Figure 2

The architecture of an ECOC-LDPC classifier under bounded gene selection. Right squares represent constituent ECOC classifiers induced from simple parity codes, left squares represent practical binary classifiers, rectangles represent "channel" functions, ellipses represent binary predictions, and small circles represent gene expression measurements. Edges are put between constituent ECOC classifiers and ideal binary predictions c _i taking care that the connectivity profile remains sparse (j < <m, u < <n). Ideal binary predictions c _i and practical ones r _i are constrained by "channel" functions f _i modeling prior statistical knowledge about binary classifiers errors, i = 1,..., n. Each practical binary classifier L _i selects a subset of v genes from a pool of p genes taking care that no more than $\frac{\lceil {\log }_{2}M\rceil }{n\cdot H\left(\frac{1}{p}\right)}$ genes get selected.