Sniffing Android Code Smells: An Association Rules Mining-Based Approach

Abstract

Interest in mobile applications (mobile apps) has grown significantly in recent years and has become an important part of the software development market. Indeed, mobile apps become more and more complex and evolve constantly, while their development time decreases. This complexity and time pressure might lead developers to adopt bad design and implementation choices, which are known as code smells. Code smells in mobile apps could lead to performance issues such as overconsumption of hardware resources (CPU, RAM, battery) or even downtime and crashes. Some tools have been proposed for the detection of code smells in Android apps, such as Paprika or a Doctor tools. These tools rely on metrics-based detection rules, which are defined manually according to code smell definitions. However, manually defined rules might be inaccurate and subjective because they are based on user interpretations. In this paper, we present a tool-based approach, called Fakie, which allows the automatic inference of detection rules by analysing code smells data using an association rules algorithm: FP-Growth. We validated Fakie by applying it on a manually analysed validation dataset of 48 opensource mobile apps. We were able to generate detection rules for a dozen code smells, with an average F-measure of 0.95. After all of that, we performed an empirical study by applying Fakie on 2,993 apps downloaded from AndroZoo, a repository of mobile apps.