ENHANCED USER SENTIMENT PREDICTION FROM MULTILINGUAL MOBILE APP REVIEWS USING ENSEMBLE LEARNING

Abstract

Mobile applications generate a vast amount of multilingual user reviews, which provide valuable insights into user satisfaction and product improvement. Current statistics indicate that over 80% of smartphone users rely on reviews before installation, while approximately 75% of developers use app store feedback to guide updates. However, manual sentiment classification of such multilingual reviews is highly timeconsuming and error-prone, while existing single-model approaches often fail to capture contextual nuances across diverse languages. The proposed methodology leverages Natural Language Processing (NLP) on multilingual mobile review datasets with advanced preprocessing and feature extraction techniques to standardize text data across multiple languages. Exploratory Data Analysis (EDA) is performed to identify sentiment distribution and user behavior patterns. For Sentiment Classification Analysis, four classifiers are employed: Logistic Regression (linear classification), Ridge Classifier (regularized classification), and an Ensemble Classifier that combines predictions using a voting mechanism for improved accuracy. Parallelly, Rating Prediction Analysis is conducted using regression models, including Linear Regression (basic model), Ridge Regression (regularized regression), and an Ensemble Regressor that integrates results for robust numerical predictions on a 1–5 scale. Finally, a Universal Real-time Prediction framework is introduced, where a user can input a review in any language and instantly obtain both the sentiment polarity (positive, negative, neutral) and predicted star rating (1–5). This dual-layer prediction not only enhances user sentiment tracking but also optimizes developer feedback mechanisms, enabling faster updates and targeted improvements. The proposed system demonstrates significant advancement in multilingual adaptability, scalability, and real-time prediction capabilities.