时间序列的典型特征 (catch22) 转换

catch22[1] 是从 hctsa [2][3] 工具箱中存在的 7000 多个时间序列特征中提取的 22 个特征的集合。对性能优于随机水平的特征相关矩阵进行了层次聚类，以消除冗余。这些聚类使用决策树分类器按平衡准确率排序，并从形成的 22 个聚类中选择一个特征，同时考虑到平衡准确率结果、计算效率和可解释性。

在本笔记本中，我们将演示如何在 ItalyPowerDemand 单变量和 BasicMotions 多变量数据集上使用 catch22 转换器。我们还将展示 catch22 与随机森林分类器一起用于分类。

参考文献:

[1] Lubba, C. H., Sethi, S. S., Knaute, P., Schultz, S. R., Fulcher, B. D., & Jones, N. S. (2019). catch22: CAnonical Time-series CHaracteristics. 数据挖掘与知识发现, 33(6), 1821-1852.

[2] Fulcher, B. D., & Jones, N. S. (2017). hctsa: 用于使用大规模特征提取进行自动化时间序列表型分析的计算框架. Cell systems, 5(5), 527-531.

[3] Fulcher, B. D., Little, M. A., & Jones, N. S. (2013). 高度比较时间序列分析：时间序列及其方法的经验结构. Journal of the Royal Society Interface, 10(83), 20130048.

1. 导入

from sklearn import metrics

from sktime.classification.feature_based import Catch22Classifier
from sktime.datasets import load_basic_motions, load_italy_power_demand
from sktime.transformations.panel.catch22 import Catch22

2. 加载数据

IPD_X_train, IPD_y_train = load_italy_power_demand(split="train", return_X_y=True)
IPD_X_test, IPD_y_test = load_italy_power_demand(split="test", return_X_y=True)
IPD_X_test = IPD_X_test[:50]
IPD_y_test = IPD_y_test[:50]

print(IPD_X_train.shape, IPD_y_train.shape, IPD_X_test.shape, IPD_y_test.shape)

BM_X_train, BM_y_train = load_basic_motions(split="train", return_X_y=True)
BM_X_test, BM_y_test = load_basic_motions(split="test", return_X_y=True)

print(BM_X_train.shape, BM_y_train.shape, BM_X_test.shape, BM_y_test.shape)

(67, 1) (67,) (50, 1) (50,)
(40, 6) (40,) (40, 6) (40,)

3. catch22 转换

单变量

catch22 特征以转换器 Catch22 的形式提供。通过它可以将转换后的数据用于各种时间序列分析任务。

c22_uv = Catch22()
c22_uv.fit(IPD_X_train, IPD_y_train)

Catch22()

请重新运行此单元格以显示 HTML 表示或信任笔记本。

transformed_data_uv = c22_uv.transform(IPD_X_train)
transformed_data_uv.head()

/opt/homebrew/Caskroom/miniforge/base/envs/sktime/lib/python3.9/site-packages/numba/cpython/hashing.py:482: UserWarning: FNV hashing is not implemented in Numba. See PEP 456 https://pythonlang.cn/dev/peps/pep-0456/ for rationale over not using FNV. Numba will continue to work, but hashes for built in types will be computed using siphash24. This will permit e.g. dictionaries to continue to behave as expected, however anything relying on the value of the hash opposed to hash as a derived property is likely to not work as expected.
  warnings.warn(msg)

	0	1	2	3	4	5	6	7	8	9	...	12	13	14	15	16	17	18	19	20	21
0	1.158630	-0.217227	8.0	0.291667	-0.625000	3.0	6.0	0.468052	0.589049	0.836755	...	3.0	1.000000	5.0	1.778748	0.750000	0.240598	NaN	NaN	0.040000	NaN
1	0.918162	-0.214762	15.0	0.208333	-0.666667	4.0	8.0	0.702775	0.196350	0.666160	...	4.0	0.869565	5.0	1.730238	0.500000	0.388217	NaN	NaN	0.111111	NaN
2	-0.273180	-0.085856	4.0	0.875000	0.250000	2.0	5.0	0.310567	0.589049	0.865073	...	2.0	0.913043	5.0	1.836012	0.666667	0.089104	NaN	NaN	0.034014	NaN
3	0.048411	-0.450080	13.0	0.166667	-0.625000	4.0	10.0	0.804047	0.196350	0.648309	...	4.0	0.869565	6.0	1.605420	0.666667	0.332436	NaN	NaN	0.111111	NaN
4	0.426379	0.572566	16.0	0.291667	-0.666667	4.0	7.0	0.675485	0.196350	0.657946	...	4.0	0.913043	6.0	1.730238	0.500000	0.318405	NaN	NaN	0.111111	NaN

5 行 × 22 列

请注意，Catch22 在 fit(x, y=None) 方法中不考虑标签 (y)，因此我们可以轻松地用单步 fit_transform 方法替换它。

c22_uv_single_step = Catch22()
transformed_data_uv_single_step = c22_uv.fit_transform(IPD_X_train)
transformed_data_uv_single_step.equals(transformed_data_uv)

True

多变量

Catch22 支持对多变量数据进行转换。默认过程将在转换之前连接每一列。

c22_mv = Catch22()
transformed_data_mv = c22_mv.fit_transform(BM_X_train)
transformed_data_mv.head()

	dim_0__0	dim_0__1	dim_0__2	dim_0__3	dim_0__4	dim_0__5	dim_0__6	dim_0__7	dim_0__8	dim_0__9	...	dim_5__12	dim_5__13	dim_5__14	dim_5__15	dim_5__16	dim_5__17	dim_5__18	dim_5__19	dim_5__20	dim_5__21
0	-0.140988	-0.268073	6.0	-0.890	0.160	2.0	3.0	0.042638	0.736311	0.314500	...	2.0	0.707071	7.0	1.907929	1.00	0.658286	0.828571	0.228571	0.012550	9.0
1	-0.387256	-0.126246	6.0	-0.920	-0.600	2.0	4.0	0.269591	0.490874	0.614552	...	2.0	0.727273	6.0	1.875354	0.50	0.206944	0.600000	0.257143	0.028935	9.0
2	0.028412	-0.224988	9.0	-0.335	-0.045	1.0	3.0	0.036650	1.030835	0.352408	...	2.0	0.818182	7.0	1.789838	0.75	0.791912	0.828571	0.228571	0.054977	11.0
3	-0.147338	-0.199523	8.0	-0.540	0.180	1.0	5.0	0.013833	1.030835	0.212988	...	2.0	0.717172	6.0	1.904917	1.00	1.191592	0.600000	0.171429	0.015611	9.0
4	-0.217645	-0.252015	7.0	-0.130	0.020	1.0	6.0	0.008072	0.883573	0.150597	...	2.0	0.707071	7.0	1.880930	1.00	3.141568	0.800000	0.200000	0.002449	10.0

5 行 × 132 列

我们也可以设置特定的列名，例如 "short_str_feat"，它将在列名中显示特征的短名称。

如果原始时间序列分布的位置和范围可能很重要，请将 catch24 = true 设置为包含额外的 Mean 和 StandardDeviation 值。

c24_mv = Catch22(col_names="short_str_feat", catch24=True)
c24_mv.fit(BM_X_train)

Catch22(catch24=True, col_names='short_str_feat')

请重新运行此单元格以显示 HTML 表示或信任笔记本。

c24_mv.transform(BM_X_train).head()

	dim_0__mode_5	dim_0__mode_10	dim_0__stretch_decreasing	dim_0__outlier_timing_pos	dim_0__outlier_timing_neg	dim_0__acf_timescale	dim_0__acf_first_min	dim_0__centroid_freq	dim_0__low_freq_power	dim_0__forecast_error	...	dim_5__stretch_high	dim_5__rs_range	dim_5__whiten_timescale	dim_5__embedding_dist	dim_5__dfa	dim_5__rs_range	dim_5__transition_matrix	dim_5__periodicity	dim_5__mean	dim_5__std
0	-0.140988	-0.268073	6.0	-0.890	0.160	2.0	3.0	0.042638	0.736311	0.314500	...	7.0	1.907929	1.00	0.658286	0.828571	0.228571	0.012550	9.0	0.054413	0.510274
1	-0.387256	-0.126246	6.0	-0.920	-0.600	2.0	4.0	0.269591	0.490874	0.614552	...	6.0	1.875354	0.50	0.206944	0.600000	0.257143	0.028935	9.0	-0.102407	0.661172
2	0.028412	-0.224988	9.0	-0.335	-0.045	1.0	3.0	0.036650	1.030835	0.352408	...	7.0	1.789838	0.75	0.791912	0.828571	0.228571	0.054977	11.0	0.031881	0.499788
3	-0.147338	-0.199523	8.0	-0.540	0.180	1.0	5.0	0.013833	1.030835	0.212988	...	6.0	1.904917	1.00	1.191592	0.600000	0.171429	0.015611	9.0	0.029537	0.248161
4	-0.217645	-0.252015	7.0	-0.130	0.020	1.0	6.0	0.008072	0.883573	0.150597	...	7.0	1.880930	1.00	3.141568	0.800000	0.200000	0.002449	10.0	0.013344	0.163754

5 行 × 144 列

4. catch22 森林分类器

对于分类任务，与 catch22 特征一起使用的默认分类器是随机森林分类器。为了便于使用，提供了基于 catch22 特征构建的、利用 sklearn 中的 RandomForestClassifier 的 Catch22Classifier 实现。

c22f = Catch22Classifier(random_state=0)
c22f.fit(IPD_X_train, IPD_y_train)

Catch22Classifier(random_state=0)

请重新运行此单元格以显示 HTML 表示或信任笔记本。

c22f_preds = c22f.predict(IPD_X_test)
print("C22F Accuracy: " + str(metrics.accuracy_score(IPD_y_test, c22f_preds)))

C22F Accuracy: 0.86

---

使用 nbsphinx 生成。Jupyter 笔记本可在此处找到。