APPGRADER : AN APP QUALITY GRADING SYSTEM BASED ON CODE-LEVEL FEATURES
- Author:
- Li, Xi
- Published:
- [University Park, Pennsylvania] : Pennsylvania State University, 2018.
- Physical Description:
- 1 electronic document
- Additional Creators:
- Zhu, Sencun
- Access Online:
- etda.libraries.psu.edu
- Restrictions on Access:
- Open Access.
- Summary:
- The current app ranking systems applied by app markets are mainly based on app rating and downloads. However, these systems have drawbacks in handling: (i) apps with abnormally high ratings and fake downloads; (ii) newly published apps with limited user feedback. Rankings of these apps may not accord with their actual quality, which will mislead users. Therefore, in an attempt to explore app ranking systems and change the under studied status quo of it, we propose AppGrader, a novel app quality grading system that ranks apps under the same category based on app functionality measured by code-level features. This system is inspired by the analysis on 18 millions app reviews which suggests that while giving ratings, most users may consider user interface and other features that can be extracted directly from app code. Therefore, our system statically analyzes app code and generates ``feature view graph'' for each app which encodes app code-level features. For app ranking, we apply Graph Convolutional Network to cluster apps into different classes based on the complexity of their corresponding feature view graphs, where each class indicates one level of app quality. According to the system evaluation from two perspectives: system accuracy and label dissimilarity, AppGrader performs well on 1440 real world apps with average accuracy of around 72\% and label dissimilarity of around 1, which indicates that AppGrader could be applied for evaluating apps with fake ratings and newly published apps.
- Genre(s):
- Dissertation Note:
- M.S. Pennsylvania State University 2018.
- Technical Details:
- The full text of the dissertation is available as an Adobe Acrobat .pdf file ; Adobe Acrobat Reader required to view the file.
View MARC record | catkey: 24468541