Currently, a lot of new types of applications are constantly emerging, and the user communication demands for different applications are also becoming diversified and personalized. To match users' frequent and changing communication demands, internet service provider (ISP) usually constantly purchases and operates new specialized network equipment, which leads to high operating cost and resource waste, and it is obviously unsustainable for network construction and development. This paper addresses the above challenge from the perspective of software-based method by reusing diverse routing functions. The suitable routing functions are selected to compose the customized routing services on communication paths of applications, in order to satisfy the user demands. Based on network function virtualization (NFV) and software defined networking (SDN), the paper proposes an adaptive routing service composition mechanism. It leverage software product line (SPL) to establish routing service product line, which serves as the basis to select routing functions and optimize routing services. In addition, based on machine learning, it establishes two-phased routing service learning model, that is, offline mode and online mode, by leveraging multilayer feed-forward neural network. It can constantly adjust and optimize routing function selection and service composition to achieve routing service customization and improve user service experience. Simulation and performance results show that the proposed model is feasible and efficient.