Machine Learning and Ensemble Approaches for Sentiment Classification of Mobile Application Reviews

Abstract

Mobile application marketplaces generate huge volumes of user reviews daily.  Automatically understanding if a review is positive or negative matters to developers, platform operators, and prospective users alike, since manually reading them one by one is not feasible. This paper reviews the empirical and methodological literature on machine-learning-based sentiment classification, with particular attention to studies using Google Play Store review data. It synthesises a substantial body of empirical work and examines the preprocessing techniques that recur across it, which includes tokenization, count vectorization, TF-IDF, and transformer-based embeddings. It also discusses, accuracy, precision, recall, F1-score, and AUC-ROC, and also reviews the core classifiers used. The Support Vector Machines, Multinomial Naïve Bayes, AdaBoost, and ensemble strategies where applied. Three major challenges with the classification of Google Play Store where also discussed. The challenges include the systematic handling of emojis, principled handling of class imbalance, and statistical validation of reported performance gains. Together, these gaps motivate further research into ensemble architectures that combine classical classifiers with modern contextual embeddings for domain-specific, imbalanced, and informally written review text.