Predictive Machine Learning Framework for Address Error Detection in Logistics System

In today's fast-paced logistics world, an incorrectly entered address can mean the difference between efficient delivery and costly delays. A misplaced typo, omissions, or an outdated postcode - incorrectly entered address data poses a major problem for companies that send thousands of shipments every day. This study aims to develop an intelligent, machine learning-based approach that enables logistics companies to regularly identify and flag suspicious addresses that could potentially cause problems before they arise. Using advanced text analysis, language understanding, and predictive modelling, the approach learns to recognize slightly deviating address patterns - often imperceptible. In this direction, several methods were trained and compared to find the most reliable solution for error detection in millions of delivery records. Specifically, transformer-based sentence embeddings were utilized to semantically encode address fields. Among the evaluated models, GTE-base demonstrated the highest semantic similarity with a cosine score of 0.945, while maintaining a low inference latency of 1.65 seconds. For the classification task, four widely used algorithms - XGBoost, LightGBM, CatBoost, and AdaBoost - were trained under both baseline and hyperparameter-optimized configurations. After optimization, CatBoost achieved the highest classification accuracy, while LightGBM obtained the highest F1 score, reflecting a strong balance between precision and recall. The study shows that the proposed approach can make a valuable contribution to better data quality and fewer delivery errors.