AEMLO: AutoEncoder-Guided Multi-Label Oversampling

TL;DR

This work tackles class imbalance in multi-label classification by introducing AEMLO, an autoencoder-guided oversampling method that learns a low-dimensional joint feature–label latent space via a deep canonical correlation objective. It combines an encoder–decoder architecture with a multi-label–aware sampling objective, enabling end-to-end generation of diverse synthetic minority instances through thresholded decoding of seeds from minority labels. Empirical results on nine benchmark datasets show that AEMLO provides consistent gains across multiple base learners and metrics (Macro-F, Macro-AUC, Ranking Loss) compared to existing sampling methods. The approach serves as a data-level balancing solution that can augment any multi-label classifier and is available as open-source software, promoting practical adoption in imbalanced MLC tasks.

Abstract

Class imbalance significantly impacts the performance of multi-label classifiers. Oversampling is one of the most popular approaches, as it augments instances associated with less frequent labels to balance the class distribution. Existing oversampling methods generate feature vectors of synthetic samples through replication or linear interpolation and assign labels through neighborhood information. Linear interpolation typically generates new samples between existing data points, which may result in insufficient diversity of synthesized samples and further lead to the overfitting issue. Deep learning-based methods, such as AutoEncoders, have been proposed to generate more diverse and complex synthetic samples, achieving excellent performance on imbalanced binary or multi-class datasets. In this study, we introduce AEMLO, an AutoEncoder-guided Oversampling technique specifically designed for tackling imbalanced multi-label data. AEMLO is built upon two fundamental components. The first is an encoder-decoder architecture that enables the model to encode input data into a low-dimensional feature space, learn its latent representations, and then reconstruct it back to its original dimension, thus applying to the generation of new data. The second is an objective function tailored to optimize the sampling task for multi-label scenarios. We show that AEMLO outperforms the existing state-of-the-art methods with extensive empirical studies.

Figures (9)

• Figure 1: Using Autoencoders to train data.
• Figure 2: The architecture of the proposed autoencoder learns the latent space $\mathcal{L}$ through the function of $\mathbf{F}_{ex}$ and $\mathbf{F}_{ey}$, and decouples $\mathcal{L}$ through $\mathbf{F}_{dx}$ and $\mathbf{F}_{dy}$.
• Figure 3: The process of generating instances
• Figure 4: The performance of the multi-label sampling methods in terms of Macro-F across five different classification methods.
• Figure 5: The performance of the multi-label sampling methods in terms of Macro-AUC across five different classification methods.