Utilization of Adaptive Machine Learning for Streaming Sentiment Analysis: The Effects of Batch and Drift Types

Abstract

The changing data patterns that continue to emerge from user reviews on digital platforms demand machine learning models to be able to adapt sustainably. This phenomenon, known as concept drift, can degrade accuracy if not handled appropriately. The study explores two adaptive learning approaches—Adaptive Random Forest (ARF) and Adaptive XGBoost (AXGB)—to classify sentiment on real-time data streams. The dataset used includes OVO reviews (2016–2025) and GoPay reviews (2023–2025) from the Google Play Store. After performing text cleanup, tokenization, stemming, and Term Frequency–Inverse Document Frequency (TF-IDF) representation, the data are streamed into two learning schemes: Big Batch Small Batch (BBSB), which combines large and small batches to detect changes more subtly, and N Drift Types Batch (NDTB), which adjusts batch sizes based on the number of drift types considered. The evaluation was carried out using accuracy metrics, computational time, and true positive rate in each sentiment category. The results showed that ARF provided the most stable performance with an average accuracy of 82%, whereas AXGB performed better in the BBSB configuration with an accuracy of 81% and faster training time. These findings confirm that both configurations can serve as practical frameworks for text stream learning and adaptive concept drift management.