Improving the review classification of Google apps using combined feature embedding and deep convolutional neural network model.

Online reviews play an integral part in making mobile applications stand out from the large number of applications available on the Google Play store. Predominantly, users consider posted reviews for appropriate app selection. Manual categorization of such reviews is both inefficient and time-consuming. Therefore, automatic analysis of the sentiments of such reviews provides fast suggestions for new users and facilitates their selection of the appropriate app. However, data imbalance is a major challenge for performing class prediction of such reviews as their distribution is sparse and often leads to low accuracy. This work proposes a framework to overcome this limitation. Extensive experiments are performed using the original and balanced data with the synthetic minority oversampling technique (SMOTE) and adaptive synthetic sampling (ADASYN). Additionally, deep learning and machine learning models are evaluated using FastText, FastText Subword, global vector (GloVe), and their combinations for word representation. Baseline machine learning models, including random forest, extra tree classifier, gradient boosting, Naive Bayes, logistic regression (LR), stochastic gradient descent (SGD), and voting classifier (VC) that combines LR and SGD, are used for comparison. The outcomes show that the convolutional neural network using a combination of word embedding techniques produces the most accurate results. [ABSTRACT FROM AUTHOR]