Wisata pantai di Pulau Madura memiliki potensi besar, namun tingginya volume ulasan digital berbasis teks bebas dari wisatawan menyulitkan identifikasi topik pembahasan dan keluhan secara manual. Penelitian ini bertujuan untuk memodelkan topik ulasan wisata pantai Madura menggunakan algoritma Latent Dirichlet Allocation (LDA), menganalisis pengaruh variasi preprocessing terhadap kualitas model, serta menentukan variasi preprocessing yang paling optimal berdasarkan nilai Coherence score. Dataset yang digunakan berupa 3.001 ulasan Google Maps hasil proses filtering dari 3.953 ulasan awal. Tahapan preprocessing yang diterapkan meliputi cleaning, tokenizing, stopword removal, stemming, dan bigram, yang kemudian dikombinasikan menjadi empat variasi preprocessing (V1–V4). Selanjutnya, data direpresentasikan menggunakan Bag of Words (BoW) dan dimodelkan menggunakan LDA dengan variasi jumlah topik K 2-10. Kualitas model dievaluasi menggunakan Coherence Score dan Perplexity, kemudian divisualisasikan menggunakan WordCloud dan pyLDAvis untuk membantu interpretasi topik. Hasil penelitian menunjukkan bahwa variasi preprocessing memberikan pengaruh terhadap kualitas model LDA. Variasi terbaik diperoleh pada V3, yaitu kombinasi Cleaning, Tokenizing, Stopword Removal, dan Stemming, dengan jumlah topik K 2, menghasilkan Coherence Score sebesar 0,5443 dan Perplexity sebesar -5,5581. Model tersebut menghasilkan dua topik utama yang menggambarkan keindahan alam dan kenyamanan lingkungan wisata, serta biaya, fasilitas, dan layanan pendukung wisata. Dengan demikian, dapat disimpulkan bahwa kombinasi Stopword Removal dan Stemming mampu menghasilkan representasi dokumen yang lebih baik dibandingkan variasi preprocessing lainnya, sedangkan penambahan bigram pada penelitian ini belum mampu meningkatkan kualitas model topic modeling menggunakan metode LDA.


Deskripsi Lain

Beach tourism on Madura Island has significant potential; however, the large volume of unstructured digital reviews submitted by tourists makes it difficult to identify discussion topics and complaints manually. This study aims to perform topic modeling on Madura beach tourism reviews using the Latent Dirichlet Allocation (LDA) algorithm, analyze the effect of preprocessing variations on model quality, and determine the most optimal preprocessing variation based on the Coherence Score. The dataset consisted of 3,001 Google Maps reviews, obtained after a filtering process from an initial 3,953 reviews. The preprocessing stages included cleaning, tokenizing, stopword removal, stemming, and bigram generation, which were combined into four preprocessing variations (V1–V4). The processed data were then represented using the Bag of Words (BoW) model and analyzed using LDA with the number of topics ranging from K 2 to 10. Model performance was evaluated using Coherence Score and Perplexity, while WordCloud and pyLDAvis were employed to facilitate topic interpretation. The results indicate that preprocessing variations significantly affect the quality of the LDA model. The best performance was achieved by Variation 3 (V3), which consists of Cleaning, Tokenizing, Stopword Removal, and Stemming, with K 2, producing a Coherence Score of 0.5443 and a Perplexity of -5.5581. This model generated two main topics representing the natural beauty and environmental comfort of beach tourism, well tourism costs, facilities, and supporting services. Therefore, it can be concluded that the combination of Stopword Removal and Stemming provides a more effective document representation than the other preprocessing variations, while the addition of bigram in this study did not improve the performance of the LDA-based topic modeling.