Figure 2

Selecting the model that minimizes the breakpoint annotation error. Training error functions for global and local models plotted against smoothing parameter λ. The original annotation data set was used to calculate the annotation error. In the top row panels, we plot E^global(λ) from Equation 6, and in the other rows, we plot $E_i^{\text{local}}\left(\lambda \right)$ from Equation 5. Each column of plots shows the error of a particular algorithm, and the minimum chosen using the global training procedure is shown using a vertical grey line. Note that the local model training error can be reduced by moving from the globally optimal smoothing parameter $\widehat{\lambda }$ to a local value ${\widehat{\lambda }}_i$, as in profile i=375 for dnacopy.sd. For the local models trained on single profiles, many smoothing parameters attain the minimum. So we use the protocol described in the “Selecting the optimal degree of smoothness” section to select the best value, shown as a black dot.