Causes of filtered ANN query failures and their relationship to filter selectivity

Determine the specific mechanisms that cause failures of filtered approximate nearest neighbor queries under standard graph-based strategies (post-filtering, pre-filtering, and traversal-time filtering), and ascertain how these failure modes depend on the selectivity of the metadata filter.

Background

The authors argue that existing filtered search strategies often treat filtering as an operational constraint and rely on standard greedy navigation, without accounting for how filtering alters the search-space geometry. As a result, queries can fail due to structural changes in the graph when nodes and edges are removed by filters.

Understanding the failure mechanisms and their dependence on filter selectivity is essential for designing robust filtered search algorithms and for explaining when and why commonly used strategies break down.