Horizontal Reduction: Instance-Level Dimensionality Reduction for Similarity Search in Large Document Databases