Solving All-k-Nearest Neighbor Problem without an Index

Abstract

Among the similarity queries in metric spaces, there are one that obtains the k-nearest neighbors of all the elements in the database (All-k-NN). One way to solve it is the naïve one: comparing each object in the database with all the other ones and returning the k elements nearest to it (k-NN). Another way to do this is by preprocessing the database to build an index, and then searching on this index for the k-NN of each element of the dataset. Answering to the All-k-NN problem allows to build the k-Nearest Neighbor graph (kNNG). Given an object collection of a metric space, the Nearest Neighbor Graph (NNG) associates each node with its closest neighbor under the given metric. If we link each object to their k nearest neighbors, we obtain the k Nearest Neighbor Graph (kNNG).The kNNG can be considered an index for a database, which is quite efficient and can allow improvements. In this work, we propose a new technique to solve the All-k-NN problem which do not use any index to obtain the k-NN of each element. This approach solves the problem avoiding as many comparisons as possible, only comparing some database elements and taking advantage of the distance function properties. Its total cost is significantly lower than that of the naïve solution.