Tourism Segmentation Using Attention and Anomaly Detection

Abstract

The RFM (Recency, Frequency, Monetary) model is used for customer segmentation based on purchasing habits. This study aims to segment tourists with an attention-based neural classifier and to quantify how outlier removal affects performance. Using preprocessed License Plate Recognition data, we derived RFM features to identify distinct patterns. A neural network with multi-attention mechanisms was trained on this data. We evaluated three outlier detection techniques—Interquartile Range, Z-score, and Isolation Forest—alongside two representative point selection algorithms: Local Density Instance Selection and k-medoids. Our results highlight Z-score as the most efficient outlier removal method, achieving a macro-averaged F1-score that is 0.05 higher and a 23.2% faster processing time compared to the other methods. Additionally, Isolation Forest excelled in identifying strategic visitors. These results could guide policymakers in personalized campaigns to promote loyalty. Future work will integrate fuzzy logic and explainable AI, incorporate individual-level data, and validate across diverse destinations.