Comparative analysis of requirements change prediction models: manual, linguistic, and neural network

Requirement change propagation, if not managed, may lead to monetary losses or project failure. The a posteriori tracking of requirement dependencies is a well-established practice in project and change management. The identification of these dependencies often requires manual input by one or more individuals with intimate knowledge of the project. Moreover, the definition of these dependencies that help to predict requirement change is not currently found in the literature. This paper presents two industry case studies of predicting system requirement change propagation through three approaches: manually, linguistically, and bag-of-words. Dependencies are manually and automatically developed between requirements from textual data and computationally processed to develop surrogate models to predict change. Two types of relationship generation, manual keyword selection and part-of-speech tagging, are compared. Artificial neural networks are used to create surrogate models to predict change. These approaches are evaluated on three connectedness metrics: shortest path, path count, and maximum flow rate. The results are given in terms of search depth needed within a requirements document to identify the subsequent changes. The semi-automated approach yielded the most accurate results, requiring a search depth of 11 %, but sacrifices on automation. The fully automated approach is able to predict requirement change within a search depth of 15 % and offers the benefits of full minimal human input.