Hotel Recommendation System in Yogyakarta based on Content-Based Filtering with FastText Modeling

Abstract

The rapid growth of tourism in Yogyakarta has resulted in a wide range of accommodation options, making it challenging for tourists to find hotels that meet their needs. Recommendation systems based on Term Frequency–Inverse Document Frequency (TF-IDF) have limitations in handling variations in word forms and informal language because their matching process relies on lexical representations. This study aims to develop a hotel recommendation system using Content-Based Filtering integrated with FastText word embeddings and to compare its performance with TF-IDF in handling variations between formal and informal language. The study uses a dataset of 100 hotels in Yogyakarta collected through web scraping and data validation. The evaluation was conducted using five pairs of formal and informal queries with equivalent search intents. For each query and method, the top five recommendations were used to calculate Mean Similarity@5 and measure score changes from formal to informal queries. The experimental results show that changes in Mean Similarity@5 for FastText tended to be smaller than those for TF-IDF across the five query pairs tested. The largest change observed for TF-IDF was 26.09%, compared with 5.41% for FastText. These findings indicate that FastText is relatively more robust to variations between formal and informal language under the evaluation scenarios considered in this study. However, the study does not use independently established ground-truth relevance labels; therefore, the evaluation results are not used to claim differences in recommendation accuracy. The analysis also reveals limitations in handling numerical attributes, such as price, and certain syntactic relationships, such as negation.