Efficient and accurate prediction of last-mile delivery time is an essential component of modern logistics systems, especially with the rapid growth of electronic commerce during the COVID-19 pandemic. This study proposes a novel approach to address the problem of delivery time prediction in complex environments where geolocation information such as latitude and longitude is fixed to a specific spot, such as collective facilities (e.g., apartment complexes, laboratories, etc.), by quantifying the similarity between addresses using the Levenshtein Distance. By incorporating similarity scores as a feature of the prediction model, we address the problem of ambiguous delivery points or shared addresses within collective facilities and improve prediction accuracy. The approach in this study transforms the delivery time prediction problem into a multi-class classification problem to predict delivery times within 24 predefined time bins. This classification approach solves the averaging problem encountered in regression-based models and enhances practicality by providing top-3 probabilistic delivery times. In our experiments, the ResNet-50 and XGBoost models performed the best, with ResNet-50 achieving top-1 accuracy of 0.619 and top-3 accuracy of 0.830. Furthermore, when including the distance-based similarity score of Levenshtein, the accuracy of top-1 increased by 5. 25% on average, demonstrating the importance of this feature. Efficient and accurate prediction of last-mile delivery time is an essential component of modern logistics systems, especially with the rapid growth of electronic commerce during the COVID-19 pandemic. This study proposes a novel approach for delivery time prediction in complex environments where geolocation information such as latitude and longitude is fixed to a specific spot, such as collective facilities (e.g., apartment complexes, laboratories, etc.), by quantifying the similarity between addresses using the Levenshtein distance. By incorporating similarity scores as a feature of the prediction model, we cope with the problem of ambiguous delivery points or shared addresses within collective facilities so as to improve prediction accuracy. The approach in this study transforms the delivery time prediction problem into a multi-class classification problem to predict delivery times within 24 predefined time bins. This classification approach solves the averaging problem encountered in regression-based models and enhances practicality by providing top-3 probabilistic delivery times. In our experiments, the ResNet-50 and XGBoost models performed the best, with ResNet-50 achieving top-1 accuracy of 0.619 and top-3 accuracy of 0.830. Furthermore, when including the Levenshtein distance-based similarity score, the top-1 accuracy increased by 5.25% on average, thereby demonstrating this feature’s importance.