Integrasi Aspect-Based Sentiment Analysis dalam Recommender System Berbasis Ulasan untuk Mengurangi Information Overload: A Systematic Literature Review

Abstract

The large volume of online reviews and the limitations of numerical ratings in capturing consumer preferences in detail have given rise to information overload, prompting the integration of Aspect-Based Sentiment Analysis (ABSA) into review-based recommender systems. This study aims to map the ABSA methods, recommender system architectures, and application domains used in such integrations, along with their reported performance, contributions, and limitations, over the 2021–2025 period. The study was conducted using a systematic literature review approach following the PRISMA 2020 guidelines. Searches across five electronic databases (ScienceDirect, Springer, Scopus, ACM Digital Library, and IEEE Xplore) initially yielded 169 articles, which were then narrowed down through deduplication, screening, and eligibility assessment to 20 studies meeting the inclusion criteria for narrative synthesis. The results show that deep learning methods, particularly Bidirectional Long Short-Term Memory (Bi-LSTM) with attention mechanisms and BERT-based approaches, are the most dominant ABSA methods, while Collaborative Filtering, either independently or hybridized with Content-Based Filtering, is the most widely adopted recommendation architecture across domains such as e-learning, e-commerce, restaurants, hospitality, and social platforms, with performance varying across RMSE, MAE, F1-score, and AUC metrics. The main contributions of this integration include improved prediction accuracy and interpretability, as well as reduced cold-start and data sparsity issues, although limitations related to manual annotation requirements and cross-domain generalization remain evident in most studies.