>>> from sklearn.decomposition import PCA
>>> from sklearn.cluster import KMeans
>>> from techminer2.factor_analysis.co_occurrence import treemap
>>> plot = treemap(
... #
... # PARAMS:
... field="author_keywords",
... association_index=None,
... #
... # ITEM PARAMS:
... top_n=20,
... occ_range=(None, None),
... gc_range=(None, None),
... custom_terms=None,
... #
... # DESOMPOSITION PARAMS:
... decomposition_estimator = PCA(
... n_components=5,
... whiten=False,
... svd_solver="auto",
... tol=0.0,
... iterated_power="auto",
... n_oversamples=10,
... power_iteration_normalizer="auto",
... random_state=0,
... ),
... #
... # CLUSTERING:
... clustering_estimator_or_dict = KMeans(
... n_clusters=6,
... init="k-means++",
... n_init=10,
... max_iter=300,
... tol=0.0001,
... algorithm="elkan",
... random_state=0,
... ),
... #
... # DATABASE PARAMS:
... root_dir="example/",
... database="main",
... year_filter=(None, None),
... cited_by_filter=(None, None),
... )
>>> # plot.write_html("sphinx/_static/factor_analysis/co_occurrence/treemap.html")