Preview

PDF, English Download (4MB) | Terms of use

Abstract

Context: Software development is an iterative process requiring continuous adaptation to user needs. However, a gap often exists between developers’ assumptions and users’ actual expectations for the software. While direct user participation is valuable for bridging this gap, practical constraints often make it difficult. Online user feedback provides an alternative source for insights but is typically unstructured and lacks context. To bridge the gap between developers and users through online feedback, two key challenges need to be tackled: (1) identifying which functionalities users discuss in their feedback and (2) understanding how users interact with these functionalities. Moreover, manually analyzing large volumes of feedback is time-consuming, highlighting the need for automation and tool support. Objective: The goal of this thesis is to introduce two approaches to tackle the above mentioned challenges. One approach, handling challenge (1), facilitates the relation of feedback and existing requirements of a software. The functionalities of a software are documented in its requirements. By relating feedback directly to the requirements, developers are able to identify the discussed functionalities of the software. The other approach, handling challenge (2), allows the analysis of users’ needs and expectations by extracting the usage information in their feedback. Usage comprises both the real-life actions of the users as well as the interactions of the user with the software. This usage information is captured through the application of the TORE framework on the user feedback and allows developers to gain a better understanding of how the users interact with the software’s functionalities. For both of these approaches, the relation and the usage information analysis, this thesis offers machine learning classifiers and tool support to reduce the manual labour required for the analysis. Methods: This thesis follows the Design Science approach consisting of solution investigation, treatment design and treatment validation. The solution investigation is conducted via two systematic mapping studies in order to identify existing machine learning classifiers and evaluate their applicability to feedback requirements relation and usage information classification. The treatment design contains goals that match the two approaches presented in this thesis: The design and implementation of an approach and accompanying automatic classifiers to relate feedback to existing requirements and to identify the usage information contained in feedback. The implemented classifiers are also evaluated on multiple manually created datasets to identify the best-performing ones. Additionally, a software prototype is presented as part of the treatment design, which offers tool support for the developed approaches. The treatment validation evaluates the developed classifiers in the context of a hypothetical deployment scenario in a company. Contributions: The main contributions of this thesis are the presented approaches for feedback requirements relation and usage information analysis as well as the classifiers for the automation of these approaches and their evaluation. Multiple manually created datasets are also provided to train and evaluate the presented classifiers. Furthermore, two mapping studies are included, which capture the current state of research towards the relation of software artifacts and detailed user feedback classification. Additionally, a prototype (Feed.UVL) is created to provide tool support for the developed approaches. A Jira plugin is also provided to integrate the tool support for the approaches into existing development workflows.