A general approach for similarity-based linear projections using a genetic algorithm

A widely applicable approach to visualizing properties of high-dimensional data is to view the data as a linear projection into two- or three-dimensional space. However, developing an appropriate linear projection is often difficult. Information can be lost during the projection process, and many linear projection methods only apply to a narrow range of qualities the data may exhibit. We propose a general-purpose genetic algorithm to develop linear projections of high-dimensional data sets which preserve a specified quality of the data set as much as possible. The obtained results show that the algorithm converges quickly and reliably for a variety of different data sets.