This page provides details about the paper:   Recommending New Features from Mobile App Descriptions.

        The training datasets are prepared for classifying whether a sentence describes a feature or not. There are two files which contain instances of feature and instances of non-feature respectively, and each line stands for an instance. The datasets are obtained from Google Play. To make it representative, we select the descriptions of the top ranked Apps for each category. To make it more accurate, we manually split the sentences in the descriptions into two classes: features and non-features. Please refer to the paper for more information. We hope that these datasets can help those ones who perform the same task.

        The datasets can be downloaded here: the dataset of features and the dataset of non-features.

        Since there is no dataset containing features for Apps, we have volunteers annotate a collection of features of Apps to evaluate the performance of feature recommenders. We download totally 8359 Apps in five categories from Google Play. Due to the large number of Apps in each category, we cannot annotate all the Apps. Hence, we randomly select 20 Apps from every category and employ volunteers to annotate their golden features. Totally, we achieve 533 golden features.

        The feature dataset can be downloaded here: FDA.

        It is non-trivial to extract features from the descriptions of Apps in free text. Hence, we develop a tool named App Feature Extractor (AFE), which consists of three components, namely data cleaner, linguistic rules filter, and feature classifier.

        The source code of AFE can be downloaded here: AFE.

        In the paper, we use Naive Bayes (NB) as the default classifier in AFE. To show whether it is effective, we introduce several commonly used classifiers to compare, i.e., Decision Tree (DT), Support Vector Machine (SVM), Random Forest (RF), and AdaBoost. Among them, DT and SVM are single classifiers. In contrast, RF and AdaBoost are ensemble classifiers. By comparing NB with these commonly used classifiers, we can know whether selecting NB as the default classifier is reasonable. The technical report shows the detailed results of these classifiers: The Comparison between different classifiers.

        To explore the real relationships among features, we download the descriptions of the top 10 ranked free and paid Apps from Google Play to analyze. We manually classify the feature-describing sentences into independent features and inter-independent feature. From the results, we can see that independent features take up about 80%. The technical report shows the detailed results of the relationships among features: Feature statistics.

        In the paper, we set the parameters in SAFER to some predefined values. To explore their effectiveness, we conduct several experiments by comparing the default values with some other values. We present the results in a technical report: Parameter selection.

        If you think these materials are useful for you, please cite our paper, Thank you!

        If you have any questions about our datasets and our paper, please contact Jingxuan Zhang. E-mail: jingxuanzhang@mail.dlut.edu.cn