Do you know why non-linear SVM was not implemented in Apache Spark? I read this page:
https://issues.apache.org/jira/browse/SPARK-4638
Look at the last comment. It says:
"Commenting here on b / c from a recent downstream thread: Non-linear kernels for SVMs in Spark would be healthy. Key barriers: Kernelized SVM training is hard to spread. Naive methods require a lot of communication. To get this feature in Spark, we need to do the right background research and write a good design. Other ML algorithms are perhaps more in demand and still need improvement (as of the date of this comment). In my opinion, tree ensembles are above all.
The question is: why is it difficult to distribute sound SVM?
Everyone knows that non-linear SVMs perform better than linear ones.
source
share