Automatic Classification of Mobile Apps Reviews for Requirement Engineering

We applied machine learning techniques in classifying health care application reviews into several types such as bug reports, new feature requests, application performance, and accuracy and user interface. There is no available free annotated data set for training and evaluating machine learning techniques, therefore, more than 7500 reviews for 10 different health-related mobile applications are annotated manually by experts in the field. Our experiments show that Multi-nominal NaiveBays can classify mobile apps reviews into bugs, new features and sentimental with an accuracy of 87%, and into a general bug, usability, security, and performance with an accuracy of 88%. The best result of the sentimental analysis system is 90%. in addition, the experiments show that the overall performance is improved when we use the data subset with high confidence labels and when two experts agree on the same label. The Re-sampling technique was successfully used to overcome the data imbalance problem in our data set, the accuracy improved to 89% for mobile apps reviews into a set of classes; bugs, security, new feature, performance, and usability and 96% for the sentimental reviews.