The present invention relates to a technique in which demand prediction of resources of virtual network functions (VNFs) that provide a core technology in a network virtualization environment is performed using machine learning technology. In the present invention, in order to predict VNF resource information, not only are the resources of the VNFs as data but also information of surrounding VNFs that are directly or indirectly related are used, and prediction is possible even in a dynamically changed network environment. In addition, service function chain (SFC) data among various pieces of network information is used to reduce a time required for machine learning according to a size of an entire network.