Abstract

Purpose: This study developed a deep learning-based mobile travel recommendation system that provides recommendations for local tourist destinations based on users' favorite travel photos. To provide recommendations, use cosine similarity to measure the similarity score between a person's image and a tourism destination gallery through the tag label vector. Label tags are inferred using an image classifier model run from a mobile user device via Tensorflow Lite. There are 40 tag labels that refer to categories, activities and objects of local tourism destinations. Methods: The model is trained using state-of-the-art mobile deep learning architecture EfficientNet-Lite, which is new in the domain of tourism recommender system. Result: This research has conducted several experiments and obtained an average model accuracy of more than 85%, using EfficientNet-Lite as its basic architecture. The implementation of the system as an Android application is proven to provide excellent recommendations with a Mean Absolute Percentage Error (MAPE) of 5.2%. Novelty: A tourism recommendation system is a crucial solution to help tourists discover more diverse tourism destinations. A content-based approach in a recommender system can be an effective way of recommending items because it looks at the user's preference histories. For a cold-start problem in the tourism domain, where rating data or past access may not be found, we can treat the user's past-travel-photos as the histories data. Besides, the use of photos as an input makes the user experience seamless and more effortless.