A Model for Multi-relational Data Mining on Demand Forecasting

Accurate demand forecasting remains difficult and challenging in today’s competitive and dynamic business environment, but even a little improvement in demand prediction may result in significant saving for retailers and manufactures. This paper aims to improve the accuracy of demand forecasting by implementing multirelational data mining process on the large data sets of stores, products, and shoppers. Most existing data mining approaches look for patterns in a single relation of a database. Multi-relational data mining can analyze data from multiple relations directly without the need to transfer the data into a single relation first. Two data mining models are proposed in this paper, which are Pure Classification (PC) model and Hybrid Clustering Classification (HCC) model. Pure Classification model uses k-Nearest Neighbor Classification technique, and Hybrid Clustering Classification first uses k-Mean Mode Clustering to define clusters and then uses k-Nearest Neighbor classification to find k most similar objects. Hybrid Clustering Classification model introduces a concept of combining existing data mining techniques on the multi-relational data sets. Experimental results show that Hybrid Clustering Classification is particularly promising for demand forecasting.