with k-means clustering as the downstream protocol.
• Results report the best selection per experiment, and the Noisy MNIST experiment uses only a 20k subset.
• Coding-consensus-to-contrastive equivalence rests on an argument with a Gaussian error assumption in the appendix proof.

Experimental setup

Datasets: LandUse-21 Caltech101-20 Scene-15 Noisy MNIST

Benchmarks: LandUse-21 Caltech101-20 Scene-15 Noisy MNIST

Metrics: ACC NMI ARI

Baselines: SC_Agg DCCA DCCAE BMVC PIC AE2Nets EERIMVC MFLVC CVCL DCP GCFAgg DealMVC

Builds on: CCA contrastive learning DCP MFLVC weak-to-strong pseudo supervision

Closest papers by meaning

• Trusted Unified Feature-Neighborhood Dynamics for Multi-View Classification (AAAI 2025 · similarity 0.85)
• Global Graph Propagation with Hierarchical Information Transfer for Incomplete Contrastive Multi-view Clustering (AAAI 2025 · similarity 0.84)
• PROTOCOL: Partial Optimal Transport-enhanced Contrastive Learning for Imbalanced Multi-view Clustering (ICML 2025 · similarity 0.84)
• Incomplete Contrastive Multi-View Clustering with High-Confidence Guiding (AAAI 2024 · similarity 0.83)
• Multi-view Granular-ball Contrastive Clustering (AAAI 2025 · similarity 0.83)
• Robust Consensus Anchor Learning for Efficient Multi-view Subspace Clustering (ICML 2025 · similarity 0.83)