PLAsTiCC dataset

This notebook gives an example of Active Anomaly Detection with coniferest and PLAsTiCC dataset.

import numpy as np
from time import monotonic
from datetime import timedelta
import matplotlib.pyplot as plt

from coniferest.datasets import plasticc_gp

data, metadata = plasticc_gp()

from coniferest.isoforest import IsolationForest
from coniferest.aadforest import AADForest
from coniferest.pineforest import PineForest

models = {
    "Isolation Forest": IsolationForest,
    "AAD Forest": AADForest,
    "Pine Forest": PineForest,
}

from coniferest.session import oracle

max_iterations = 100
max_trials = 20

steps = np.arange(1, max_iterations + 1)

anomalies = {}
wall_times = {}
for name, model_class in models.items():

    wall_time = []
    anomalies_ = []
    for i in range(max_trials):
        session = oracle.create_oracle_session(data, metadata,
                                               model = model_class(random_seed = i),
                                               max_iterations = max_iterations)

        start_time = monotonic()

        session.run()

        wall_time.append(timedelta(seconds = monotonic() - start_time))

        labels = (1 - np.asarray(list(session.known_labels.values()), dtype=int)) // 2
        anomalies_.append(np.cumsum(labels) / steps)

    anomalies[name] = np.asarray(anomalies_)
    wall_times[name] = np.median(np.asarray(wall_time))

for name in models.keys():
    q5 = np.quantile(anomalies[name], 0.05, axis = 0)
    median = np.median(anomalies[name], axis = 0)
    q95 = np.quantile(anomalies[name], 0.95, axis = 0)

    plt.plot(steps, median,
             alpha=0.75,
             label="{} ({:.0f} sec.)".format(name, wall_times[name].total_seconds()))
    plt.fill_between(steps, q5, q95, alpha=0.5)

plt.legend()
plt.xlabel("Iteration")
plt.ylabel("Fraction of anomalies")

Text(0, 0.5, 'Fraction of anomalies')