PSIS Cross Validation with WAIC
Sumio Watanabe Homepage
Key words: PSISCV, ISCV, WAIC
Aki Vehtari, Andrew Gelman, and Jonah Gabry
proposed Pareto Smoothed Importance Sampling Cross Validation (PSISCV).
Aki Vehtari, Andrew Gelman, and Jonah Gabry,
``Efficient implementation of leave-one-out cross-validation and WAIC
for evaluating fitted Bayesian models," http://arxiv.org/abs/1507.04544
Let us compare PSISCV with WAIC
from the viewpoint of statistical estimation of the generalization error
in a simple regression problem.
If a leverage sample is contained in a training set
(such a case is called influential observation),
then ISCV has the infinite variance. PSISCV was proposed for such a case.
We compare PSISCV, ISCV, and WAIC in influential observations.
(Conclusion) PSISCV, ISCV, and WAIC were
compared from the viewpoint of statistical estimators of the generalization error.
(1) From the viewpoint of statistical estimation
of the generalization error, the difference among PSISCV, ISCV, and WAIC is smaller
than the fluctuation of the generalization error, even in influential observation.
(2) In experiments, E|PSISCV-GE| was smaller than E|ISCV-GE|.
(3) In experiments, E|WAIC-GE| was smaller than E|ISCV-GE|.
(4) If the standard deviation of a leverage sample
is 100 times as large as other samples, then E|PSISCV-GE| was almost equal to E|WAIC-GE|.
(5) If otherwise, E|PSISCV-GE| was larger than E|WAIC-GE|.
It seems that there is a trade-off structure in estimation of the generalization error
in influential observation.
If we have MCMC posterior samples, then it is easy to calculate WAIC, ISCV, and PSISCV.
Thus I recommend that all criteria had better be calculated and compared.