Query optimization is a main component of graph databases and triple stores that host knowledge graphs (KGs). Traditional symbolic optimizers rely on heuristics and cost models that are prone to inaccuracies, leading to suboptimal execution plans. Recent advances in machine learning (ML) provide promising solutions to address these limitations by learning from data and queries to enhance cardinality estimation, cost prediction, and plan enumeration. This work surveys the emerging landscape of ML-based query optimization over KGs, including learned, neuro-symbolic, and (fully) neural approaches. We discuss the architecture and trade-offs of these systems, present preliminary results, and highlight open challenges for future research.