Genetically Modified Logistic Regression with Radial Basis Function for Robust Software Effort Prediction

Abstract

Existing logistic regression technique applies log-based function on the dataset and provides the necessary result for analysis. It has been for years for prediction analytics. This project develops a genetically modified evolutionary logistic regression technique using radial basis functions (GLR-RBF). This methodology employs three crucial stages. First stage is an evolutionary stage employing two fitness functions. Initially, the dataset obtained from the repository is modified using Genetic Algorithm (GA), followed by creation and simulation of radial basis neural network (RBFNN). The RBFNN is iteratively trained and simulated. The weight matrix generated in each stage is tested with a fitness function, which define the rule for selection of the best individual that will be appended to the dataset. The process of appending the weight matric to the dataset marks the second stage. In the third stage, the new set of covariates is fed into logistic regression operator which classifies the attributes based on effort response variable. This method has been tested in field of remote sensing image classification and other large dataset. This project also tests this technique for software effort prediction, which is regarded as the crucial activity before the commencement of the software development cycle. The model GLR-RBF was found be competitive when compared with similar standard models. The measure obtained from AUC indicates that GLR-RBF has reached the state-of-the-art.