Figure 8.

Selection of model for bicluster using cross-validation (CV). The ordinate represents an estimate of prediction error (Err_CV) from tenfold CV (the mean of the error in the 10 leave-out samples used is the CV error estimate). The shrinkage parameter t allows us to select subsets of predictors continuously. We evaluate our fitted model for a range of values of t (with t = 0 [the null model] and t = 1 [the ordinary least squares solution]). The error bars denote the standard error of Err_CV (the standard deviation of the 10 leave-out samples' error estimates). The red line shows the value of t selected for our final model for this cluster - the most parsimonious model within 1 standard error of the minimum on the Err_CV versus t curve.