Over the past decade, network traffic classification has undertaken substantial advancements, evolving from traditional flow-based and port-based methods to more sophisticated approaches. Early techniques relied on port numbers and protocol signatures, which became ineffective with the increasing prevalence of dynamic port usage. This led to the adoption of statistical methods that leveraged flow-level features such as packet size, flow duration, and inter-arrival times, improving classification accuracy. Machine learning (ML) techniques further enhanced the capabilities of these statistical methods by automating feature selection and enabling the recognition of complex traffic patterns. Despite this, challenges remain in achieving real-time classification, handling encrypted traffic, and ensuring scalability across diverse network environments. In this research, we recommend the use of Expectation-Maximization (EM) clustering for network traffic classification, which offers a robust approach to handling uncertainty in cluster assignments and effectively identifying complex traffic patterns with an accuracy of around 94%. This paper provides a comprehensive survey of network traffic classification methods, tracing their evolution from conventional techniques to modern approaches, with a focus on the importance of employing statistical approaches augmented by machine learning algorithms, and highlights the potential of EM clustering for improving classification performance. Additionally, it discusses the challenges and explores future directions for robust and efficient traffic classification in contemporary networks.