Sales Forecast and Design Generation for Textile Products Using Machine Learning

Abstract

In the retail context, incorrect determination of the quanti- ties to be purchased of each item from vendors, either by excess or by default, can result in unnecessary storage costs or lost sales, respectively. Both of these situations should be avoided by firms, hence the need for efficient determination of purchase quantities. In this study, we applied artificial intelligence methods to predict sales based on several features. We developed four models (decision tree, random forest, XGBoost and artificial neural network) based on the sales data of John Walkden & Company, a distribution subsidiary of the VLISCO group, which spe- cializes in the import and exclusive distribution of four (04) brands of textile products: VLISCO, WOODIN, UNIWAX and GTP. It has been shown that the random forest algorithm offers the best performance on these sales data. In addition, we have also addressed the conception of designs by proposing two models (Generative Adversarial Networks and Deep Convolution Generative Adversarial Networks) based on artificial intelligence techniques. These models are able to generate designs from noise in order to automate the creation of these designs which, until now, was done by humans. We trained these two models using a public “Quickdraw” dataset and at the end of our experiments, it appears that the Deep Convolution Generative Adversarial Networks model provides the best performance.