Classification And Analysis Of Non-Functional Requirements In App User Reviews Based On The System Model

With the rapid popularization of mobile devices,mobile apps and app stores are widely used in our daily life.Users can not only download and install the apps in the app stores,but also provide their feedback on the apps by giving scores and uploading reviews.User reviews usually contain various types of information,such as feature requirements,bug reports,and user experience.As an important source to elicit user requirements and other useful information,user reviews can help software developers identify missing features and improve software quality.Existing research on eliciting requirements from user reviews mainly focuses on functional aspects.However,nonFunctional Requirements(NFRs)from diverse sources are closely related to software quality,and play a critical role during software development.The most common used approach for NFRs classification is to categorize NFRs into different quality characteristic classes according to the quality model in software quality requirements and evaluation standards.Broy proposed a new view to rethink the nature of NFRs.According to the structured system model,NFRs can be divided into behavioral requirements that describe the behavioral properties of the system and representational requirements that describe the system representation.Then behavioral requirements can be further classified as different classes based on the system view and the behavior theory.NFRs classification based on the system model can make the representation of requirements more precise and explicit.Based on the above research background,this thesis conducted an exploratory and comparative study on NFRs in App user reviews.We first identified different quality characteristic classes of NFRs according to the quality model,and further classified and analyzed NFRs from the perspective of the system model,aiming to get a deeper understanding on the nature of NFRs in App user reviews and the desired properties of the system they describe.We also compared the distribution of NFRs in App user reviews and industrial requirements specifications to explore the similarities and differences in the distribution of NFRs from different sources.In addition,using our manual classification on NFRs based on the system model,this thesis carried out automatic classification experiments on the dataset.We combined the classification technique TF-IDF with two classic machine learning algorithms Na(?)ve Bayes and Logistic Regression,and combined the classification technique Word2 Vec with the deep learning text classification algorithm Text CNN,which constitute three classification models.Then we conducted experiments to evaluate their performance on NFRs classification.The research and experiment dataset in this thesis come from the user reviews of i Books in Apple store and Whats App in Google Play.The research results of this thesis show that(1)in App user reviews,users primarily report quality aspects on Reliability and Usability.Over 50% NFRs address interface behavior of the systems(the same type of behavior as functional requirements do),and thus we concluded that most NFRs in App user reviews are essentially not non-functional;(2)overall the distributions of NFRs with respect to the system view and the behavior theory in App user reviews and industrial requirements specifications are similar,but the distributions of NFRs classified as architecture and state,and the distributions of NFRs in certain quality characteristic classes(e.g.,the Usability class)show a big difference;(3)when performing automatic classification experiments on the dataset based on the system model,all the three classification models used in this thesis achieve good performance on NFRs classification.Compared with the traditional machine learning algorithms Na(?)ve Bayes and Logistic Regression,the classification model that combines Word2 vec with deep learning algorithm Text CNN has the best classification performance,with a highest F-Measure(0.960).