ENH: Add OTOS oversampling algorithm

pyloras/__init__.py

@@ -1,3 +1,4 @@
 from ._loras import LORAS
 from ._prowras import ProWRAS
 from ._gamus import GAMUS
+from ._otos import OTOS

pyloras/_otos.py

@@ -0,0 +1,94 @@
+from imblearn.over_sampling.base import BaseOverSampler
+from imblearn.utils import Substitution
+from imblearn.utils._docstring import (
+ _random_state_docstring,
+ _n_jobs_docstring,
+)
+from sklearn.svm import LinearSVC
+import numpy as np
+
+from ._common import check_random_state, safe_random_state
+
+
+@Substitution(
+ sampling_strategy=BaseOverSampler._sampling_strategy_docstring,
+ random_state=_random_state_docstring,
+ n_jobs=_n_jobs_docstring
+)
+class OTOS(BaseOverSampler):
+ def __init__(
+ self,
+ *,
+ sampling_strategy="auto",
+ svc_reg=1.0,
+ ot_reg=1.0,
+ tradeoff=1.0,
+ random_state=None,
+ max_iter=100,
+ ):
+ super().__init__(sampling_strategy=sampling_strategy)
+ self.svc_reg = svc_reg
+ self.ot_reg = ot_reg
+ self.tradeoff = tradeoff
+ self.random_state = random_state
+ self.max_iter = max_iter
+
+ def _fit_resample(self, X, y):
+ import ot
+ random_state = check_random_state(self.random_state)
+ X_res = [X.copy()]
+ y_res = [y.copy()]
+ svc = LinearSVC(
+ loss="hinge", C=self.svc_reg, random_state=safe_random_state(random_state)
+ )
+ for minority_class, samples_to_make in self.sampling_strategy_.items():
+ if samples_to_make == 0:
+ continue
+ X_p = X[y == minority_class]
+ X_n = X[y != minority_class]
+ n_p = X_p.shape[0]
+ n_n = X_n.shape[0]
+ n_r = samples_to_make
+ one_r = np.ones((n_r, 1))
+ one_n = np.ones((n_n, 1))
+ # set initial distribution for mu_r and mu_p
+ mu_r = np.asarray([1.0 / n_r] * n_r)
+ mu_p = np.asarray([1.0 / n_p] * n_p)
+ T = mu_r[:, None] @ mu_p[:, None].T
+ # manufactor a binary classification problem
+ _y = np.empty_like(y)
+ _y[y == minority_class] = 0
+ _y[y != minority_class] = 1
+ svc.fit(X, _y)
+ w = svc.coef_.T
+
+ hingelosses = np.concatenate(
+ [
+ np.atleast_1d(max(1 - y_i * svc.coef_ @ x_row, 0.0))
+ for y_i, x_row in zip(y[y == minority_class], X_p)
+ ]
+ )
+ mu_p = np.exp(hingelosses)
+ mu_p /= mu_p.sum()
+
+ D_r = np.diag(1 / mu_r)
+ X_r = D_r @ T @ X_p
+ # C_p = np.apply_along_axis(c_row, axis=-1, arr=X_r)
+ C_p = np.asarray(
+ [
+ [np.linalg.norm(x_row - row) for row in X_p]
+ for x_row in X_r
+ ]
+ )
+ wwT = w @ w.T
+ Theta = self.tradeoff * C_p.T - X_p @ np.kron(one_r.T, wwT @ X_n.T @ one_n + n_n * w) @ D_r
+ Phi = X_p @ wwT @ X_p.T
+ Psi = D_r.T @ D_r
+
+ for _ in range(self.max_iter):
+ transport_cost = Theta.T + n_n * Psi @ T @ Phi
+ T = ot.sinkhorn(mu_r, mu_p, transport_cost, self.ot_reg)
+ X_res.append(D_r @ T @ X_p)
+ y_res.append([minority_class] * n_r)
+
+ return np.concatenate(X_res), np.concatenate(y_res)

pyproject.toml

@@ -27,6 +27,9 @@ classifiers = [
  "Operating System :: MacOS :: MacOS X",
 ]
 
+[project.optional-dependencies]
+otos = ["POT"]
+
 [project.urls]
 source = "https://github.com/zoj613/pyloras"
 tracker = "https://github.com/zoj613/pyloras/issues"

requirements-dev.txt

@@ -2,5 +2,6 @@
 build==0.10.0
 imbalanced-learn==0.10.1
 numpy==1.23.2
+POT==0.9.0
 pytest==7.3.1
 pytest-cov==4.0.0

tests/test_otos.py

@@ -0,0 +1,23 @@
+import numpy as np
+import pytest
+from sklearn.datasets import make_classification
+
+from pyloras import OTOS
+
+
+@pytest.fixture
+def data():
+ return make_classification(n_samples=150, n_features=4, n_informative=4,
+ n_redundant=0, n_repeated=0, n_classes=3,
+ n_clusters_per_class=2,
+ weights=[0.01, 0.05, 0.94],
+ class_sep=0.8, random_state=0)
+
+
+def test_otos(data):
+ X, y = data
+ rng = np.random.RandomState(12345)
+ otos = OTOS(random_state=rng)
+ X_res, y_res = otos.fit_resample(X, y)
+ _, y_counts = np.unique(y_res, return_counts=True)
+ np.testing.assert_allclose(y_counts[0], y_counts[1])