The PI model, its predictions, and parameters have been used in a number of ways to derive biological insight.
Having determined the posterior distribution of model predictions, there is now a direct link between different predictions and parameters, which can be exploited by determining how predictions relate to each other and to the model parameters.
While this issue invalidates many of the reported evaluation results, it also impedes pre-deployment performance prediction and parameter adjustment.
We have used cross-validation and an additional base pair shifting simulations to show that binding preference prediction and parameter optimization do not result in any (optimistic) bias, or overfitting, in binding prediction accuracy.
The filter coefficients are updated when prediction and parameter errors are larger than user-specified bounds.
Therefore, we first evaluated all LUR models using a bootstrap approach to determine the sensitivity of model prediction and parameter estimates to monitor sampling.
Our proposed approach, GM-SMCC, utilizes the generative model with both network and label regularization for protein function prediction, and parameter estimation is different from original PLSA [ 31] or previous work utilizing PLSA with manifold learning for unsupervised data clustering [ 32].
In order to investigate how well the different methods do in practice in terms of prediction accuracy and parameter convergence, we implemented Viterbi training, Baum-Welch training and stochastic EM training for three small example HMMs.
Note that this numerical analysis is partially based on the bias-variance decomposition of the estimated model predictions and estimated parameters, thus it requires the knowledge of the nominal (true) parameter vector.
We explicitly discuss the trade-off between improved predictions and additional parameters.
