ANALYSIS OF MOBILE APPLICATION USER PREFERENCES BASED ON MACHINE LEARNING METHODS

Abstract

This study examined classical machine learning methods and neural networks for analyzing user preferences in mobile applications. The target variable and preference criterion was the average app rating. A dataset from the open source Kaggle was used, followed by data cleaning and preprocessing. A comparative analysis was conducted on three classical machine learning methods (linear regression, random forest, XGBoost) and three neural network models (ANN, CNN, RNN) to predict users’ average app ratings based on seven features. As the dataset was relatively small and of simple structure, some neural network models could not fully realize their potential. The XGBoost model demonstrated the best performance, highlighting its usefulness for this type of data. The CNN model performed slightly worse, as it is designed to capture significant patterns in complex datasets. The most important features for predicting user ratings were identified, including types, installations, genres, categories, and others. In future work on decision-making tasks aimed at improving user engagement, this study can assist in selecting appropriate models and input features to focus on when designing or enhancing an application.