This paper presents a comprehensive review of machine learning (ML) in optical fiber communications, particularly in channel modeling. It discusses the evolution from conventional methods to ML-based approaches that aim to enhance predictive and computational efficiency. Specifically, the discussions categorize ML methodologies into data-driven and principle-driven approaches. The former treats channel modeling as a “black box” providing rapid modeling capabilities at the expense of transparency and substantial data requirements. In contrast, the latter integrate physical principles into the ML-based system, enhancing model interpretability and reducing data dependency. In addition, the emergence of hybrid models that combine the strengths of both approaches is explored. This classification provides a structured overview of how ML is reshaping channel modeling in optical fiber communications, underscoring its potential to improve system design and exploring advanced nonlinear dynamics in optical fiber communication systems.