1

I'm attempting to use boosted regression trees to look at the effect different fish groups have on coral using the dismo package in R following Elith 2008 and 2017

However the training data correlation and the cv correlation looks to be low, is there an acceptable cutoff I should be aiming for?

Using this code

testmodel <- gbm.step(data=brtdata, gbm.x = c(10:12), gbm.y = 13,
                            family = "poisson", tree.complexity = 5,
                            learning.rate = 0.0005, bag.fraction = 0.6,
                            max.trees=5000)

I get this result

fitting final gbm model with a fixed number of 3450 trees for jcount

mean total deviance = 19.611 
mean residual deviance = 13.077


estimated cv deviance = 15.676 ; se = 1.432


training data correlation = 0.58 
cv correlation =  0.409 ; se = 0.071


elapsed time -  0.17 minutes
kjetil b halvorsen
  • 77,844
Peran Bray
  • 11
  • I haven't use BRTs of family = "poisson" however, I expect interpretation of the results is similar to bernoulli. Calculate the deviance explained to see how well your model is fitting the data, please see this link – Jo Harris Jul 28 '20 at 15:25
  • Alternatively, you could try the ggPerformace function of the ggBRT R package which can help you assess how well your model is performing. – Jo Harris Jul 28 '20 at 15:33

0 Answers0