Reducing the Dimensions of Attributes by selection and Aggregation

Abstract

Given a matrix of objects and attributes, dual scaling analysis (DSA) [1] is known to be a promising method of reducing the dimensionality while preserving the underlying structure between objects and attributes [2], [3]. However, due to the computational complexity of matrix inversion, DSA suffers from a scalability problem for data with tens of thousands of attributes, as is often the case in information retrieval applications. The problem thus becomes how to reduce the dimension of the original data at the pre-processing stage, to make analysis feasible. Our study calculates the comparative data losses of two schemes for such dimension reduction, feature selection and feature aggregation, and proposes a procedure for combining these two schemes. We also evaluate performance using HTTP log data