Big data transfer in next-generation scientific applications is now commonly carried out over connections with guaranteed bandwidth provisioned in High-performance Networks (HPNs) through advance bandwidth reservation. To use HPN resources efficiently, provisioning agents need to carefully schedule data transfer requests and allocate appropriate bandwidths. Such reserved bandwidths, if not fully utilized by the requesting user, could be simply wasted or cause extra overhead and complexity in management due to exclusive access. This calls for the capability of performance prediction to reserve bandwidth resources that match actual needs. Towards this goal, we employ machine learning algorithms to predict big data transfer performance based on extensive performance measurements, which are collected over a span of several years from a large number of data transfer tests using different protocols and toolkits between various end sites on several real-life physical or emulated HPN testbeds. We first identify a comprehensive list of attributes involved in a typical big data transfer process, including end host system configurations, network connection properties, and control parameters of data transfer methods. We then conduct an in-depth exploratory analysis of their impacts on application-level throughput, which provides insights into big data transfer performance and motivates the use of machine learning. We also investigate the applicability of machine learning algorithms and derive their general performance bounds for performance prediction of big data transfer in HPNs. Experimental results show that, with appropriate data preprocessing, the proposed machine learning-based approach achieves 95% or higher prediction accuracy in up to 90% of the cases with very noisy real-life performance measurements.