A Generative Neighborhood-Based Deep Autoencoder with an Extended Loss Function for Robust Imbalanced Classification

Deep learning models have demonstrated remarkable performance in classification tasks; however, real-world applications often grapple with constraints such as limited labeled data and significant class imbalance. These constraints can result in unstable predictions and reduced performance. To tackle this challenge, three distinct approaches have emerged: data-level methods, model-level methods, and hybrid methods. Data-level methods make use of generative models, typically grounded in Generative Adversarial Networks, which rely on extensive data resources. In contrast, model-level methods leverage domain expertise and may be less accessible to users lacking such specialized knowledge. Hybrid methods combine elements of both these approaches. In this work, we introduce GENDA-XL, a generative neighborhood-based deep autoencoder featuring an extended loss function. GENDA-XL places emphasis on learning latent representations via supervised similarity learning and it integrates a pretrained classification model to associate each generated sample with its corresponding label. Through comprehensive experiments conducted across various image and time-series datasets, we illustrate the effectiveness of our method.