I am using boosted regression trees (BRT) to investigate the environmental influences of shark presence at a single location. I have a year-long dataset of acoustic tag detections and environmental data (e.g. current speed and direction etc.) which I have split into 1hrly time bins, i.e. the number of detections every hour and the hourly mean for each environmental variable. Having researched BRTs extensively, it seems that temporal autocorrelation (serial correlation) is not addressed (although spatial autocorrelation is). Having built several models I used the acf() function and a Durbin–Watson test on the model's residuals and it is showing some degree of autocorrelation (DW test = 1.09). Is serial correlation a problem with BRTs? If so, can anyone suggest a way to 'fix' the issue? Many thanks!
DR Harris--you may want to peruse the following links;
https://arxiv.org/pdf/1502.02513.pdf
Boosted Regression Trees - MDPI
📷www.mdpi.com › pdf
PDFby H Gu - 2019 -
Article Cross-validation strategies for data with temporal, spatial,...
Kind regards--
Thank you Ijaz Durrani, but this article seems to only address spatial autocorrelation whereas I am concerned about temporal autocorrelation (serial correlation).
Hello DR Harris: have a look at the following links: https://www.researchgate.net/post/Does_random_forest_account_for_temporal_auto-correlation https://www.researchgate.net/post/In_BRT_boosted_regression_trees_model_for_instance_using_GBM_R_package_how_do_you_make_custom_non_random_CV_folds https://pdfs.semanticscholar.org/ada8/058a03fdf61fc6a81131c8ff0cd011b440cd.pdf
Kind regards--
- Badges
- Science topic