Scoring samples based on the nudge mechanism

RANDOM_SEED = 42
VERBOSE = False

# generate train / test data
from examples.utils.dataset import generate_train_data, generate_test_data
X_train, y_train = generate_train_data(n_normal=100, n_abnormal=20, random_seed=RANDOM_SEED)
X_test, y_test = generate_test_data(n_normal=100, n_abnormal=20, random_seed=RANDOM_SEED*2)

import matplotlib.pyplot as plt
import numpy as np
from pan import ParallelAnomalousNudge

# train PAN
model = ParallelAnomalousNudge.with_derived_estimators(X_train, y_train, nu=.1, random_seed=RANDOM_SEED, verbose=VERBOSE)
model.fit(X_train, y_train)

ParallelAnomalousNudge(estimators={np.int64(0): SomRepresentationEstimator(nu=0.1,
                                                                           random_seed=42,
                                                                           som_representation=<minisom_representation.som_representation.SomRepresentation object at 0x7fc2911ae810>),
                                   np.int64(1): SomRepresentationEstimator(nu=0.1,
                                                                           random_seed=42,
                                                                           som_representation=<minisom_representation.som_representation.SomRepresentation object at 0x7fc2904e79b0>)},
                       nu=0.1, random_seed=42, scaler=StandardScaler())

In a Jupyter environment, please rerun this cell to show the HTML representation or trust the notebook.
On GitHub, the HTML representation is unable to render, please try loading this page with nbviewer.org.

fig, axes = plt.subplots(2, 1, figsize=(10, 10), sharex=True, sharey=True)
axes = axes.flatten()

scores = model._score_component_normal(X_test)
y_values = [np.ones_like(scores[y_test == 0]), np.ones_like(scores[y_test == 1])]

ax = axes[0]
ax.set_title(r"Anomaly score considering only the Normal representation", loc="left")
ax.text(0, .8, s="inlier", va="center", ha="center", fontsize=12, color="white", bbox=dict(color="seagreen", boxstyle='round,pad=.6'), zorder=4)
ax.text(min(scores), .8, s="outlier", va="center", ha="center", fontsize=12, color="white", bbox=dict(color="indianred", boxstyle='round,pad=.6'), zorder=4)
ax.violinplot(scores, positions=[1], vert=False)
ax.scatter(scores[y_test == 0], y=y_values[0], marker="o", edgecolor="k", color="none", s=400, alpha=.75, label="True normal", zorder=2)
ax.scatter(scores[y_test == 1], y=y_values[1], marker="x", color="crimson", s=200, alpha=.75, label="True anomaly", zorder=3)
ax.legend()
ax.margins(.1)

scores = model.score_samples(X_test)

ax = axes[1]
ax.set_title(r"Nudged anomaly score", loc="left")
ax.text(0, .8, s="inlier", va="center", ha="center", fontsize=12, color="white", bbox=dict(color="seagreen", boxstyle='round,pad=.6'), zorder=4)
ax.text(min(scores), .8, s="outlier", va="center", ha="center", fontsize=12, color="white", bbox=dict(color="indianred", boxstyle='round,pad=.6'), zorder=4)
ax.violinplot(scores, positions=[1], vert=False)
ax.scatter(scores[y_test == 0], y=y_values[0], marker="o", edgecolor="k", color="none", s=400, alpha=.75, label="True normal", zorder=2)
ax.scatter(scores[y_test == 1], y=y_values[1], marker="x", color="crimson", s=200, alpha=.75, label="True anomaly", zorder=3)
ax.legend()
ax.margins(.1)

plt.yticks([])
plt.show()

/home/runner/work/PAN/PAN/examples/demonstrate_06_inspect_score_samples_nudged_aspect.py:41: MatplotlibDeprecationWarning: vert: bool was deprecated in Matplotlib 3.11 and will be removed in 3.13. Use orientation: {'vertical', 'horizontal'} instead.
  ax.violinplot(scores, positions=[1], vert=False)
/home/runner/work/PAN/PAN/examples/demonstrate_06_inspect_score_samples_nudged_aspect.py:53: MatplotlibDeprecationWarning: vert: bool was deprecated in Matplotlib 3.11 and will be removed in 3.13. Use orientation: {'vertical', 'horizontal'} instead.
  ax.violinplot(scores, positions=[1], vert=False)

scores_0 = model._score_component_normal(X_test)
scores_n = model.score_samples(X_test)
nudged_inds = scores_n < scores_0
print(f"Number of nudged samples: {nudged_inds.sum()}")

Number of nudged samples: 13

y_values = [np.ones_like(scores[y_test == 0]), np.ones_like(scores[y_test == 1])]
plt.figure(figsize=(8, 5))
plt.title(r"Effect of nudge", loc="left")
plt.text(0, .8, s="inlier", va="center", ha="center", fontsize=12, color="white", bbox=dict(color="seagreen", boxstyle='round,pad=.6'), zorder=4)
plt.text(min(scores), .8, s="outlier", va="center", ha="center", fontsize=12, color="white", bbox=dict(color="indianred", boxstyle='round,pad=.6'), zorder=4)
plt.violinplot(scores, positions=[1], vert=False)
plt.scatter(scores_0[nudged_inds], y=np.ones_like(scores_0[nudged_inds]), marker=".", edgecolor="k", color="k", s=300, alpha=.4, zorder=2, label="Normalcy score")
plt.scatter(scores_n[nudged_inds], y=np.ones_like(scores_n[nudged_inds]), marker=7, color="purple", s=300, alpha=.5, zorder=3, label="Nudged score")
plt.legend()
plt.margins(.1)
plt.yticks([])
plt.show()

/home/runner/work/PAN/PAN/examples/demonstrate_06_inspect_score_samples_nudged_aspect.py:76: MatplotlibDeprecationWarning: vert: bool was deprecated in Matplotlib 3.11 and will be removed in 3.13. Use orientation: {'vertical', 'horizontal'} instead.
  plt.violinplot(scores, positions=[1], vert=False)