Analisis Sentimen Ulasan Game dengan KNN: Perbandingan Rating dan Kamus Sentimen

Abstract

The growth of the global gaming industry makes sentiment analysis of user reviews a crucial tool for understanding satisfaction and identifying technical issues. This study aims to evaluate three labelling methods (rating-based, Sentiwords_id, and InSet) for classifying the sentiment of Indonesian-language reviews for the game Zenless Zone Zero (ZZZ) using the K-Nearest Neighbor (KNN) algorithm. The study analyzes 4,282 reviews from the Google Play Store, which underwent a Data Preprocessing stage, including Null Handling, Cleaning, Case Folding, Tokenization, Stopword Removal, and Stemming. The KNN's performance for each labelling method was evaluated using accuracy, precision, recall, and F1-score metrics on 80:20 train-test split. The labelling results reveal different sentiment perceptions: the rating-based method tends toward positive, InSet toward negative, while Sentiwords_id is dominated by the positive and neutral classes. The KNN performance evaluation shows that rating-based labelling achieved the highest accuracy (72%), excelling on the positive class (86% recall) but performing poorly on the neutral class (9% recall). Conversely, the lexicon-based labelling methods (both 69% accuracy) have specific strengths: InSet in negative detection (81% recall) and Sentiwords_id in recognizing the neutral class (83% recall). Main challenges of this study include the lexicon's limitations in handling slang and game-specific terms, as well as the inconsistency between ratings and text. This study is expected to provide empirical evidence on performance trade-offs among automatic labelling methods to aid in identifying player satisfaction and advancing the quality of game development.