Software Effort Estimation Using Deep Learning and Fuzzy Modelling

Building a successful software application with high quality needs a better software process model. One of the factors that impact software development process is estimating the most likely human effort needed to accomplish software project. Machine learning and statistical algorithms have been widely used to build such estimation model. But a little attention has been paid to the applicability of Deep Convolutional Neural Network to better estimate software effort. One reason that hinder using deep learning is that most of the software datasets contains samples in the form of vector, not a matrix. To handle this challenge and reduce uncertainty in software measurement, we use Fuzzy modelling and Fuzzy c-means clustering to build a convenient data form for each instance. Simply, we used Fuzzy c-means to partition samples in the dataset into different clusters then map these cluster on each feature dimension using the concept of Fuzzy membership function. This process helps yielding a matrix representation for each data sample, which can fit as input to the deep learning model. The proposed model has been evaluated over multiple benchmark datasets obtained from PROMISE repository. The results are promising and in general more adequate than many popular estimation models. However, there is a need to investigate the impact of changing number of clusters and features on the proposed model.