Enhancing DCNN model performance: Experimental analysis based on data augmentation

Deep Convolutional Neural Networks (DCNNs), as a state-of-the-art machine learning technique, have attracted significant attention due to their remarkable performance in data classification, clustering, and pattern recognition. Owing to the DCNN model’s ability to utilize convolution kernels for feature extraction from input data while effectively exploiting local correlations enables it to achieve optimal prediction results, particularly when forecasting large datasets. However, practical scenarios often involve small sample sizes, which significantly diminishes model accuracy. In response to this limitation, this study introduces three data augmentation algorithms aimed at enhancing DCNN precision in contexts with limited sample data. Notably, the Generative Adversarial Network (GAN) augmentation algorithm not only expands the dataset but also incorporates a discriminator to assess the authenticity of generated samples in comparison to the original ones. This methodology ensures that the augmented data closely approximates real samples, thereby facilitating superior predictive performance by the DCNN model. By integrating three oversampling methods with DCNN and applying them to the identification of acid radical ion ligand binding sites, we trained the DCNN using augmented datasets produced through various oversampling techniques. The predictive model derived from the training dataset was independently validated on the test set, revealing that the outcomes achieved with the GAN-DCNN algorithm were optimal and surpassed previous predictions. This finding underscores the exceptional effectiveness of GAN-DCNN in addressing small sample classification challenges.