Hi,
I am facing a hierarchically structured dataset, and I am not sure of the right way to analyses it with cforest, if their is one. - - BACKGROUND & PROBLEM We are analyzing the behavior of some social birds facing different temperature conditions. The behaviors of the birds were recorder during many sessions of 2 hours. Conditional RF (cforest) are quite useful for this analysis since, we have a large number of variables describing the temperature during the 2 hours, they are rather highly correlated, and we expect they have some non linear effects on the behavior. For the other behaviors, for each individual and each session of 2 hours, we recorded the frequency. For each session of 2 hours, we have only one value for the variables related to the temperature, since these variables are for example minimal and maximal temperature, median temperature, and different measures of the variance of the temperature. Visually the dataset thus looks like this: Y_behaviour_frequency Individual Session X1 X2 X3 ... 0.5 ind1 S1 5 10 7 ... 0.55 ind2 S1 5 10 7 ... 0.2 ind3 S1 5 10 7 ... ... S1 5 10 7 ... 0.3 ind1 S2 15 7 50 ... 0.01 ind5 S2 15 7 50 ... ... S2 15 7 50 ... 0.4 ind1 S3 2 8 5 ... 0.05 ind3 S3 2 8 5 ... 0.1 ind4 S3 2 8 5 ... 0.2 ind5 S3 2 8 5 ... ... S3 2 8 5 ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... If I run a classical cforest on this dataset, explaining Y_behaviour_frequency with Individual, Session and all the X... variables, I end up with some conditional relative importances similar to the attached plot: They are all very very low, but none is negative. The absence of negative conditional relative importance is annoying since we were selecting variables using the threshold of minus two times the lowest conditional relative importance. - - QUESTIONS 1) have you ever faced one of these situations of - all very low conditional relative importances - all positives conditional relative importances - hierarchically structured dataset analyzed with cforest ? I think, but I am not sure, the very low but all positive conditional importance might come from the hierarchically structured dataset: Since RF are based on bootstraps, when bootstrapping in at each iteration, all sessions or almost all sessions of 2 hours are sampled, although they are the main source of variation. The bootstrap would need to be itself hierarchic, first sampling the sessions and then sampling the individual in the sampled session of 2 hours. 2) It's easy to perform such kind of hierarchic bootstrap in R, but have you ever heard about it in a random forest ? The question was asked 4 years ago: here: https://stats.stackexchange.com/questions/62840/random-forest-and-cluster-level-bootstrapping and here: https://stats.stackexchange.com/questions/93156/random-forest-on-multi-level-hierarchical-structured-data but the main track "hie-ran-forest" also called "HieRanFor" seems aborted. (https://r-forge.r-project.org/R/?group_id=2021) Thanks for your help, cheers. hugo -- - no title specified Hugo Mathé-Hubert BU-G19 postdoc eawag (Swiss Federal Institute of Aquatic Science and Technology) Evolutionary Ecology <http://www.eawag.ch/en/department/eco/main-focus/evolutionary-ecology/>- About me <http://www.eawag.ch/en/aboutus/portrait/organisation/staff/profile/hugo-mathe-hubert/> Überlandstrasse 133 P.O.Box 611 8600 Dübendorf, Switzerland - - - - - - - - - - - - - - - - - - Thoughts appear from doubts and die in convictions. Therefore, doubts are an indication of strength and convictions an indication of weakness. Yet, most people believe the opposite. - - - - - - - - - - - - - - - - - - Les réflexions naissent dans les doutes et meurent dans les certitudes. Les doutes sont donc un signe de force et les certitudes un signe de faiblesse. La plupart des gens sont pourtant certains du contraire. ______________________________________________ [hidden email] mailing list -- To UNSUBSCRIBE and more, see https://stat.ethz.ch/mailman/listinfo/r-help PLEASE do read the posting guide http://www.R-project.org/posting-guide.html and provide commented, minimal, self-contained, reproducible code. Rplot.pdf (6K) Download Attachment |
Free forum by Nabble | Edit this page |