Modeling Users Feedback Using Bayesian Methods for Data-Driven Requirements Engineering

Abstract

Data-driven requirements engineering represents a vision for a shift from the static traditional methods of doing requirements engineering to dynamic data-driven user-centered methods. App developers now receive abundant user feedback from user comments in app stores and social media, i.e., explicit feedback, to feedback from usage data and system logs, i.e, implicit feedback. In this dissertation, we describe two novel Bayesian approaches that utilize the available user's to support requirements decisions and activities in the context of applications delivered through software marketplaces (web and mobile). In the first part, we propose to exploit implicit user feedback in the form of usage data to support requirements prioritization and validation. We formulate the problem as a popularity prediction problem and present a novel Bayesian model that is highly interpretable and offers early-on insights that can be used to support requirements decisions. Experimental results demonstrate that the proposed approach achieves high prediction accuracy and outperforms competitive models. In the second part, we discuss the limitations of previous approaches that use explicit user feedback for requirements extraction, and alternatively, propose a novel Bayesian approach that can address those limitations and offer a more efficient and maintainable framework. The proposed approach (1) simplifies the pipeline by accomplishing the classification and summarization tasks using a single model, (2) replaces manual steps in the pipeline with unsupervised alternatives that can accomplish the same task, and (3) offers an alternative way to extract requirements using example-based summaries that retains context. Experimental results demonstrate that the proposed approach achieves equal or better classification accuracy and outperforms competitive models in terms of summarization accuracy. Specifically, we show that the proposed approach can capture 91.3% of the discussed requirement with only 19% of the dataset, i.e., reducing the human effort needed to extract the requirements by 80%.