Penerapan Model LSTM dan CNN Untuk Klasifikasi Sentimen Pada Ulasan Aplikasi Roblox

Abstract

In the rapidly evolving digital era, online gaming platforms like Roblox have transformed into complex interactive social spaces where users interact, create, and co-build virtual experiences. This study aims to compare the performance of two deep learning architectures Convolutional Neural Network (CNN) and Long Short-Term Memory (LSTM) in classifying user reviews of the Roblox app from the Google Play Store into positive, negative, and neutral sentiment categories. The dataset comprises 5,000 Indonesian-language reviews collected via web scraping using the google_play_scraper library. Preprocessing involved text cleaning, case folding, tokenization, normalization of informal words, stopword removal, stemming, and lexicon-based sentiment labeling. Data were converted to numerical representations using Tokenizer and padding, then split into 80% training and 20% testing subsets. CNN achieved superior performance with 89% accuracy, 0.88 precision, 0.81 recall, and 0.83 F1-score, outperforming LSTM (86.60% accuracy, 0.80 precision, 0.82 recall, 0.81 F1-score). CNN effectively extracts spatial patterns in text, while LSTM captures temporal word dependencies. This research affirms CNN's superiority for short Indonesian text sentiment analysis, provides a deep learning benchmark for gaming app reviews, and offers practical implications for Roblox developers to interpret user feedback for feature enhancements.