Travel DNA: Persona-Driven Destination Prediction with Sentiment-Enhanced Hybrid Modeling

Abstract

This study develops a sentiment analysis-based personalized travel destination prediction system using 12.57 million user reviews from the online accommodation platform Airbnb. The proposed system operates on a dataset covering 20 European cities through a five-stage pipeline: (i) data extraction and preprocessing, (ii) sentiment analysis with the multilingual transformer-based XLM-RoBERTa language model, (iii) traveler profiling with the K-Means clustering algorithm, (iv) trajectory mining with Markov chains, and (v) a three-signal hybrid prediction engine. The proposed hybrid architecture combines three signals: a Markov signal that models inter-city transition probabilities, a history signal reflecting users’ revisit tendency, and a sentiment-city affinity signal that matches users’ sentiment profiles with city characteristics. These signals are combined with weights optimized per traveler persona. In experiments conducted on 10,436 users (20.9% coverage rate) who reviewed during the test period and had travel history in the training period, the correct destination was found in the first recommendation (Hit@1) at 24.42%, in the top three (Hit@3) at 40.27%, in the top five (Hit@5) at 50.80%, and the Mean Reciprocal Rank (MRR) measuring ranking quality was 33.56%. Ablation study measuring each component’s contribution reveals that the hybrid model outperforms all individual signals across every metric and that the proposed sentiment-city affinity signal achieves a 76% relative improvement in Hit@1 over the popularity-based baseline.