<p>Assessing credit risk using machine learning tools is critical for minimizing bank losses, which can arise from financial defaults or missed opportunities. Since financial losses are more detrimental than missed opportunities, this paper proposes an integrated approach for imbalanced credit scoring. We leverage a Variational Autoencoder (VAE) with an Integrated Neural Network classifier (VAE-INN) designed to address class imbalance during feature extraction. Unlike traditional methods that separately apply data augmentation and feature selection, potentially altering the original data distribution or introducing biases, our method inherently tackles class imbalance within the latent space, guided by a weighted loss function. While traditional methods attempt to address imbalance through separate steps such as data augmentation or feature selection, these approaches risk distorting the data distribution or discarding features crucial for identifying minority class instances. In particular, feature selection, commonly used to enhance class separability, can inadvertently favor the majority class when applied to imbalanced data. Our method promotes fairer and more robust classification by shaping the latent space to represent both classes equally. We validate our approach on an imbalanced real-world dataset, focusing on reducing Type II errors. Experimental results show that our VAE-INN model with weighted loss significantly outperforms fragmented techniques, achieving substantial Type II error reduction and demonstrating robustness under severe imbalance ratios, thus enhancing the detection of true credit defaults and mitigating financial risk.</p>