Research map / Probabilistic and scientific modeling / Molecules and proteins

Towards Escaping from Class Dependency Modeling for Multi-Dimensional Classification

ICML 2025 · Probabilistic and scientific modeling › Molecules and proteins · approach: Probabilistic embeddings and relational inference

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 DCOM, a multi-dimensional classification method that avoids explicit class dependency modeling by learning a latent factor that makes class variables approximately conditionally independent given features and latent, then trains per-dimension MLPs.

Problem

Multi-dimensional classification requires modeling dependencies among intercoupled class variables, which is hard to do precisely; existing approaches model class dependencies directly.

Approach

Instead of modeling class dependencies, learn a latent factor Z (via a perturbed-input encoder) so that class variables become partially conditionally independent given X and Z, measured by an instance-based conditional mutual information loss; then each dimension is predicted by a simple neural network.

Key results

• DCOM significantly outperforms the seven compared approaches in 79.6% of cases for Hamming Score (78 wins, 19 ties, 1 loss across 98 pairwise comparisons, per Table 6 totals).
• On unstructured data sets DCOM achieves best Hamming Score on BP4D (.785) and DeepFashion (.785), second only to SLEM on SEWA.
• Ablation: removing the conditional-independence loss (DEV1 or DEV2) leads to statistically worse performance on HS and EM under Wilcoxon tests.

Measured effects reported in the paper

Surprising findings

• Chain-free CP beats DCOM on Enb in Hamming Score, with DCOM inferior to CP in only 3 cases overall, suggesting strongly correlated dimensions favor modeling the full joint space.
• SLEM is best on SEWA, which the authors attribute to label sparsity given many classes per dimension.

Limitations

• Theoretical justification relies on an assumption (Assumption 3.4) and small-joint-probability approximation that the authors themselves describe as only likely valid.
• Performance drops relative to CP on strongly correlated datasets such as Enb.
• Latent dimensionality must not be too low to retain feature information.

Experimental setup

Datasets: Flare1 Oes97 Jura Oes10 Enb Song BeLaE Voice Thyroid CoIL2000 TIC2000 Flickr

Benchmarks: Flare1 Oes97 Jura Oes10 Enb Song BeLaE Voice Thyroid CoIL2000 TIC2000 Flickr

Metrics: hamming score exact match sub-exact match

Baselines: BR CP DLEM EDCC LEFA PIST SLEM

Builds on: binary relevance class powerset PIST variational autoencoder

Closest papers by meaning

• Multi-Dimensional Classification via kNN Feature Augmentation (AAAI 2019 · similarity 0.86)
• Trading off Consistency and Dimensionality of Convex Surrogates for Multiclass Classification (NeurIPS 2024 · similarity 0.82)
• Self-Paced Unified Representation Learning for Hierarchical Multi-Label Classification (AAAI 2024 · similarity 0.82)
• CFDM: Contrastive Fusion and Disambiguation for Multi-View Partial-Label Learning (AAAI 2025 · similarity 0.81)
• Limited-Supervised Multi-Label Learning with Dependency Noise (AAAI 2024 · similarity 0.81)
• Reducing Class-wise Confusion for Incremental Learning with Disentangled Manifolds (CVPR 2025 · similarity 0.81)