Optimizing the Distance Computation Order of Multi-Feature Similarity Search Indexing

Abstract

Multi-feature search is an effective approach to similarity search. Unfortunately, the search efficiency decreases with the number of features. Several indexing approaches aim to achieve efficiency by incrementally reducing the approximation error of aggregated distance bounds. They apply heuristics to determine the distance computations order and update the object’s aggregated bounds after each computation. However, the existing indexing approaches suffer from several drawbacks. They use the same computation order for all objects, do not support important types of aggregation functions and do not take the varying CPU and I/O costs of different distance computations into account. To resolve these problems, we introduce a new heuristic to determine an efficient distance computation order for each individual object. Our heuristic supports various important aggregation functions and calculates cost-benefit-ratios to incorporate the varying computation costs of different distance functions. The experimental evaluation reveals that our heuristic outperforms state-of-the-art approaches in terms of the number of distance computations as well as search time.