Approach to Compare Point Distribution Patterns Produced by Dimension Reduction Techniques

Visualization techniques based on multidimensional data projection techniques are often applied to explore heterogeneous data from multiple sensors of the system. They could be used to produce graphical patterns of both normal and abnormal states of system functioning. However, it is necessary to define metrics and criteria to assess the applicability of different multi-dimensional data projection techniques to detect typical patterns and outliers. This paper analyzes approach based on Delaunay triangulation compare and evaluate distribution of points. The authors also discuss the choice of metric depending on type of multidimensional data projection technique. They are tested on the dataset describing functioning of HVAC system that contains different modes of system functioning.