TravelVibe: Aspect-Based Sentiment Analysis of Hotel Booking Application Reviews Using Support Vector Machines

Abstract

Online reviews strongly influence travel decisions, yet their volume and informal language make manual interpretation impractical. This paper presents TravelVibe, a web-based aspect-based sentiment analysis system for reviews of Tripadvisor, Booking.com, and Agoda mobile applications. A total of 30,000 English-language Google Play Store reviews posted from 2022 to 2025 were collected. Reviews were cleaned, spell-corrected, lemmatised, and represented using unigram and bigram term frequency-inverse document frequency features. Non-negative matrix factorisation supported the derivation of four aspect categories—Customer Support, Hotel Experience, Price and Deals, and Travel Planning—whereas VADER compound scores supplied initial positive, neutral, and negative sentiment labels. Two calibrated linear support vector machine classifiers were trained for sentiment and aspect prediction. The Synthetic Minority Over-sampling Technique was applied only to the training data to address unequal class distributions. The baseline sentiment classifier achieved 80.99% accuracy; after balancing, accuracy increased to 91.42%, neutral-class recall rose from 0.50 to 0.98, and neutral F1-score rose from 0.58 to 0.95. Aspect accuracy increased from 88.45% to 88.94%, with improved balance for underrepresented aspects. Fourteen functional test cases passed, covering review analysis, filters, platform dashboards, and downloadable visualisations. The results indicate that TF-IDF and linear SVM remain effective for interpretable review analytics, although the use of automatically generated labels and semantically overlapping aspects limits the strength of external validity claims