Comparing the Behaviour of Ensemble Algorithms for Route Optimization in Last-Mile Deilivery Considering the Weather Condition and Holiday Effect

Abstract

Delivery Time Prediction (DTP) is a crucial factor in last mile logistics. A variety of studies were conducted under this domain using statistics, machine learning and deep learning approaches. The main intention of these kinds of systems is to measure the accuracy and the computational time. However, these improvements come at the cost of significantly increased implementation and operation expenses which are not affordable for small and medium scale businesses. Moreover, DTP considering dynamic factors such as weather, traffic conditions and holidays remains a challenge. Considering the above factors, this paper proposes a novel method of DTP based on the origin, destination geographical points (OD DTP) which is fitting for short distances. According to the case study analysis conducted on the delivery time of New York yellow cab data set using Light Gradient Boosting (LGB), Extreme Gradient Boosting (XGB), Cat Boosting Algorithms and Random Forest (RF), proved that Booting algorithms are much more capable of building DTP model with the exogenous factors such as weather conditions and holiday effect. The feature importance data explained that temperature, humidity, and wind directions are the most important factors within other selected climate criteria). Overall, the trip distance and the trip direction are the most important features when predicting short distance delivery time. The detailed analysis of the selected algorithm behavior concludes that, in terms of evaluation criteria (computational time, overfitting, accuracy, feature importance) LGB is good for model training which has short iteration rounds with small data sets, and the XGB is good for more complex predicting model which deal with large and complex data.