[Paper Review] Geometric Near-neighbor Access Tree (GNAT) revisited

Kimmo Fredriksson|arXiv (Cornell University)|May 19, 2016

Data Management and Algorithms11 references3 citations

TL;DR

This paper revisits the Geometric Near-neighbor Access Tree (GNAT), proposing memory-efficient enhancements through adaptive hyperplane and ball partitioning, fixed-point range table compression, and dynamic tree construction. The result is reduced memory usage and improved search performance, with GNATTY (GNAT with ball partitioning and optimization) achieving up to 10x faster queries than baseline GNAT while using less than 1/4 the memory in some cases.

ABSTRACT

Geometric Near-neighbor Access Tree (GNAT) is a metric space indexing method based on hierarchical hyperplane partitioning of the space. While GNAT is very efficient in proximity searching, it has a bad reputation of being a memory hog. We show that this is partially based on too coarse analysis, and that the memory requirements can be lowered while at the same time improving the search efficiency. We also show how to make GNAT memory adaptive in a smooth way, and that the hyperplane partitioning can be replaced with ball partitioning, which can further improve the search performance. We conclude with experimental results showing the new methods can give significant performance boost.

Motivation & Objective

• To address the long-standing criticism that GNAT has excessive memory usage, which stems from overly coarse analysis.
• To improve GNAT's search efficiency while simultaneously reducing its memory footprint through novel partitioning and data representation techniques.
• To introduce memory adaptivity in GNAT, allowing dynamic adjustment of parameters like arity and pruning thresholds based on available memory.
• To demonstrate that ball partitioning can outperform traditional hyperplane partitioning in terms of query performance, especially for string data.
• To enable practical deployment of GNAT in external memory and dynamic settings by optimizing construction and insertion costs.

Proposed method

• Replacing hyperplane partitioning with ball partitioning to improve flexibility and query performance, especially in high-dimensional or non-Euclidean spaces.
• Introducing a memory-adaptive parameter α that controls pruning thresholds, allowing trade-offs between memory usage and distance evaluations.
• Using fixed-point (1-byte) representation for range table entries instead of 32-bit floats, reducing memory by 75% with minimal performance loss.
• Applying lazy construction: range tables and subtrees are built on-demand during search, amortizing construction cost.
• Enhancing EGNAT with nearest-neighbor indexing over pivots to accelerate query processing in high-arity settings.
• Proposing bulk-loading with on-demand node creation and dynamic promotion of leaf buckets to internal nodes to reduce initial construction overhead.

Experimental results

Research questions

• RQ1 Can the memory usage of GNAT be significantly reduced without sacrificing search efficiency?
• RQ2 Does replacing hyperplane partitioning with ball partitioning lead to measurable performance improvements in proximity search?
• RQ3 Can fixed-point compression of range tables maintain high query accuracy while reducing memory consumption by up to 75%?
• RQ4 How does adaptive memory control (via parameter α) affect the trade-off between memory usage and distance evaluations?
• RQ5 Can lazy construction and dynamic node promotion improve the practicality of GNAT in external and dynamic memory settings?

Key findings

• GNATTY with ball partitioning and α=0.5 achieves up to 10x faster query performance than baseline GNAT for the same memory budget.
• Using 1-byte fixed-point range tables (GNATTY-FP) reduces memory usage by 75% compared to 32-bit floats, with only a minor drop in performance.
• For the same performance level, GNAT with hyperplane partitioning requires m≈135 for color histograms and m≈1120 for strings, while GNATTY with α=0.5 or 0.6 achieves comparable results with much lower memory.
• Ball partitioning outperforms hyperplane partitioning in both vector and string datasets, especially at higher search radii and lower arities.
• Lazy construction and on-demand range table computation reduce initial construction cost and amortize overhead, improving scalability for large datasets.
• The proposed techniques are directly applicable to EGNAT, improving its construction and insertion costs while maintaining high query efficiency.