I am building species distribution models with boosted regression trees and support vector machines using a large number of Presence-(Pseudo)absence data (> 10.000 plots) Since my goal is not to predict occurences but rather probabilities of occurence, I would like to use support vector regression rather than classification.

However, even though the results look promising so far, I am not sure if treating the binary Y as a continuous variable makes sense statistically. I am familiar using SVM for classifiation (e.g. of remote sensing data) but have never used the ε-regression method before. I am using the e1071-package in R with an RBF-kernel, 19 predictors in total.

From how I see it, the application should be relatively simple, since like SVM classification, SVR works by transforming the data into a feature space, then performing linear regression in that vector space with the linear ε-insensitive loss function ignoring errors that are within the ε distance of the observed value by treating them as zero. After considering the scaling of predictions as well as the prevalence of presence-data, I believe it should be possible to interpret results as occurence probabilities. However, since I'm not coming from a data mining - background, I feel like it's easy to miss some basic concept of SVR that I haven't understood so far. In fact, I haven't found any literature about SVR treating a binary variable as a continuous. Any advise or hints on that matter would be highly appreciated. Thank you


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Browse other questions tagged or ask your own question.