我有一个UTM坐标的地理数据集'coordinates'，我正在对其执行HDBSCAN，并希望使用sklearn的GridSearchCV验证不同的参数，使用DBCV。当我手动评估HDBSCAN的参数时，我得到了以下结果，这比sklearn的GridSearchCV要好：
clusters = hdbscan.HDBSCAN(min_cluster_size=75, min_samples=60, 
                           cluster_selection_method='eom', gen_min_span_tree=True, 
                           prediction_data=True).fit(coordinates)
获得的DBCV分数：0.2580606238793024
当使用sklearn的GridSearchCV时，它选择模型参数，以获得更低的DBCV值，即使手动选择的参数在参数字典中。另外，当我尝试使用RandomizedSearchCV时，使用不同范围的参数，我能够获得0.28的DBCV值，但未记录使用了哪些参数。
*更新：当我运行RandomizedSearchCV和GridSearchCV时，选择的'最佳'模型始终是参数网格的第一个条目或第一个随机样本。例如，在下面的代码中，它总是选择min_samples和min_cluster_size的第一个条目。我怀疑是因为它遇到了错误。当我添加error_score="raise"时，它会引发TypeError，这可能与它无法与y进行比较有关，但这是无监督聚类，没有数据标签。
> TypeError: _BaseScorer.__call__()缺少1个必需的位置参数：'y_true'
from sklearn.model_selection import RandomizedSearchCV
from sklearn.model_selection import GridSearchCV
import hdbscan
from sklearn.metrics import make_scorer
import logging # to further silence deprecation warnings
logging.captureWarnings(True)
### GridSearch CV模型调整 ###
logging.captureWarnings(True)
hdb = hdbscan.HDBSCAN(gen_min_span_tree=True).fit(coordinates)
# 指定要从中抽样的参数
grid = {'min_samples': [50,55,60,65,70,75,80,90,100,110],
                  'min_cluster_size':[40,45,50,55,60,65,75,80,85,90,95,100],  
                  'cluster_selection_method' : ['eom','leaf'],
                  'metric' : ['euclidean','manhattan'] 
                 }
#validity_scroer = "hdbscan__hdbscan___HDBSCAN__validity_index"
validity_scorer = make_scorer(hdbscan.validity.validity_index,greater_is_better=True)
grid_search = GridSearchCV(hdb
                           ,param_grid=grid
                           ,scoring=validity_scorer)
grid_search.fit(coordinates)
print(f"Best Parameters {grid_search.best_params_}")
print(f"DBCV score :{grid_search.best_estimator_.relative_validity_}")
>     最佳参数 {'cluster_selection_method': 'eom', 'metric': 'euclidean', 'min_cluster_size': 40, 'min_samples': 50}
>     DBCV分数：0.22213170637127946

# Naive grid search implementation by Mueller and Guido, Introduction to Machine Learning with Python
best_score = 0
for min_cluster_size in [40, 45, 120, 50, 55, 130, 140, 150, 155, 160]:
    for min_samples in [40, 45, 50, 85, 55, 60, 90, 100, 110, 115, 120]:
        for cluster_selection_method in ['eom', 'leaf']:
            for metric in ['euclidean']:
                # for each combination of parameters of hdbscan
                hdb = hdbscan.HDBSCAN(min_cluster_size=min_cluster_size, min_samples=min_samples,
                                      cluster_selection_method=cluster_selection_method, metric=metric, 
                                      gen_min_span_tree=True).fit(coordinates)
                # DBCV score
                score = hdb.relative_validity_
                # if we got a better DBCV, store it and the parameters
                if score > best_score:
                    best_score = score
                    best_parameters = {'min_cluster_size': min_cluster_size, 
                                       'min_samples': min_samples, 'cluster_selection_method': cluster_selection_method,
                                       'metric': metric}
print("Best DBCV score: {:.3f}".format(best_score))
print("Best parameters: {}".format(best_parameters))