Accelerating Metric Filtering by Improving Bounds on Estimated Distances.

Filtering is a fundamental strategy of metric similarity indexes to minimise the number of computed distances. Given a triple of objects for which distances of two pairs are known, the lower and upper bounds on the third distance can be set as the difference and the sum of these two already known distances, due to the triangle inequality rule of the metric space. For efficiency reasons, the tightness of bounds is crucial, but as angles within triangles of distances can be arbitrary, the worst case with zero and straight angles must also be considered for correctness. However, in data of real-life applications, the distribution of possible angles is skewed and extremes are very unlikely to occur. In this paper, we enhance the existing definition of bounds on the unknown distance with information about possible angles within triangles. We show that two lower bounds and one upper bound on each distance exist in case of limited angles. We analyse their filtering power and confirm high improvements of efficiency by experiments on several real-life datasets.