Where classic approaches still win
Translational models such as TransE-style families remain useful for lightweight link prediction, especially when interpretability and low compute cost are priorities.
What has changed recently
Transformer-based encoders and graph-aware contrastive training have improved representation quality for sparse and heterogeneous enterprise graphs.
Notable trends include:
- stronger handling of relation composition
- better zero-shot behavior on long-tail entity types
- improved transfer across domains with shared schema fragments
Practical model selection
Choose based on workload:
- online retrieval + ranking: optimize latency and memory footprint
- offline analytics: prioritize expressiveness and calibration
- safety-critical domains: emphasize explainability and provenance hooks
Recommendation for teams
Start with a baseline embedding model and benchmark on your own task set. The best architecture is usually determined by graph quality and query patterns rather than benchmark leaderboards.