Research On The Evaluation Of App Description

With the development of Android operating system, Android applications (Apps) release faster and faster, meanwhile, the downloads of Android applications increase with a high speed. The number of Android applications in Google Play has been more than 2 million in 2016. In the Google Play Store, an introduction page is associated with every Android application for users to acquire its details, including the name, developers, screenshots, description, reviews, etc. To attract more downloads, there are many guidelines about how to write a good app description to drive more downloads. However, most of these guidelines are too abstract or broad, and it is hard to define a high quality or good app description. Meanwhile, there is no tool to evaluate the quality of app description for providing some advice for app developers.To evaluate the quality of app description, this paper constructs attributes by data-driven method to train evaluation models. The datasets of this paper are app descriptions from 5 categories of Google Play, namely Music & Audio, News & Magazines, Photography, Travel & Local, Weather,100 Apps per category. Then, we invite 30 volunteers to evaluate these app descriptions and explain reasons. Next, we construct some attributes from these reasons to train the machine learning models with the quality levels mapped by ratings from volunteers.With the reasons provided by volunteers for the ratings of app descriptions, we construct 16 features in total and determine how to calculate their values based on our experience and related research work. Next, we choose support vector machine, decision tree, random forest and logistics to train all the data in the sample with all the features values as input and the quality levels as output. Finally, the support vector machine obtains up to 58% accuracy.In addition, as all the features are constructed by us, we analyze their importance with the feature selection tool of LibSVM and decision tree models of Weka, the results of unsupervised learning model agree with that of supervised model in terms of the order about feature importance. There are some features which are important for all the categories, namely the length of app description (the number of words), the difficulty of every words (the number of character per word), the length of sentences and the ratio of feature description with all the app description. We expect that our results can be helpful for app developers when preparing app descriptions before releasing a new app.