Consensus-driven tensor learning for multi-source multi-instance multi-label classification

Y.L. Chen

doi:10.1016/j.eswa.2026.132954

Consensus-driven tensor learning for multi-source multi-instance multi-label classification

2026-09-29

1

PRE

AI

DOI：10.1016/j.eswa.2026.132954

Original

Original request for help

Save

Abstract

En 中文

Multi-source multi-instance multi-label learning (MSMIML) seeks to predict fine-grained instance-level labels under coarse bag-level supervision from heterogeneous sources. However, propagating coarse bag-level annotations into reliable instance-level semantics is inherently difficult due to cross-source feature heterogeneity and the fundamental bag-instance supervision mismatch. Existing methods remain confined to pairwise correlation modeling and decouple fusion from classification, lacking tensor-based representations to capture complex instance-label-source dependencies. To address these limitations, we propose CTM3L, a consensus-driven tensor learning framework for MSMIML. Specifically, a low-rank tensor-based bag-instance semantic decoupling module is proposed to propagate coarse bag-level labels into latent instance-level semantics. Heterogeneous features from each source are encoded using a Broad Learning System, and the resulting representations are structured into a multi-source feature tensor. A tensor tri-mode projection (TMP) maps this tensor into a latent instance-level semantic space, thereby enabling joint fusion of heterogeneous sources and instance-level label prediction via high-order instance-label-source interactions. Within TMP, the label-mode projection functions as a tensor-oriented classifier, while the source-mode projection estimates source contributions to drive an adaptive label fusion mechanism that consolidates source-specific predictions into consensus instance-level labels. These consensus labels in turn regularize instance semantic learning via cross-source consistency constraints, forming a self-consistent optimization loop. An efficient alternating optimization algorithm with closed-form updates is designed for CTM3L. Experiments show it improves average precision by 8.0-10.1 percentage points over strong baselines at both bag and instance levels.

Keywords:

Multi-source multi-instance multi-label learning

Multi-source information fusion

Tensor modeling

Label consensus

Weak supervision

Consensus-driven tensor learning for multi-source multi-instance multi-label classification

Abstract

Journal

Organization

Cited Papers