4

does anyone know how exactly the Deviances (Poisson, Gamma, Tweedie) are computed in H2O? I cannot find the functions. For interpretation purposes I would like to know the calculations.

Thank you!

Zugi
  • 41
  • 1
  • 2
    It's defined in here: https://github.com/h2oai/h2o-3/blob/master/h2o-algos/src/main/java/hex/glm/GLMMetricBuilder.java but it's still not clear to me what the explicit formula is – Dan Mar 06 '18 at 15:36

2 Answers2

2

The formulas are in the H2O User Guide in the GLM section. Here are links for the deviance formulas for

  1. Linear: $$D=\sum^{N}_{i=1}{(y_{i}-\hat{y}_{i})^{2}}$$
  2. Logistic: $$-D=\sum^{N}_{i=1}{\left(y_{i}\log{(\hat{y}_{i})}+(1-y_{i})\log{(1-\hat{y}_{i})}\right)}$$
  3. Poisson: $$D=-2\sum^{N}_{i=1}{\left(y_{i}\log{\left(\frac{y_{i}}{\hat{y}_{i}}\right)}-\left(y_{i}-\hat{y}_{i}\right)\right)}$$
  4. Gamma: $$D=2\sum^{N}_{i=1}{\left(-\log{\left(\frac{y_{i}}{\hat{y}_{i}}\right)}+\frac{\left(y_{i}-\hat{y}_{i}\right)}{\hat{y}_{i}}\right)}$$
  5. Tweedie: where $p\neq1$ and $p\neq2$ $$D=-2\sum^{N}_{i=1}{\left(\frac{y_{i}\left(y_{i}^{1-p}-\hat{y}_{i}^{1-p}\right)}{1-p}-\frac{\left(y_{i}^{2-p}-\hat{y}_{i}^{2-p}\right)}{2-p}\right)}$$

Note for Tweedie there is an error in the docs, you can see the formula actually used in the h2o source, line 395 here: https://github.com/h2oai/h2o-3/blob/master/h2o-algos/src/main/java/hex/glm/GLMModel.java. The formula in this answer shows the correct formula rather than the formula in the docs.

Dan
  • 1,288
  • 2
  • 12
  • 30
Erin LeDell
  • 765
  • 3
  • 11
0

The deviance equations in H2O should be following the general specifications used typically in actuarial sciences. See for instance Modern Actuarial Risk Theory by Kaas et al. page $308$ for the Tweedie Deviance and pages $246$ and $247$ for Poisson and Gamma Deviances. They correspond to the same equations in the accepted answer above, save the weights.

vosirus
  • 1