Improved Data Segmentation Architecture for Early Size Estimation using Machine Learning

Abstract

Software size estimation plays an important role in project management. According to the report given by Standish Chaos, about 65% of software projects are beyond companies budget or overdue; which could have been saved if an early estimation was imposed. Though the software size can’t be measured directly, but it is related to effort and hence a low effort will lead to low size. The calculation of effort depends upon how the data is organized or segmented. This research paper focuses on the improvement of data segmentation in order to reduce the effort and parallel the size. In order to improve the segmentation architecture, the project data is divided based on the similarity indexes which the projects have in between their attributes. Three similarity measures were used namely Cosine Similarity (CS), Soft Cosine Similarity (SC) and a hybrid similarity index which combines CS and SC. Based on these similarity indexes, the project data is divided into groups by K-means algorithm. In order to estimate the size, the co-relation between the formed groups are calculated. To calculate the correlation, Mean Square Error (MSE), Square Error (SE), and Standard Deviation (STD) is calculated and the normalized parameters are used to evaluate the software size.