Hybrid deep learning approach for software defect prediction: Diffusion kernel attention and gudermannian networks

Abstract

Software defect prediction (SDP) is an important endeavor in software engineering which is defined as the exploration of possible defects in a software module before it is released for functioning so as to minimize the time consuming and resource in taking use of referred to as software maintenance and to ensure quality control. Static approaches used in most previous studies of SDP are based on the use of exclusively quantitative parameters and standard machine learning (ML) approaches, which severely hampers the ability to model nonlinear dependencies between metrics and thus provides low accuracy of picked features and low possibilities for generalization. To meet these challenges, this work presents a new Diffusion Kernel Attention Network integrated with Gudermannian Neural Networks employing Fire Hawk Optimizer, abbreviated as DKANGNN-FHO. From the collection of six open source formationary Java projects in the PROMISE dataset, which engages diverse applications and defect rates, data acquisition forms the initial pragmatic workflow step. Preprocessing also includes techniques that require scaling of features such as effort, time, and complexity uniformly. Feature extraction utilizes the Geometric Algebra Transformer (GATr) to saturate software metrics as multivectors and perform dimensionality reduction through attention maps and geometric tokenization. A dependency structure is captured by the phrase ‘DKANGNN’, while highest achievable hyperparameters reliably come from ‘FHO’. The method developed in this paper has better accuracy of 99.4% and excellent stability as compared with the existing techniques and with the help of F1 score of the method 99.8%, Specificity of the method is 99.6% and MSE of the proposed is 0.1%.