Implementation:Rapidsai Cuml KMeans DBSCAN HDBSCAN Init

Knowledge Sources	cuML cuML Docs
Domains	Machine_Learning, Clustering
Last Updated	2026-02-08 00:00 GMT

Overview

Concrete tool for configuring KMeans, DBSCAN, and HDBSCAN GPU-accelerated clustering algorithms via their constructor parameters.

Description

These constructors initialize the three primary clustering estimators in cuML:

• KMeans.__init__ configures the number of clusters, initialization method (scalable-k-means++, random), convergence tolerance, and batch processing parameters.
• DBSCAN.__init__ configures the neighborhood distance epsilon, minimum samples for core points, distance metric, and memory batch budget.
• HDBSCAN.__init__ configures the minimum cluster size, cluster selection method (EOM vs leaf), KNN build algorithm, and prediction data generation.

Usage

Import and instantiate these classes to create clustering estimators. Configure hyperparameters based on the dataset size, expected cluster count, and noise characteristics.

Code Reference

KMeans.__init__

Source Location

• Repository: cuML
• File: python/cuml/cuml/cluster/kmeans.pyx
• Lines: 482-504

Signature

def __init__(
    self,
    *,
    n_clusters=8,
    max_iter=300,
    tol=1e-4,
    verbose=False,
    random_state=None,
    init='scalable-k-means++',
    n_init='auto',
    oversampling_factor=2.0,
    max_samples_per_batch=32768,
    output_type=None,
):

Import

from cuml import KMeans
# or
from cuml.cluster import KMeans

DBSCAN.__init__

Source Location

• Repository: cuML
• File: python/cuml/cuml/cluster/dbscan.pyx
• Lines: 281-299

Signature

def __init__(
    self,
    *,
    eps=0.5,
    min_samples=5,
    metric='euclidean',
    algorithm='brute',
    verbose=False,
    max_mbytes_per_batch=None,
    output_type=None,
    calc_core_sample_indices=True,
):

Import

from cuml import DBSCAN
# or
from cuml.cluster import DBSCAN

HDBSCAN.__init__

Source Location

• Repository: cuML
• File: python/cuml/cuml/cluster/hdbscan/hdbscan.pyx
• Lines: 802-836

Signature

def __init__(
    self,
    *,
    min_cluster_size=5,
    min_samples=None,
    cluster_selection_epsilon=0.0,
    max_cluster_size=0,
    metric='euclidean',
    alpha=1.0,
    p=None,
    cluster_selection_method='eom',
    allow_single_cluster=False,
    gen_min_span_tree=False,
    verbose=False,
    output_type=None,
    prediction_data=False,
    build_algo='brute_force',
    build_kwds=None,
    device_ids=None,
):

Import

from cuml import HDBSCAN
# or
from cuml.cluster import HDBSCAN

I/O Contract

KMeans Inputs

Name	Type	Required	Description
n_clusters	int	No (default 8)	Number of clusters to form.
max_iter	int	No (default 300)	Maximum Lloyd iterations.
tol	float	No (default 1e-4)	Convergence threshold on center shift.
init	str	No (default 'scalable-k-means++')	Initialization: 'scalable-k-means++', 'k-means	', 'k-means++', or 'random'.
n_init	int or str	No (default 'auto')	Number of initializations to run.
oversampling_factor	float	No (default 2.0)	Factor for scalable k-means++ oversampling.
max_samples_per_batch	int	No (default 32768)	Samples per distance computation batch.

DBSCAN Inputs

Name	Type	Required	Description
eps	float	No (default 0.5)	Maximum neighborhood distance for core point calculation.
min_samples	int	No (default 5)	Minimum number of neighbors for a core sample.
metric	str	No (default 'euclidean')	Distance metric: 'euclidean', 'cosine', or 'precomputed'.
max_mbytes_per_batch	float or None	No (default None)	Memory budget in MB per batch for distance computation. None uses all available GPU memory.

HDBSCAN Inputs

Name	Type	Required	Description
min_cluster_size	int	No (default 5)	Minimum number of samples in a cluster.
cluster_selection_method	str	No (default 'eom')	Cluster extraction: 'eom' (Excess of Mass) or 'leaf'.
build_algo	str	No (default 'brute_force')	KNN graph construction: 'brute_force' or 'nn_descent'.
prediction_data	bool	No (default False)	If True, caches data needed for approximate_predict.

Outputs

Name	Type	Description
KMeans instance	KMeans	Configured KMeans estimator ready for fitting.
DBSCAN instance	DBSCAN	Configured DBSCAN estimator ready for fitting.
HDBSCAN instance	HDBSCAN	Configured HDBSCAN estimator ready for fitting.

Usage Examples

from cuml.cluster import KMeans, DBSCAN, HDBSCAN

# KMeans for known cluster count
kmeans = KMeans(n_clusters=5, max_iter=500, init='scalable-k-means++')

# DBSCAN for density-based discovery
dbscan = DBSCAN(eps=0.3, min_samples=10, metric='euclidean')

# HDBSCAN for variable-density clusters
hdbscan = HDBSCAN(min_cluster_size=15, cluster_selection_method='eom', prediction_data=True)

Related Pages

Implements Principle

• Principle:Rapidsai_Cuml_Clustering_Algorithm_Configuration

Requires Environment

• Environment:Rapidsai_Cuml_CUDA_GPU
• Environment:Rapidsai_Cuml_Python_RAPIDS_Stack