时序聚类

方法使用参见官方文档：
https://tslearn.readthedocs.io/en/latest/gen_modules/tslearn.clustering.html#module-tslearn.clustering

from tslearn.clustering import GlobalAlignmentKernelKMeans, TimeSeriesKMeans, KShape
from tslearn.metrics import sigma_gak
from tslearn.preprocessing import to_time_series_dataset
from tslearn.generators import random_walks

def cluster_method(x, method, max_iter, n_cluster=3, seed=100):
    if method == 'KShape':
        x = TimeSeriesScalerMeanVariance(mu=0., std=1.).fit_transform(x)  #数据标准化
        model = KShape(n_clusters=n_cluster, max_iter=max_iter, n_init=1, random_state=seed).fit(x)
    elif method == 'KMeans_euclidean':
        model = TimeSeriesKMeans(n_clusters=n_cluster, metric="euclidean", max_iter=max_iter,
                      random_state=seed).fit(x)
    elif method == 'KMeans_dtw':
        model = TimeSeriesKMeans(n_clusters=n_cluster, metric="dtw", max_iter=max_iter, 
                                 max_iter_barycenter=100,random_state=seed).fit(x)
    elif method == 'KMeans_softdtw':
        model = TimeSeriesKMeans(n_clusters=n_cluster, metric="softdtw", max_iter=max_iter, 
                                 max_iter_barycenter=100,metric_params={"gamma": .5}, random_state=seed).fit(x)
    elif method == 'KernelKMeans':
        model = GlobalAlignmentKernelKMeans(n_clusters=n_cluster,
                                     sigma=sigma_gak(input_data),
                                     n_init=20,
                                     verbose=False,
                                     random_state=seed).fit(x)        
    return model

def input_data_process(method):
    if method in ('KMeans_euclidean','KShape'):  # 要求时序等长
        x = random_walks(n_ts=50, sz=32, d=1)
    else:  # 其他方法序列可不等长
        x = to_time_series_dataset([[1, 2, 3, 4],[1, 2, 3],[2, 5, 6, 7, 8, 9]])  # to_time_series_dataset可将list转换成时序聚类模型需要的输入格式
    return x

if __name__ == '__main__':
    method = 'KMeans_euclidean'
    input_data = input_data_process(method=method)
    model = cluster_method(x=input_data, method=method, n_cluster=2, max_iter=100, seed=100)
    pred = model.predict(input_data)
    pred