MCGI: Geometry-Aware Graph Indexing for Billion-Scale Vector Search
A new indexing technique, Manifold-Consistent Graph Indexing (MCGI), has been developed by researchers for conducting Approximate Nearest Neighbor (ANN) searches on a billion-scale. This geometry-aware method tackles the issue of Euclidean-Geodesic mismatch in high-dimensional environments by utilizing Local Intrinsic Dimensionality (LID) to adjust search strategies according to the intrinsic geometry of the data. Unlike traditional algorithms that apply a uniform approach to dimensions, MCGI varies its beam search budget through real-time geometric assessments, which decreases reliance on specific hyperparameters. Theoretical evaluations indicate that MCGI maintains strong approximation by upholding manifold-consistent topological connectivity, making it suitable for disk-resident applications on standard hardware.
Key facts
- MCGI stands for Manifold-Consistent Graph Indexing.
- It is a geometry-aware and disk-resident indexing method for ANN search.
- It uses Local Intrinsic Dimensionality (LID) to adapt search strategies.
- It addresses the Euclidean-Geodesic mismatch in high-dimensional spaces.
- MCGI modulates beam search budget based on geometric analysis.
- It reduces sensitivity to data-specific hyperparameters.
- Theoretical analysis confirms robust approximation via manifold-consistent topological connectivity.
- The method targets billion-scale vector search on disk-resident systems.
Entities
—