Memory-efficient clustering
Large datasets that cannot be conveniently retained as individual points during clustering.
Scikit-Learn · Class
Birch
Birch is a memory-efficient, online clustering estimator that incrementally builds a Clustering Feature Tree. It can produce final clusters or compress data into subclusters for a later clustering algorithm.
- Since
- 0.16.1
How it works
Birch incrementally compresses samples into a Clustering Feature Tree. It merges each incoming sample into the nearest eligible subcluster, updates that subcluster’s summary statistics, and splits nodes when their subcluster count exceeds the branching factor; leaf subclusters can then be used directly or passed to a final clustering estimator.
Syntax
class sklearn.cluster.Birch(*, threshold=0.5, branching_factor=50, n_clusters=3, compute_labels=True)Parameters
| Name | Type | Description |
|---|---|---|
threshold | float | Radius limit for the subcluster formed by merging a sample with its closest subcluster; default 0.5. |
branching_factor | int | Maximum number of CF subclusters per node; default 50. |
n_clusters | int, clustering estimator, or None | Number of clusters for the final step, a clustering estimator, or None; default 3. |
compute_labels | bool | Whether to compute labels during fitting; default True. |
Returns
A fitted Birch estimator; fit_predict returns cluster labels, while transform returns distances to cluster centroids.
Key points
- Birch is an unsupervised clustering estimator implementing the BIRCH algorithm.
- The estimator supports incremental, memory-conscious learning and is positioned as an alternative to MiniBatchKMeans.
- Its tree stores compressed subcluster information, with leaf centroids serving as representative outputs.
- branching_factor limits how many Clustering Feature subclusters a node may contain before splitting.
- A new sample joins the closest subcluster only when the resulting radius remains below threshold.
- fit constructs the Clustering Feature Tree from the supplied input matrix.
Common mistakes
- BIRCH is a poor choice for high-dimensional inputs; the guide recommends MiniBatchKMeans when the feature count exceeds about twenty.
- A small threshold causes more subcluster splits, while a larger threshold permits broader subclusters.
- With n_clusters=None, Birch stops after producing leaf subclusters rather than applying a final global clustering step.
Example: Cluster synthetic data with Birch
from sklearn.cluster import Birch
from sklearn.datasets import make_blobs
X, _ = make_blobs(n_samples=300, centers=3, random_state=0)
model = Birch(n_clusters=3)
labels = model.fit_predict(X)
print(labels.shape)
This fits a BIRCH clustering model to synthetic data and returns one cluster label per sample.
When to use it
Online learning
Streaming or chunked input processed through repeated partial_fit calls.
Clustering preprocessing
Reducing many observations to representative leaf subclusters before another clustering method.
Scalable low-dimensional clustering
Datasets with relatively few features where a scalable alternative to MiniBatchKMeans is suitable.
Interview questions
QWhat does Birch construct, and why is it useful?
Birch builds a memory-efficient Clustering Feature Tree rather than retaining every input point. Its leaf subclusters provide compressed representatives that can be used as final clusters or supplied to another clustering algorithm.
Follow-up: How would you process data in chunks and choose the final cluster count? Initialize with n_clusters=None, call partial_fit repeatedly, set n_clusters with set_params, then call partial_fit once without arguments.
QWhat are the roles of threshold and branching_factor?
threshold controls the permitted radius when a new sample is merged into its closest subcluster. branching_factor caps the number of CF subclusters in each node; exceeding it triggers a node split.
Follow-up: What practical effect does lowering threshold have? It promotes splitting and therefore generally creates more, smaller subclusters.
QHow does Birch perform its final clustering step?
n_clusters specifies the number of clusters in the final clustering step, and it may be an integer, a clustering estimator, or None. compute_labels controls whether labels are calculated during fitting.
Follow-up: When is Birch useful as preprocessing? Set n_clusters=None to reduce many samples to leaf subclusters, then cluster those representatives separately.
Related
MiniBatchKMeansAgglomerativeClusteringpartial_fitfitpredicttransform