Network traffic classification has been extensively used in Quality-of-Service (QoS) control, intrusion detection, and other areas of network communications and cybersecurity. To classify traffic, Neural Networks (NN) have been adopted and achieved promising performances. There are two major approaches in the NN-based traffic classification, i.e., using raw traffic data and using flow features. This paper first proposed a novel searching algorithm to find the optimal hyperparameters of the NN with the consideration of the characteristics of the network traffic data. With the optimized NNs, a comprehensive comparison was conducted between the raw-date-based and the feature-based traffic classification. The experimental results showed that the former achieved higher accuracy and precision, which means that even with partial information and less data pre-processing, NN performs more effectively and efficiently in extracting features for traffic classification.