[docs]class LodaDetector(BaseDetector):
[docs] def __init__(
self, window_len: int = 100, random_cuts_num: int = 100,
):
"""Multivariate LODA Detector :cite:`DBLP:journals/ml/Pevny16`.
Args:
window_len (int, optional): The length of window. Defaults to 100.
random_cuts_num (int, optional): The number of random experiments. Defaults to 100.
"""
super().__init__()
self.window_len = window_len
self.window = deque(maxlen=window_len)
self.random_cuts_num = random_cuts_num
self.bins_num = int(1 * (window_len ** 1) * (np.log(window_len) ** -1))
self._weights = np.ones(random_cuts_num) / random_cuts_num
self.components_num = None
self.nonzero_components_num = None
self.zero_components_num = None
self._projections = None
self._histograms = None
self._limits = None
def fit(self, X: np.ndarray):
self.window.append(X)
if self.index == 0:
self.components_num = len(X)
self.nonzero_components_num = int(np.sqrt(self.components_num))
self.zero_components_num = (
self.components_num - self.nonzero_components_num
)
elif len(self.window) == self.window.maxlen:
self._projections = np.random.randn(
self.random_cuts_num, self.components_num
)
self._histograms = np.zeros([self.random_cuts_num, self.bins_num])
self._limits = np.zeros([self.random_cuts_num, self.bins_num + 1])
for i in range(self.random_cuts_num):
rands = np.random.permutation(self.components_num)[
: self.zero_components_num
]
self._projections[i, rands] = 0.0
projected_data = self._projections[i, :].dot(
np.array(self.window).T
)
self._histograms[i, :], self._limits[i, :] = np.histogram(
projected_data, bins=self.bins_num, density=False
)
self._histograms[i, :] += 1e-12
self._histograms[i, :] /= np.sum(self._histograms[i, :])
return self
def score(self, X: np.ndarray):
score = 0
for i in range(self.random_cuts_num):
projected_data = self._projections[i, :].dot(np.array(X).T)
inds = np.searchsorted(
self._limits[i, : self.bins_num - 1],
projected_data,
side="left",
)
score += -self._weights[i] * np.log(self._histograms[i, inds])
score = score / self.random_cuts_num
return float(score)
