DDA SMOTE: Deep Distance-Aware SMOTE Technique for Imbalance Data

Class imbalance problems can cause significant challenges in machine learning and deep learning applications. While earlier machine learning methods have relied on simple oversampling methods to handle this problem, such a strategy does not map well into the existing deep learning frameworks, especially in fields like computer vision, where acquiring large, high-quality image datasets. Although many oversampling techniques have been developed to address imbalanced data problems, most of these methods fail to preserve the original characteristics of complex data, such as images, which is crucial for deep learning models. We introduce a Deep Oversampling Technique, a new oversampling algorithm designed specifically for deep learning architectures. We aimed to develop an effective, straightforward model that integrates the synthetic minority oversampling technique (SMOTE) method with a deep learning approach that can produce meaningful high-quality images. Our method consists of the following steps: ۱. Train the Autoencoder on the imbalance dataset; ۲. Incorporate a penalty loss alongside the reconstruction loss for the autoencoder; ۳. Implement synthetic minority oversampling technique (SMOTE) for oversampling and ۴. Optimizing generated features and leveraging the trained Autoencoder to create new images thereby balancing the dataset. The main advantage is that this method doesn’t need a discriminator, as typically required in generative adversarial models. Moreover, it employs optimization to improve the quality of synthetic data while aiming to avoid generating overly similar samples.

کلیدواژه ها:

Imbalance data ، Deep learning ، Oversampling ، optimization ، image generation ، Autoencoder