<p>Link prediction is a core task in network science and machine learning. This survey offers an updated synthesis from classical similarity indices to modern graph representation learning, including embedding methods, graph neural networks, and emerging Graph Transformers. We formalize the task and evaluation protocol, compare methods across network settings (static and dynamic graphs, multiplex and heterogeneous networks, knowledge and weighted graphs), and analyze trade-offs in scalability, accuracy, and interpretability. The review distills practical guidance—using calibrated heuristic baselines, leveraging unsupervised embeddings for large graphs, favoring GNNs when attributes and relation types are informative, and considering Transformers for long-range dependencies and multimodal contexts—while highlighting open challenges in temporal generalization, sparsity and cold-start, explainability, and integration with large language models. Together, the survey provides an updated map of methods, benchmarks, and lessons to inform model selection and motivate future research.</p>