Privacy-Enhanced Point-of-Interest Recommendations Using Synthetic Data: A Comparative Study of CTGAN and Clustering Techniques

Recommender systems are designed to provide users with personalized suggestions and promotions by filtering vast amounts of information. However, these systems often face significant privacy and data sparsity challenges. Point-of-interest (POI) recommender systems, which focus on suggesting physical locations to users, are particularly affected by these issues due to location data's sensitive and scattered nature. This paper presents an examination of privacy-preserving approaches through data synthesis techniques, specifically employing Conditional Tabular Generative Adversarial Networks (CTGAN) and clustering. We investigate the performance of recommender systems trained solely with synthetic data generated by CTGAN or clustering for physical users. Our study compares the effectiveness of recommending by using original data, CTGAN-generated data, and clustering-generated centroids. Experimental results on real-world datasets demonstrate that integrating synthetic data can maintain system accuracy while resolving privacy issues such as data transportation and user anonymization, demonstrating the potential of GAN-based methods to advance POI recommendation systems.