Time Series Clustering Experiments

In this notebook we will show how to run a simple time series clustering experiment using the tsml-eval package. Time series clustering is the task of grouping time series into similar groups.

​

import numpy as np
import pandas as pd
from aeon.benchmarking import get_estimator_results
from aeon.datasets import load_classification
from aeon.visualisation import plot_critical_difference
from sklearn import preprocessing
from tsml.datasets import load_minimal_chinatown
from tsml.dummy import DummyClusterer

from tsml_eval.evaluation.metrics import clustering_accuracy_score
from tsml_eval.evaluation.storage import load_clusterer_results
from tsml_eval.experiments import (
    experiments,
    get_clusterer_by_name,
    run_clustering_experiment,
)

X, y = load_minimal_chinatown()

# set_clusterer can be used to find various clusterers by string, but
# any aeon, tsml or sklearn clusterer can be used in the experiments function
clusterer = get_clusterer_by_name("DummyClusterer")

# record memory usage every 0.1 seconds, just here for notebook speed
# does not need to be changed for usage
experiments.MEMRECORD_INTERVAL = 0.1

run_clustering_experiment(
    X,
    y,
    clusterer,
    "./generated_results/",
    dataset_name="Chinatown",
    resample_id=0,
)

A function is also available to load the dataset as well as run an experiment, see load_and_run_clustering_experiment in tsml_eval.experiments.

Both experiment functions will output a results file in the {results_dir}/{clusterer_name}/Predictions/{dataset_name}/ directory. These files can be loaded individually, or used as a collection in the evaluation module. See the evaluation notebook for more details.

clr = load_clusterer_results(
    "./generated_results/DummyClusterer/Predictions/Chinatown/trainResample0.csv"
)
print(clr.predictions)
print(clr.clustering_accuracy)
print(clr.rand_index)
print(clr.mutual_information)

[0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.
 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.]
0.5
0.48717948717948717
0.0

A function is also available to load the dataset as well as run an experiment, see load_and_run_clustering_experiment in tsml_eval.experiments.

A common use-case is wanting to compare a new algorithm against provided benchmark results. The tsml group stores their publication results and provides an aeon function to load them. An example of this is shown below for clustering.

clu = DummyClusterer()
datasets = ["ArrowHead", "GunPoint", "ItalyPowerDemand", "Chinatown", "BeetleFly"]

# find accuracy for each of our datasets on our estimator
results = {}
for d in datasets:
    X, y = load_classification(d)
    y = preprocessing.LabelEncoder().fit_transform(y)
    clu.fit(X)
    y_pred = clu.predict(X)
    results[d] = clustering_accuracy_score(y, y_pred)

results

{'ArrowHead': 0.38388625592417064,
 'GunPoint': 0.5,
 'ItalyPowerDemand': 0.5009124087591241,
 'Chinatown': 0.7134986225895317,
 'BeetleFly': 0.5}

benchmarks = ["kmeans-dtw", "kmedoids-dtw", "kmedoids-msm"]
res = get_estimator_results(
    datasets=datasets,
    estimators=benchmarks,
    task="clustering",
    measure="clacc",
)
res

{'kmeans-dtw': {'ArrowHead': 0.5257142857142857,
  'GunPoint': 0.5133333333333333,
  'ItalyPowerDemand': 0.5014577259475219,
  'Chinatown': 0.6880466472303207,
  'BeetleFly': 0.5},
 'kmedoids-dtw': {'ArrowHead': 0.52,
  'GunPoint': 0.5133333333333333,
  'ItalyPowerDemand': 0.5335276967930029,
  'Chinatown': 0.6822157434402333,
  'BeetleFly': 0.6},
 'kmedoids-msm': {'ArrowHead': 0.5314285714285715,
  'GunPoint': 0.52,
  'ItalyPowerDemand': 0.5189504373177842,
  'Chinatown': 0.6909620991253644,
  'BeetleFly': 0.85}}

res["Dummy"] = results
table = pd.DataFrame(res)
table

	kmeans-dtw	kmedoids-dtw	kmedoids-msm	Dummy
ArrowHead	0.525714	0.520000	0.531429	0.383886
GunPoint	0.513333	0.513333	0.520000	0.500000
ItalyPowerDemand	0.501458	0.533528	0.518950	0.500912
Chinatown	0.688047	0.682216	0.690962	0.713499
BeetleFly	0.500000	0.600000	0.850000	0.500000

plt, _ = plot_critical_difference(np.array(table), list(table.columns))
plt.show()

---

Generated using nbsphinx. The Jupyter notebook can be found here.