WRSVM

What is WRSVM?

WRSVM combines class-weighted regularization with per-class slack relaxation, producing a multiclass SVM that remains accurate when (a) class sizes are severely imbalanced and (b) a fraction of training labels are corrupted. The framework supports four decomposition strategies (Crammer-Singer, SimMSVM, One-vs-One, One-vs-Rest) crossed with four formulations: SVM, RSVM, WSVM, and WRSVM.

Install

Pick your language. Full troubleshooting is in each package’s INSTALL.md.

# core install
pip install wrsvm

# from a local clone
cd wrsvm_package
pip install -e .

# optional extras
pip install "wrsvm[gpu]"      # PyTorch kernel backend
pip install "wrsvm[gurobi]"   # Gurobi solver
pip install "wrsvm[dev]"      # pytest + coverage

from wrsvm import WRSVMClassifier, inject_outliers_minority
from sklearn.datasets import load_iris
from sklearn.preprocessing import StandardScaler

X, y = load_iris(return_X_y=True)
X = StandardScaler().fit_transform(X)

y_noisy = inject_outliers_minority(X, y, outlier_rate=0.3, seed=0)

clf = WRSVMClassifier(C=100.0, gamma=0.1, upsilon=0.2)
clf.fit(X, y_noisy)
print("accuracy:", clf.score(X, y))

# 1) Python backend (one-time)
reticulate::install_python(version = "3.11.9")
reticulate::py_install("wrsvm", pip = TRUE)

# 2) R wrapper
remotes::install_github("annicenajafi/wrsvm", subdir = "WRSVMr")

# or from a local clone:
remotes::install_local("path/to/WRSVMr")

library(WRSVMr)

data(iris)
X <- scale(as.matrix(iris[, 1:4]))
y <- as.integer(iris$Species)

y_noisy <- inject_outliers_minority(X, y, outlier_rate = 0.3, seed = 0)

fit <- wrsvm_fit(X, y_noisy,
                 strategy = "cs",
                 C = 100, gamma = 0.1, upsilon = 0.2)

mean(wrsvm_predict(fit, X) == y)

using Pkg

# from a local clone (recommended while iterating)
Pkg.develop(path = "path/to/WRSVM.jl")
Pkg.instantiate()

# or activate the package's own environment for reproducibility
Pkg.activate("path/to/WRSVM.jl")
Pkg.instantiate()

using WRSVM, Statistics

X = randn(100, 4)
y = rand(1:3, 100)
y_noisy = inject_outliers_minority(X, y; outlier_rate = 0.3, seed = 0)

model = solve_crammer_singer(X, y_noisy;
                             C = 100.0, gamma = 0.1, upsilon = 0.2)

preds = predict_cs(model, X)
println("accuracy = ", mean(preds .== y))

Decomposition strategies

Downloads

Citation

@article{najafi2026wrsvm,
  title   = {Open-Source Python, R, and Julia Toolkit for Robust Multiclass Classification Under Class Imbalance and Label Noise},
  author  = {Najafi, Annice and Razzaghi, Talayeh},
  year    = {2026}
}