SimAndro: an effective method to compute similarity of Android applications

As the number of Android applications (apps) is increasing dramatically, users face a serious problem to find relevant apps to their needs. Therefore, there is an important demand for app search engines or recommendation services where developing an accurate similarity method is a challenging issue. Contrary to malware detection, very fewer efforts have been devoted to similarity computation of apps. Furthermore, all the existing methods use the features obtained only from the app stores such as description and rating, which could be inaccurate, varied in different stores, and affected by language barrier; they totally neglect useful information clearly capturing the app’s functionalities and behaviors that can be mined from the apps themselves such as the API calls and manifest information. In this paper, we propose an effective method called SimAndro to compute the similarity of apps, which extracts the features based on the information obtained only from apps themselves and the Android platform without using information obtained from third-party sources such as app stores. SimAndro performs both feature extraction and similarity computation where the API calls, manifest information, package name, and strings are used as features. To compute the similarity score of an app-pair, a separate similarity score is computed based on each feature, and a weighted linear combination of these four scores is regarded as the final similarity score by utilizing an automatic weighting scheme based on TreeRankSVM. The results of extensive experiments with three real-world datasets and a dataset constructed by human experts demonstrate the effectiveness of SimAndro.