Keywords:Variational Autoencoder, Unsupervised Learning, Clustering
For E-Commerce(EC) business, highly accurate demand-forecasting is important. However, the trends of receiving orders of items vary by the attributes of them, so it is necessary to forecast by the suitable models for each trends. The purpose of this paper is to cluster the items by the suitable representation for trends of receiving orders for realizing highly accurate demand-forecasting. Using real data in EC business, we evaluate the clustering based on the annual trends of receiving orders. We employ dimensional reduction by Variational Autoencoder(VAE) and clustering in latent space by Gaussian Mixture Model(GMM). And we introduce degree of forecasting difficulty as an new index. By our experiments, we confirm that the result of clustering is valid with degree of forecasting difficulty.