Figure 1
From: Infinite mixture-of-experts model for sparse survival regression with application to breast cancer
Dummy coding illustration. On the top-left, categorical observations for 2 patients are shown for whom 2 biomarkers (X 1 and X 2) are measured for expression values. Each biomarker (categorical variable) can have three possible values (high, med and low). The top-right side shows the transformed covariate matrix after the dummy coding procedure has been applied. The resulting design matrix represents each variable as a group of dummy-variables. Hence identifying key features from the original matrix is translated to the problem of identifying key groups of dummy variables. The bottom-right shows the transformed matrix after using a polynomial contrast coding procedure. The resulting contrast-coded matrix uses (K − 1)order+1 columns for an interaction as opposed to (K)order+1 columns in a a dummy-coded matrix where K is the number of categories for a variable and order denotes the order number of the interaction (zeroth, first, second etc).