DEVELOPMENT OF A BACKORDER INVENTORY PREDICTIVE MODEL USING RECURRENT NEURAL NETWORK

Abstract

Inventory management is one of the important business processes which ensure that the supply of raw materials and finished goods remain continuous throughout the business operations. Increasing demand of products is a common cause of out of product inventory, and the adoption of backordering to satisfy outstanding customer orders after its occurrence cannot be undermined. However, wrong management of backorders incurs several issues such as delay in product delivery, low customer satisfaction, and many more. Therefore, it is necessary to ascertain products with high tendencies of shortage beforehand in order to undertake proactive measures and potentially mitigate both tangible and intangible costs. Hence, this research work proposes a backorder predictive model using recurrent neural network (RNN) on large and imbalanced inventory dataset. The research work involve data collection, conversion, normalization and classification. Inventory data of product information from Kaggle repository was used for this study. The data was preprocessed using Min-Max Scaler, while three data balancing methods - Synthetic Minority Oversampling Technique (SMOTE), Adaptive Synthetic (ADASYN) and Random Under Sampling (RUS) were applied on the imbalanced data one after the other and their output were fed into RNN to predict which item goes on backorder. The evaluation of the result obtained showed ADASYN+ RNN had performed better with 0.901 precision, 0.879 recall, and 0.889 F1-Score. The proposed model when compared with other machine learning algorithms shows significant impact on prediction of product backorder.