def hac_clustering_experiment(datasets,
save_results_path=None,
preprocess=None,
verbose=True,
return_datasets=False):
affinity_names = ['l1', 'l2', 'cosine', 'canberra', 'correlation', 'rbf']
linkage = ['complete', 'average']
if return_datasets:
updated_datasets = {}
result = np.zeros([len(datasets), len(affinity_names), len(linkage), 3])
dataset_names = list(datasets.keys())
for i in range(len(dataset_names)):
dataset_name = dataset_names[i]
if verbose:
print(f'\t\t\t {dataset_name}')
if (preprocess is not None) and (dataset_name in preprocess):
T, labels = preprocess[dataset_name](datasets[dataset_name])
T = numpy_tools.flatten_np(T)
else:
T = numpy_tools.flatten_np(datasets[dataset_name]['data'])
labels = datasets[dataset_name]['labels']
if return_datasets:
updated_datasets[dataset_name] = {'data': T, 'labels': labels}
n_clusters = len(np.unique(labels))
clustAlg = AgglomerativeClustering(n_clusters=n_clusters)
for k_affinity in range(len(affinity_names)):
current_affinity = affinity_names[k_affinity]
if verbose:
print(f'\t\t Affinity: {current_affinity}')
for k_linkage in range(len(linkage)):
current_linkage = linkage[k_linkage]
if verbose:
print(f'\t Linkage: {current_linkage}')
clustAlg.linkage = current_linkage
if ((current_affinity == 'canberra')
or (current_affinity == 'correlation')
or (current_affinity == 'rbf')):
clustAlg.affinity = 'precomputed'
if current_affinity == 'rbf':
D = pairwise_distances(T, metric='euclidean')
D = -np.exp(-D)
else:
D = pairwise_distances(T, metric=current_affinity)
pred = clustAlg.fit_predict(D)
else:
clustAlg.affinity = current_affinity
pred = clustAlg.fit_predict(T)
result[i, k_affinity, k_linkage,
0] = adjusted_rand_score(labels, pred)
result[i, k_affinity, k_linkage,
1] = adjusted_mutual_info_score(
labels, pred, average_method='arithmetic')
result[i, k_affinity, k_linkage,
2] = fowlkes_mallows_score(labels, pred)
if save_results_path is not None:
np.savez_compressed(os.path.join(save_results_path),
result=result,
affinity_names=affinity_names,
linkage=linkage,
dataset_names=dataset_names,
n_clusters=n_clusters)
if verbose:
print(f'ARI={result[i, k_affinity, k_linkage, 0]:.3f} '
f'AMI={result[i, k_affinity, k_linkage, 1]:.3f} '
f'FMI={result[i, k_affinity, k_linkage, 2]:.3f} ')
if return_datasets:
return result, affinity_names, linkage, dataset_names, n_clusters, updated_datasets
return result, affinity_names, linkage, dataset_names, n_clusters
