AI 中文总结
针对图ANN索引在插入/删除变动下召回率下降的问题，提出基于可导航性退化信号触发局部边修复的策略，在固定修复预算下显著提升尾部召回，优于固定周期修复。
AI 中文摘要
图近似最近邻（ANN）索引（如HNSW、DiskANN/Vamana）在插入/删除变动下会丢失召回率，因为删除操作会使经过已移除节点的贪心搜索路径成为孤路。生产系统通过按固定计划修复图（每X次操作合并一次）来恢复可导航性。我们研究是否基于测量的可导航性退化信号（而非盲时钟）触发局部边修复，能更好地使用固定修复预算。在两个真实ANN数据集（SIFT-128和Fashion-MNIST-784）上，在受控的突发变动流下，并在匹配的摊销修复预算（相等合并次数）下比较修复策略，信号触发的修复帕累托优于固定节奏修复。增益集中在稀缺预算下的最坏情况（尾部）召回：大约一次合并时，它使最小召回@10在四个流种子间提高+0.014（SIFT）到+0.050（Fashion-MNIST），95%置信区间排除零，而平均召回增益很小（<0.005）。优势遵循清晰的漂移严重性梯度——对于更稀疏、更脆弱的图更大——当索引健壮或预算充足时消失。廉价的探测召回信号是真实召回的有效领先指标（Spearman rho ~= 0.95）。我们贡献了该机制、一个将修复调度与修复开销分离的预算匹配评估协议，以及一个开放、可复现的变动-修复工具。我们故意不声称平均召回改进或新索引；召回与修复成本界限以及数据分布漂移耦合留作未来工作。
英文摘要
Graph approximate-nearest-neighbor (ANN) indexes (HNSW, DiskANN/Vamana) lose recall under insert/delete churn, because deletions orphan the greedy-search paths that route through removed nodes. Production systems restore navigability by repairing the graph on a fixed schedule (consolidate every X operations). We asked whether triggering local edge repair on a measured navigability-degradation signal, rather than a blind clock, spends a fixed repair budget better. At matched repair budget, it does not. On two real ANN datasets (SIFT-128 and Fashion-MNIST-784) under a bursty churn stream, compared against a fixed-cadence baseline actually run at the triggered policy's realized consolidation count, the tail-recall advantage is indistinguishable from zero at every operating point, graph degree, and index scale; at several points the clock is better. We trace our earlier positive result to an interpolated baseline: recall is sharply concave in repair budget -- one consolidation captures over half of all achievable gain -- so reading the baseline off a straight line between zero and four passes understates it by more than the effect claimed. Evaluated by the statistic we pre-registered -- correlation with the subsequent recall drop rather than with the concurrent recall level -- the probe signal is also not a leading indicator. What remains is useful: an exact live-set recall oracle, a reproducible churn harness, a drift-severity regime map, and a budget-parity protocol that makes this error detectable. We report the negative result and the trap, because the trap generalizes: any "at matched budget X" comparison whose baseline is read off an interpolated curve, at the scarce end of a concave response, will manufacture an effect favouring the proposal.
Comments10 pages, 2 figures. v2 is a substantial correction: v1's main result and mechanism claim are withdrawn. v1 compared against an interpolated, never-executed fixed-cadence baseline; running it erases the effect. Four pre-registered controls, asserted in v1 as passing, were unimplemented; all now run, and budget parity fails. Code: https://github.com/samyama-ai/updatable-graph-index
