Using Support Vector Machine on Large Data

Question

I'm building a model based on a database with around 90000 observations and 100+ variables. My target variable is binary (0 or 1).

I'm using SAS Miner and I want to test a few high performance techniques.

I already tested Random Forest and got very successful results.

Now, I'm giving a try at using Support Vector Machine (SVM) and Radial Basis Function (RBF). However, I'm experiencing a lot of issues running those procedures those due to insufficient memory.

I researched a bit about both procedures but I still have no clue if It's possible to run SVM and RBF with such large data.

If that's not possible, when should I use SVM and RBF?

Thanks in advance!

Edit 1: I was able to run SVM with 22000 observations database and a small amount of variables (five). Still, no success in performing with the bigger data and RBF.

Answer 1

I am not familiar with SaS, but Kernels consume a good amount of memory, you may try dimensionality reduction with PCA, in order to improve memory utilization. Stochastic gradient descent shines in bigger datasets, I mean a neural network. Another recent option worth exploring is Gaussian Process which also uses kernels, I have better results with GP than Random Forest, there is a combination of variational inference for GPs that aims for cases like yours, one library to check is Gpflow, which relies on Tensorflow, well this is Python. I believe if you want to scale and have recent algorithms Python is a must, I hope this helps.