Total Unique Variation Analysis

A statistical method for determining the total unique variance in multidimensional data, ordering channels from most to least informative.

the TUVA analysis is a statistical technique for determining the total unique variance in multidimensional data, ordering channels from most to least informative. The TUVA analysis was developed in the context of brain-computer interface development but may be applied to other areas.

The set of data to be analysed consists of a set of channels, each of which contains many samples which may be different locations in space or time, represented as a (samples × channels) matrix of data. This set of data is partitioned into two groups and channels are sequentially added from the first group to the second, beginning with the pair of channels with the lowest pairwise correlation, until the first group is empty. At each step in the sequence, the channel selected for addition to the second group is the channel with the least absolute correlation with its least-squares prediction under multiple linear regression from the channels already included in the second group.

This code includes an embedded demo which shows how this analysis can be applied to sensor field and sensor time-series data.