IncrementalDBSCAN

incdbscan is an implementation of IncrementalDBSCAN, the incremental version of the DBSCAN clustering algorithm.

IncrementalDBSCAN lets the user update the clustering by inserting or deleting data points. The algorithm yields the same result as DBSCAN but without reapplying DBSCAN to the modified data set.

Thus, IncrementalDBSCAN is ideal to use when the size of the data set to cluster is so large that applying DBSCAN to the whole data set would be costly but for the purpose of the application it is enough to update an already existing clustering by inserting or deleting some data points.

The implementation is based on the following paper. To see what's new compared to the paper, jump to Notes on the IncrementalDBSCAN paper.

Ester, Martin; Kriegel, Hans-Peter; Sander, Jörg; Wimmer, Michael; Xu, Xiaowei (1998). Incremental Clustering for Mining in a Data Warehousing Environment. In: Proceedings of the 24rd International Conference on Very Large Data Bases (VLDB 1998).

Table of Contents

• Highlights
• Installation
• Usage
• Performance

Highlights

The incdbscan package is an implementation of the IncrementalDBSCAN algorithm by Ester et al., with about 40 unit tests covering diverse cases, and with additional corrections to the original paper.

Installation

incdbscan is on PyPI, and can be installed with pip:

pip install incdbscan

The latest version of the package requires at least Python 3.11.

Usage

The algorithm is implemented in the IncrementalDBSCAN class.

There are 3 methods to use:

• insert for inserting data points into the clustering
• delete for deleting data points from the clustering
• get_cluster_labels for obtaining cluster labels

All methods take a batch of data points in the form of an array of shape (n_samples, n_features) (similar to the scikit-learn API).

from sklearn.datasets import load_iris
X = load_iris()['data']
X_1, X_2 = X[:100], X[100:]

from incdbscan import IncrementalDBSCAN
clusterer = IncrementalDBSCAN(eps=0.5, min_pts=5)

# Insert 1st batch of data points and get their labels
clusterer.insert(X_1)
labels_part1 = clusterer.get_cluster_labels(X_1)

# Insert 2nd batch and get labels of all points in a one-liner
labels_all = clusterer.insert(X_2).get_cluster_labels(X)

# Delete 1st batch and get labels for 2nd batch
clusterer.delete(X_1)
labels_part2 = clusterer.get_cluster_labels(X_2)

For a longer description of usage check out the notebook developed just for that!

Performance

Performance has two components: insertion and deletion cost. The results below are based on measurements using data sets in the 1K-10K size range.

The cost of inserting a new data point with IncrementalDBSCAN is quite small and grows slower than the cost of applying (scikit-learns's) DBSCAN to a whole data set. In other words, given that we have a data set D clustered with IncrementalDBSCAN, and we want to see which cluster would a new object P belong to, it is faster to insert P into the current IncrementalDBSCAN clustering than to apply DBSCAN to the union of D and {P}.

The cost of deleting a data point with IncrementalDBSCAN is quite small and grows slower than the cost of applying DBSCAN to the data set minus that data point. In other words, given that we have a data set D clustered with IncrementalDBSCAN, and we want to see what happens to the clustering after removing an object P from the data set, it is faster to delete P from the existing IncrementalDBSCAN clustering than to apply DBSCAN to the difference of D and {P}.

These results do not imply that it is efficient to cluster a whole data set with IncrementalDBSCAN insertions. If we measure the time to cluster a data set of a given size with DBSCAN versus IncrementalDBSCAN, IncrementalDBSCAN will be slower compared to DBSCAN. The advantage of IncrementalDBSCAN is the efficiency of making additional changes (insertions, deletions) to an already clustered data set, not clustering from scratch.

A typical performance number is that clustering 8,000 data points takes about 600-700 ms with this implementation.

See this notebook about performance for more details.