Model Selection using GMRQ

The generalized matrix Rayleigh quotient (GMRQ) is a specific application of the variational principle (adapted from quantum mechanics) for Markov state models and a useful tool for model parameter selection.

The variational principle yields a rigorous way of comparing two different Markov models for the same underlying stochastic process when using different state decompositions. Even under the assumption that you have access to infinite sampling, there is still some error associated with approximating the true continuous eigenfunctions of your modeled process with the indicator functions, as is the case with Markov state models. If we interpret the variational theorem as the measure of the quality of this approximation, the state decomposition that leads to a Markov model with larger leading dynamical eigenvalues is consequently the better state decomposition. If you wish to see the full derivation of this quantity, please refer to [1].

Using this method, we can generate single scalar-valued scores for a proposed model given a supplied data set. This allows for the use of separate testing and training data sets to quantify and avoid statistical overfitting. This method extends these tools, making it possible to score trained models on new datasets and to perform hyperparameter selection. PLEASE NOTE: You cannot use GMRQ to optimize the MSM lag time. Changing the lag time fundamentally alters the model’s eigenfunctions, which no longer makes it a useful scoring function. The number of timescales used to score the model must also be constant and user- selected.

Algorithms

decomposition.tICA.score(sequences[, y])	Score the model on new data using the generalized matrix Rayleigh quotient
msm.MarkovStateModel.score(sequences[, y])	Score the model on new data using the generalized matrix Rayleigh quotient
msm.ContinuousTimeMSM.score(sequences[, y])	Score the model on new data using the generalized matrix Rayleigh quotient

References

[1]	McGibbon, Robert T., and Vijay S. Pande. Variational cross-validation of slow dynamical modes in molecular kinetics J. Chem. Phys. 142, 124105 (2015).