Aspect-Based Sentiment Classification using Unsupervised Learning Approach for Malaysia Restaurant Review

The increasing reliance on online restaurant reviews for decision-making highlights the need for accurate and detailed sentiment analysis. However, most existing systems focus only on overall polarity and rely heavily on labeled datasets, limiting their applicability in real-world, lowresource environments. To address this gap, this paper proposes an unsupervised framework for aspect-based sentiment classification of Malaysian restaurant reviews. The methodology combines K-means clustering for aspect grouping, Latent Dirichlet Allocation (LDA) for topic modeling, and the VADER sentiment analyzer for polarity detection. Using a dataset of 3,569 English-language reviews from Google and TripAdvisor, the system identifies meaningful review aspects and assigns sentiment scores without requiring prior annotation. K-means achieved a silhouette score of $\mathbf{0. 3 6 9}$, while LDA topic coherence scores ranged from 0.293 to 0.396. VADER sentiment classification yielded an accuracy of $\mathbf{8 0 \%}$, precision of 0.74, recall of 0.80, and $\mathbf{F 1}$-score of 0.76. An interactive dashboard was also developed to visualize sentiment distributions, topic trends, and keyword usage. This study demonstrates that interpretable sentiment analysis can be conducted using unsupervised techniques, especially in settings where labeled data is limited.