Causality-Aware Feature Disentanglement Model for Shared and Class-Specific Information Learning

Abstract

Multi-label learning requires models to simultaneously capture information shared across labels and information that is uniquely associated with individual labels. Conventional feature-selection and representation-learning approaches can improve predictive performance, but they may inadequately distinguish genuine label-relevant information from redundant, correlated, or spurious associations. This paper proposes a Causality-Aware Feature Disentanglement Model (CAFDM) for learning shared and class-specific representations through causal feature analysis, correlation-aware disentanglement, and label-conditioned representation learning. The proposed framework separates an input representation into a common component that captures information useful across multiple labels and label-specific components that preserve discriminative information associated with individual labels. Causal relevance is incorporated to prioritize stable feature relationships rather than relying exclusively on statistical correlation. The framework is conceptually grounded in multi-label feature learning, causality-based feature selection, multidimensional correlation analysis, and representation separation. The proposed architecture further incorporates contrastive learning principles to increase the distinction between shared and label-specific representations while reducing redundancy. Analytical findings indicate that causal disentanglement can provide a more interpretable representation structure, improve resistance to spurious correlations, and support label-specific decision making without discarding globally useful information. The model is particularly applicable to multi-label classification, image annotation, text classification, recommendation, and biological function prediction. The study also identifies important limitations, including causal discovery uncertainty, computational complexity, dependence on label quality, and the absence of direct experimental validation. Overall, the framework provides a theoretically motivated direction for integrating causal reasoning with disentangled multi-label representation learning.