An Interpretable Semantic Intelligence Framework for Concurrent Tourist Behaviour Modeling and Tourism Demand Forecasting

Abstract

The continuous expansion of digital tourism platforms has produced an extensive collection of online customer feedback, creating new opportunities to understand traveller preferences and anticipate tourism demand. Despite this abundance of information, accurately interpreting unstructured textual reviews remains a significant challenge. Traditional tourism forecasting methods mainly rely on historical trends and numerical data, making them less effective in capturing the underlying opinions, emotions, and behavioural patterns expressed by travellers. Furthermore, tourism datasets often suffer from imbalanced class distributions, where certain visitor categories and rating levels have considerably fewer samples. Conventional classification techniques, such as Linear Discriminant Analysis (LDA), Quadratic Discriminant Analysis (QDA), and Histogram Gradient Boosting (HGB), are limited in their ability to extract contextual semantic information and perform consistently on skewed datasets. To address these challenges, this research presents GPS-Tourism, an intelligent prediction framework that integrates Google Pathways Language Model (PaLM) embeddings, the Synthetic Minority Oversampling Technique (SMOTE), and a Sparse Linear Integer Model (SLIM) Classifier. Initially, PaLM converts textual reviews into rich 768-dimensional semantic representations that preserve contextual meaning. SMOTE subsequently balances the training data by generating synthetic minority samples, enabling improved learning across underrepresented classes. Finally, the SLIM-based ensemble of oblique decision trees performs the classification of tourist ratings and demand categories. Experimental findings demonstrate that the proposed framework achieves higher predictive accuracy than LDA, QDA, and HGB, providing tourism stakeholders with a reliable decisionsupport system for demand forecasting, strategic resource planning, and personalized marketing initiatives.