Analisis Sentimen Ulasan Pengguna Aplikasi KitaLulus Berdasarkan Skala Rating Menggunakan Algoritma Support Vector Machine (SVM)

Abstract

The KitaLulus job-search application has been downloaded more than 10 million times and holds an average rating of 4.7 on the Google Play Store, yet the large volume of user reviews makes manual interpretation of user perception difficult. This study applies the Support Vector Machine (SVM) algorithm to classify the sentiment of KitaLulus user reviews into five classes based on the 1-5 rating scale. Data were collected through web scraping using the google-play-scraper library, yielding 15,000 reviews that were reduced to 10,536 after data cleaning. The research followed the Sample, Explore, Modify, Model, Assess (SEMMA) framework, comprising text preprocessing (cleaning, case folding, tokenization, slang normalization, stopword removal, and stemming), TF-IDF feature weighting, SVM modeling with 10-fold cross validation, and hyperparameter tuning using GridSearchCV across three kernel types (Linear, Radial Basis Function, and Polynomial). The initial LinearSVC evaluation achieved a mean accuracy of 74.17%. Hyperparameter tuning selected the RBF kernel with C=1 and gamma=scale as the best configuration, which, when evaluated on the full dataset using cross_val_predict, produced an accuracy of 70.10%, with the highest F1-Score on the rating-5 class (0.86) and considerably lower performance on the minority rating-2 class (0.04) due to class imbalance. The study demonstrates that SVM is reasonably effective for multiclass sentiment classification of application reviews, although handling imbalanced data remains an important consideration for future research.