Easing the Dimensionality Curse by Stretching Metric Spaces

Queries over sets of complex elements are performed extracting features from each element, which are used in place of the real ones during the processing. Extracting a large number of significant features increases the representative power of the feature vector and improves the query precision. However, each feature is a dimension in the representation space, consequently handling more features worsen the dimensionality curse. The problem derives from the fact that the elements tends to distribute all over the space and a large dimensionality allows them to spread over much broader spaces. Therefore, in high-dimensional spaces, elements are frequently farther from each other, so the distance differences among pairs of elements tends to homogenize. When searching for nearest neighbors, the first one is usually not close, but as long as one is found, small increases in the query radius tend to include several others. This effect increases the overlap between nodes in access methods indexing the dataset. Both spatial and metric access methods are sensitive to the problem. This paper presents a general strategy applicable to metric access methods in general, improving the performance of similarity queries in high dimensional spaces. Our technique applies a function that "stretches" the distances. Thus, close objects become closer and far ones become even farther. Experiments using the metric access method Slim-tree show that similarity queries performed in the transformed spaces demands up to 70% less distance calculations, 52% less disk access and reduces up to 57% in total time when comparing with the original spaces.