TL;DR
The paper introduces a multi-scale metric for strings that combines angle distances between n-gram count vectors at different scales using exponential weights. It provides algorithms, theoretical guarantees, and empirical validation for clustering applications.
Existing string distance metrics like edit distance and n-gram based measures have limitations: edit distance is computationally expensive and may not capture semantic similarity well, while single-scale n-gram approaches ignore structural information at multiple levels. There is a need for a string metric that captures similarity across multiple granularities while remaining computationally efficient and theoretically well-founded.
The authors define a weighted angle distance metric d_ρ that aggregates angle distances (cosine-based) between n-gram count vectors across all scales n, using exponential weights ρ^n to control the contribution of each scale. They provide a linear-time suffix-tree algorithm for efficient computation, establish that the metric satisfies metric properties, prove stability guarantees including robustness to tandem-repeat stutters, and characterize the isometries of the metric space.
What the paper shows.
The metric d_ρ is benchmarked in DBSCAN clustering against edit distance and n-gram baselines. The paper demonstrates that the weighted angle distance achieves competitive or superior clustering performance while maintaining computational efficiency through the linear-time suffix-tree algorithm. Specific numerical results on clustering benchmarks are not detailed in the abstract.
The abstract does not provide specific quantitative results or detailed performance comparisons, making it difficult to assess the practical magnitude of improvements over baselines. The paper's scope appears limited to clustering evaluation; applicability to other string matching tasks is not discussed. The choice of exponential weight parameter ρ and its sensitivity is not addressed in the abstract.
Generated by claude-haiku-4-5 · Apr 25, 2026 · Read the PDF for authoritative content.