Improving unloading time prediction through driver and customer segmentation

Abstract

Modern-day society is driven by transportation networks. Now it is easier than ever to order your daily necessities through online platforms. The research interest in this thesis focus on the delivery aspect lies with ordering. A successfully completed delivery means a properly addressed vehicle routing problem. The data set that is involved in the study refers to a large amount of perishable good cases that are delivered through large trucks. Each truck caters to 8-10 customers in a day. Since these are perishable goods delivered to people in the foodservice industry, they expect a sound ETA of their delivery to plan ahead for meal preparations.

To provide an ETA in a multi-stop route there are two variables to be solved. One is the travel time between stops, which modern-day map services would output without a hassle. However, the next important thing is the unloading time needs to calculate with the historical data. The study suggests a way to involve customer profiling and driver profiling so that unloading time prediction can be done with those two variables along with the delivery volume of the stop.

Modeling these two variables into a regression model was a challenge on its own due to their large dimension of them. Segmentation of the said variables and using segment mean yielded better results in regression compared to using a label encoding technique blindly which introduced an orderly nature to features from the id itself. Furthermore, once segment means were clustered based on their distribution and provided a cluster identifier that justifies the orderly nature, models were able to yield their least MSE.

Finally, this study highlights the importance involving of the customer site and the driver's experience in the unloading time. Also, this study has presented a way of representing such variables with a high cardinality in a meaningful manner so that model can be built with less error. This will provide a good starting point for further analysis on similar research interests in the future