Fine-Grained Categorization of Mobile Applications Through Semantic Similarity Techniques for Apps Classification

The number of Android apps is constantly on the rise. Existing stores allow selecting apps from general named categories. To prevent miscategorization and facilitate user selection of the appropriate app, a closer examination of the categories’ content is required to discover hidden subcategories of apps. Recent work focuses on exploring the granularity of the categories, but a validation of the categories’ content against miscategorized apps is missing. In this research, we apply semantic similarity to apps’ descriptions to uncover similarity and hierarchical clustering to search for misclassified apps. Furthermore, we apply Latent Dirichlet Allocation (LDA) algorithm to explore the existence of possible subcategories and to classify apps. Our empirical research is conducted using two data sets: 9,265 apps from Google Play Store, and 300 apps from App Store. Results confirm the existence of misclassified apps on markets and suggest the existence of multiple fine-grained categories. Our experiments outperform other LDA-based classification approaches achieving 0.61 precision. Moreover, the analysis hints the presence of misclassified apps might decrease the performance of existing classifiers.