Human Activity Recognition in AAL Environments Using Random Projections
Table 2
Summary of feature selection/dimensionality reduction methods in HAR.
Method
Advantages
Disadvantages
Complexity
PCA
High dimensionality reduction; reduction of noise; lack of redundancy of data due to orthogonality of components
The covariance matrix is difficult to be evaluated accurately; even the simplest invariance could not be captured by the PCA unless the training data explicitly provides for it
, where are data points, each represented with features
ReliefF
Low computational complexity
Unstable due to random selection of instances
ā
Rankfeatures
Features highly correlated with already selected features are less likely to be included
It assumes that data classes are normally distributed
It depends upon class separability criterion
CFS
It evaluates a subset of features rather than individual features
It fails to select locally predictive features when they are overshadowed by strong, globally predictive features