The machine learning algorithm is gaining prominence in traffic identification research as it offers a way to overcome the shortcomings of port-based and deep packet inspection, especially for P2P-based Skype. However,recent studies have focused mainly on traffic identification based on a full-packet dataset, which poses great challenges to identifying online network traffic. This study aims to provide a new flow identification algorithm by taking the sampled flow records as the object. The study constructs flow records from a Skype set as the dataset, considers the inherent NETFLOW and extended flow metrics as features, and uses a fast correlation-based filter algorithm to select highly correlated features. The study also proposes a new NFI method that adopts a Bayesian updating mechanism to improve the classifier model. The experimental results show that the proposed scheme can achieve much better identification performance than existing state-of-the-art traffic identification methods, and a typical feature metric is analyzed in the sampling environment. The NFI method improves identification accuracy and reduces false positives and false negatives compared to other methods.