by Guangming Lang
~1 min read

Categories

  • da

The relationship between the number of rows (n) and the number of columns (p) is important for choosing which predictive modeling techniques to use when providing predictive analytics solutions. If n < p, there’s always a risk that a relevant predictor found in the training data set will not be reproducible. In addition, some predictive modeling techniques such as multiple linear regression (and multivariate regression analysis for more than one outcome variables) or linear discriminant analysis cannot be directly used when n < p. Yet, some models such as regression tree, classification tree, random forest and the k nearest neighbors algorithm (KNNs) can be used directly when n < p.

When p is big, i.e., when there’re many predictors, we’re almost guaranteed to have some of them correlated. Predictive models such as partial least squares regression manages correlated predictors by default but is more stable if the predictors are on similar scales. Recursive partitioning such as tree based models has a less stable partitioning structure when facing correlated predictors but is immune to predictors of different scales.