Prediction of Store Demands by Decision Trees and Recurrent Neural Networks Ensemble with Transfer Learning

Abstract

Simple vehicle routing problem (VRP) algorithms today achieve near-optimal solution and solve problems with a large number of nodes. Recently, these algorithms are upgraded with additional constraints to respect an increasing number of real-world conditions and, further on, adding a predictive character to the optimization. A distinctive contribution lies in taking into account the predictions of orders that are yet to occur. Such problems fall under time series approaches that are most often obtained using statistical methods or historical data heuristics. Machine learning methods have proven to be superior to statistical methods in most of the literature. In this paper, machine learning techniques for predicting the mass of total daily orders for individual stores are further elaborated and tested on historical data of a local retail company. Among the tested methods are Gradient Boosting Decision Tree methods (XGBoost and LightGBM) and methods of Recurrent Neural Networks (LSTM, GRU and their variations using transfer learning). Finally, an ensemble of these methods is performed, which provides the highest prediction accuracy. The final models use the information on historical order quantities and time-related slack variables.