Dynamic Exploration Graph speeds up nearest neighbor search on evolving data

Approximate Nearest Neighbor Search (ANNS) underpins image search, recommendation systems, and other large-scale retrieval pipelines. Graph-based algorithms have become the go-to for balancing precision and latency, but they assume static data—real-world datasets continuously add and remove items, which degrades the index structure. The Dynamic Exploration Graph (DEG), developed by Nico Hezel, Kai Uwe Barthel, and colleagues, directly tackles this. DEG extends their earlier Exploration Graph with two key innovations: a vertex deletion algorithm that provably preserves graph connectivity, and a data distribution-agnostic expansion method. This keeps the graph balanced and well-connected, even under constant insertions and deletions, avoiding the index rebuild or performance collapse typical of static methods.

The authors evaluated DEG in both streaming and online scenarios, showing it outperforms existing dynamic graph-based ANNS approaches in construction time and search efficiency. Impressively, DEG matches state-of-the-art static algorithms, meaning it offers broad applicability without sacrificing performance on non-evolving datasets. The work will appear in the Proceedings of MMM 2025 and is available on arXiv with DOI 10.48550/arXiv.2607.27640. For professionals managing real-time multimedia databases or recommendation engines, DEG suggests a future where indexes stay fresh without periodic reconstruction—reducing operational overhead and improving query latency as data churns continuously.