JSUpySVM

This is a multithreaded, multiclass implementation of a weighted, reduced convex hull-based support vector machine algorithm based on this paper by Goodrich et al. I have written the algorithm in Java, and that is mostly complete. Currently, I am working on developing the package with rJava. Uncommented demos can be found at this repo, which include a live OCR demo. My presentation can be found at my website.

As of now, the algorithm can be accessed by calling:

WSVM(df, kernel = "linear",
                  mu = NA,
                  q = 2.0,
                  gamma = 0.01,
                  ep = 1e-2,
                  max_its = 2000,
                  non_sep = 1e-5)

Where df is a dataframe with classes as the first column, and features as the rest. kernel has optiosn of "linear", "rbf", "monomial", and "polynomial". mu is calculated by default when the value is NA as $\frac{1}{0.9\kappa}$ where $\kappa$ equals the sum of all weights in the smallest by weight class. Unfortunately it is not possible to assign weights yet (although the implementation in Java supports it).

q is the exponent of the monnomial/polynomial kernel, gamma is for the rbf kernel. ep is $\epsilon$. max_its is the maximum iterations. non_sep is the threshold of deciding if hulls are nonseparable.

Predictions can be done with:

out$predict(vector)

Where out is the Java object returned from the WSVM function and vector is a feature vector with features in the same order as df.

Installation

require(devtools)
install_github(“mhmdmodan/JSUpySVM”, args="--no-multiarch")

If issues arise with Java/R architecture mismatch, clone the repo and install it using the build tab in RStudio.