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

T-MDML: Triplet-based Multiple Distance Metric Learning for Multi-Instance Multi-Label Classification with Label Correlation

AAAI 2025 · Label-efficient and robust learning › Recognition and long-tail · approach: Partial and multi-label 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 T-MDML, a triplet-based multiple distance metric learning method for multi-instance multi-label classification that learns global and label-specific metrics plus a label correlation estimated from bag and label spaces, applied to MIMLkNN and other distance-based MIML learners.

Problem

Similarity-based multi-instance multi-label (MIML) classification relies on distances computed only in feature space, ignoring label information, and existing DML methods for MIML learn a single metric that cannot capture label-specific properties.

Approach

Learns a global Mahalanobis transformation shared across labels plus a per-label transformation, optimized with a smooth sigmoid-like triplet loss over bag triplets, with a regularizer that ties label transformations according to a label correlation computed from both the bag-feature kNN graph and label manifold.

Key results

• Applying T-MDML to MIMLkNN improves Hamming Loss on all 10 datasets (e.g., H. marismortui 0.0921 vs 0.1099 for MIMLkNN in Table 3 terms of Hamming Loss).
• T-MDML+MIMLkNN ranks first overall (average rank 2.40) across Ranking Loss, Coverage, and Average Precision compared with seven MIML algorithms.
• Label correlation improves most metrics on nine of ten datasets, with larger gains on protein datasets that have many labels; news is the exception.

Measured effects reported in the paper

Surprising findings

• Label correlation gives no gain on the news dataset, where performance is slightly worse on most metrics.

Limitations

• Some baselines (e.g., MIMLBoost) fail or do not finish on the largest protein datasets, and T-MDML's own scaling is not analyzed.
• Evaluation is restricted to distance-based MIML learners and a fixed set of 10 datasets.
• Gains from label correlation are small or absent on some datasets (e.g., news).

Experimental setup

Datasets: scene MSRC v2 news birds reuters H. marismortui P. furiosus A. vinelandii C. elegans D. melanogaster

Benchmarks: scene MSRC v2 news birds reuters H. marismortui P. furiosus A. vinelandii C. elegans D. melanogaster

Metrics: Hamming Loss Ranking Loss One Error Coverage Average Precision

Baselines: MIMLkNN MIMLBRkNN MIMLIBLR MIMLDGC MIMLBoost MIMLSVM KISAR

Builds on: MIMLkNN large margin nearest neighbor Mahalanobis distance metric learning average Hausdorff distance

Closest papers by meaning

• Implicit Relative Labeling-Importance Aware Multi-Label Metric Learning (AAAI 2025 · similarity 0.88)
• Tensorized Multi-View Multi-Label Classification via Laplace Tensor Rank (ICML 2025 · similarity 0.87)
• Multi-Instance Multi-Label Classification from Crowdsourced Labels (AAAI 2025 · similarity 0.87)
• Multi-View Multi-Label Classification via View-Label Matching Selection (AAAI 2025 · similarity 0.86)
• Double-Layer Hybrid-Label Identification Feature Selection for Multi-View Multi-Label Learning (AAAI 2024 · similarity 0.85)
• Correlation-Induced Label Prior for Semi-Supervised Multi-Label Learning (ICML 2024 · similarity 0.85)