A PaLM-Embedding and SLIM-Driven Model for High-Precision Tourist Behaviour and Demand Prediction

Abstract

The rapid digitization of the travel industry has generated vast repositories of unstructured customer reviews, yet extracting actionable intelligence for demand forecasting remains a complex challenge. Historically, tourism management relied on basic statistical models and time-series analysis to predict visitor influx. However, these traditional systems frequently fail to account for the nuanced sentiment and Behavioral shifts reflected in modern digital footprints. The core problem lies in the high dimensionality of natural language and the inherent class imbalance found in tourism datasets, where certain customer segments or extreme ratings are underrepresented. Conventional machine learning approaches, such as Linear Discriminant Analysis (LDA) and Quadratic Discriminant Analysis (QDA), often lack the architectural depth to process semantic context or handle skewed data distributions effectively, leading to suboptimal prediction accuracy. To address these limitations, there is a critical need for a framework that fuses deep linguistic understanding with robust statistical rebalancing. This research proposes GPS-Tourism, a hybrid high-precision model integrating Google Pathways Language Model (PaLM) embeddings, the Synthetic Minority Oversampling Technique (SMOTE), and a Sparse Linear Integer Model (SLIM) Classifier. In the proposed system, PaLM converts raw reviews into dense 768-dimensional semantic vectors, while SMOTE synthetically balances the training manifold to ensure minority behavioral patterns are not ignored. Finally, the SLIM architecture is implemented as an ensemble of oblique trees that executes the final classification of tourist ratings and demand segments. Experimental results demonstrate that this fusion significantly outperforms QDA, LDA, and Histogram Gradient Boosting (HGB) models. The significance of this research lies in its ability to provide tourism stakeholders with a granular, high-precision tool for resource allocation and personalized marketing, ultimately bridging the gap between qualitative sentiment and quantitative demand forecasting.