Fig. 3

Multi-block group sparsity structure for proposed reference-based predictor. A multi-block multi-class classifier of K=3 classes applies a K×2p matrix W=[W(r e f),W(t a r g e t)] to the combined vector of probes on reference and target gene chips in order to form the vector of scores for each of the K=3 states. The two blocks of the classifier correspond to the block of weights W(r e f), applied to the reference sample, and the block of weights W(t a r g e t), applied to the target sample. The classifier decision rule is to assign the state label that corresonds to the maximum score. The blocks share sparsity, denoted by the black columns in the weight matrix, which designate columns that are identically zero. Variables associated with these zero columns are not selected by the classifier