Learning Filter-Aware Distance Metrics for Nearest Neighbor Search with Multiple Filters

We study the problem of approximate nearest neighbor (ANN) search under multiple conjunctive attribute filters — a common setting in enterprise search and recommendation systems. Standard ANN indexes are built with a single global distance metric and struggle when queries are restricted to subsets of the data defined by combinations of filters. We propose a method that learns filter-aware distance metrics, adapting the embedding space to the distribution of each filter combination. Our approach significantly improves retrieval quality and efficiency compared to generic metric baselines.

Keywords: Nearest Neighbor Search, Metric Learning, Approximate Nearest Neighbor, Filtered Search, Distance Metrics

Recommended citation: Sutradhar, A., Gupta, S., Krishnaswamy, R., Xu, H., Rastogi, A., & Srinivasa, G. (2025). "Learning Filter-Aware Distance Metrics for Nearest Neighbor Search with Multiple Filters." arXiv preprint arXiv:2511.04073.
Download Paper