Is it a work or leisure travel? Applying text classification to identify work-related travel on social networks

TL;DR

The paper addresses identifying whether a trip is leisure or work-related to enhance travel recommendations. It investigates transformer-based automatic text classification (BERT, RoBERTa, BART) applied to TripAdvisor reviews, using a binary leisure/work label and careful handling of data imbalance. Across a 5-fold cross-validation setup, RoBERTa yields the strongest Macro-F1, with all models showing strong Micro-F1 scores and no statistically significant differences, suggesting transformer-based ATC is effective for inferring travel purpose from textual data. The work highlights potential improvements in POI recommendations by incorporating travel purpose, while acknowledging limitations of reliance on review text alone and suggesting future feature integration.

Abstract

In today's digital era, the use of Social Networks (SNs) and Location-Based SNs (LBSNs) has become integral for travelers seeking Points of Interest (POI) and sharing travel experiences. This trend is supported by the fact that a significant majority of American travelers utilize SNs during their trips. However, the abundance of information available on these platforms presents a challenge in identifying the best options. To address this issue, Recommender Systems (RS) are commonly employed to suggest POIs based on user history, with the integration of contextual information enhancing the quality of recommendations. Notably, incorporating user travel purpose, which is often overlooked but holds potential in characterizing travelers' behavior, can lead to more tailored recommendations. In this study, we propose a model to predict whether a trip is leisure or work-related, utilizing state-of-the-art Automatic Text Classification (ATC) models such as BERT, RoBERTa, and BART to enhance the understanding of user travel purposes and improve recommendation accuracy in specific travel scenarios.