A Predictive Analysis of Customer Use of Mobile Fitness Application based on Hybrid Machine Learning

Abstract

The researcher proposed an investigation of customer reviews of mobile fitness apps listed on the Google Play Store. This study aims to provide a nuanced understanding of user sentiments beyond simple star ratings. Additionally, the researcher highlighted the importance of considering the specific challenges in analysing app reviews. The method adopted for analysing mobile fitness app reviews started with web scraping from the Google Play Store to systematically collect review data of the top 100 mobile fitness applications, of which 30 were selected based on more than 1500 reviews per application. Sentiment scoring and aggregation methods were deployed on cleaned datasets, utilizing natural language processing techniques to regulate personal evidence and emotional tone in each review. Such a process calculates individual sentiment scores by applying ensemble machine learning methods, including Random Forest, Extra Trees, AdaBoost, Gradient Boost, and XGBoost, which enhance the accuracy and reliability of sentiment analysis results. The accuracy of each output was measured using the Rank, Accuracy, AUC, Precision, Recall, and F1-score for each application. Comprehensive analysis depicted Extra Trees and XGBoost as the best models. The maximum accuracy of 89.1% indicates that the assumption of a gap between numeric and sentiment ratings is significant, which was further confirmed by applying a t-test on the outputs of each algorithm. This research has been effective in determining XGBoost as the best possible method that effectively controls overfitting and underfitting. The findings suggest that the XGBoost and Extra Trees algorithms demonstrate superior performance in predicting sentiment ratings from numeric ratings. The high accuracy and ability of these models to control overfitting and underfitting makes them particularly suitable for this type of analysis. Future research could explore the application of these algorithms in other domains or investigate potential improvements to enhance their predictive capa