Novel approach for nearest neighbor search in high dimensional space

Index structures for nearest neighbor search in high-dimensional metric space are mostly built by partitioning the data set based on distances to certain reference point(s). Using the constructed index, the search is limited to a smaller number of the partitions in a way to avoid exhaustive search. However, the approaches already described in the literature either ignore the property of the data distribution or produce non-disjoint partitions; this greatly aspects the search efficiency. In this paper, we propose a new index structure, which overcomes the above disadvantages. The proposed tree structure is constructed by recursively dividing the data set into a nested set of approximate equivalence classes. We also propose a new reference point selection method using principal component analysis (PCA). The conducted analysis and the reported test results demonstrate that the proposed index structure, empowered by the PCA-based reference selection strategy, gives an optimal partition of the data set and greatly improves the search efficiency compared to the VP-tree, which is one of the approaches well documented in the literature.