Quickstart#

This page runs milia on a synthetic population of meters: sixty meters, ninety days, and three kinds of daily pattern. Every cell runs when the documentation is built.

A population of series#

from milia.data import profiles, synthetic

population = synthetic.generate(meters=60, days=90, steps=48, rate=0.03)
shapes = profiles.shape(population.pivot)  # the peak-normalised mean day of each meter
shapes.shape
(60, 48)

Cluster#

The number of groups is chosen from the data, and a trust gate keeps a split that is only an artefact of one sample.

from milia import make
from milia.cluster import selection, validation

scores = validation.sweep(shapes, range(2, 8), resampling=True)
k = selection.recommend(scores, strategy="trust")
labels = make("clusterer", "kmeans", n_clusters=k).fit_predict(shapes)
k, [int((labels == c).sum()) for c in range(k)]
(3, [20, 20, 20])

Detect#

An ensemble of detectors is fitted on normal days. A threshold calibrated on held-out normal days bounds the share of normal days that are flagged.

from milia.anomaly.ensemble import Ensemble
from milia.features.window import vector

days = profiles.tensor(population.pivot)[0]  # one meter: days by steps
features = [vector(day) for day in days]
detector = Ensemble().fit(features[:60]).calibrate(features[60:80], alpha=0.1)
flags = detector.flag(features[80:])
int(sum(flags)), len(flags)
(1, 10)

Rank#

A retriever returns the nearest labelled cases to a query, and says whether the query is close enough to trust them.

import numpy as np

from milia.recommend.retrieval import KNNRetriever

retriever = KNNRetriever(n_neighbors=3).fit(shapes[1:], np.zeros(len(shapes) - 1))
distances, neighbours = retriever.index_.kneighbors(shapes[:1])
neighbours[0], distances[0].round(3)
(array([ 5, 50, 53]), array([0.026, 0.031, 0.032]))

Next, read the key concepts, then the user guide for the task you need.