Theoretically efficient and practical parallel DBSCAN


Keywords
cluster, clustering, density, dbscan, data-clustering, data-mining, dbscan-clustering, machine-learning, parallel-algorithm, python
License
MIT
Install
pip install dbscan==0.0.9

Documentation

Theoretically-Efficient and Practical Parallel DBSCAN

arXiv build

Overview

This repository hosts fast parallel DBSCAN clustering code for low dimensional Euclidean space. The code automatically uses the available threads on a parallel shared-memory machine to speedup DBSCAN clustering. It stems from a paper presented in SIGMOD'20: Theoretically Efficient and Practical Parallel DBSCAN.

Our software on 1 thread is on par with all serial state-of-the-art DBSCAN packages, and provides additional speedup via multi-threading. Below, we show a simple benchmark comparing our code with the DBSCAN implementation of Sklearn, tested on a 6-core computer with 2-way hyperthreading using a 2-dimensional data set with 50000 data points, where both implementation uses all available threads. Our implementation is more than 32x faster. We also show a visualization of the clustering result on a smaller data set.

Data sets with dimensionality 2 - 20 are supported by default, which can be modified by modifying DBSCAN_MIN_DIMS and DBSCAN_MAX_DIMS in the source code.

timing example

Tutorial

Option 1: Use the Python binding

There are two ways to install it:

  • Install it using PyPI: pip3 install --user dbscan (you can find the wheels here).
  • To build from scratch for testing: pip3 install -e . from the project root directory.

An example for using the Python module is provided in example.py. It generates the clustering example above.

Python API

from dbscan import DBSCAN
labels, core_samples_mask = DBSCAN(X, eps=0.3, min_samples=10)

Input

  • X: A 2-D Numpy array containing the input data points. The first dimension of X is the number of data points n, and the second dimension is the data set dimensionality (the maximum supported dimensionality is 20).
  • eps: The epsilon parameter (default 0.5).
  • min_samples: The minPts parameter (default 5).

Output

  • labels: A length n Numpy array (dtype=np.int32) containing cluster IDs of the data points, in the same ordering as the input data. Noise points are given a pseudo-ID of -1.
  • core_samples_mask: A length n Numpy array (dtype=np.bool) masking the core points, in the same ordering as the input data.

We provide a complete example below that generates a toy data set, computes the DBSCAN clustering, and visualizes the result as shown in the plot above.

import numpy as np
from sklearn.datasets import make_blobs
from sklearn.preprocessing import StandardScaler

# #############################################################################
# Generate sample data
centers = [[1, 1], [-1, -1], [1, -1]]
X, labels_true = make_blobs(n_samples=750, centers=centers, cluster_std=0.4,
                            random_state=0)
X = StandardScaler().fit_transform(X)

# #############################################################################
# Compute DBSCAN

# direct call of the C API:
from dbscan import DBSCAN
labels, core_samples_mask = DBSCAN(X, eps=0.3, min_samples=10)

# OR calling our sklearn API:
# from dbscan import sklDBSCAN as DBSCAN
# db = DBSCAN(eps=0.3, min_samples=10).fit(X)
# core_samples_mask = np.zeros_like(db.labels_, dtype=bool)
# core_samples_mask[db.core_sample_indices_] = True
# labels = db.labels_

# #############################################################################
# Plot result
import matplotlib.pyplot as plt

n_clusters_ = len(set(labels)) - (1 if -1 in labels else 0)
n_noise_ = list(labels).count(-1)
# Black removed and is used for noise instead.
unique_labels = set(labels)
colors = [plt.cm.Spectral(each)
          for each in np.linspace(0, 1, len(unique_labels))]
for k, col in zip(unique_labels, colors):
    if k == -1:
        # Black used for noise.
        col = [0, 0, 0, 1]

    class_member_mask = (labels == k)

    xy = X[class_member_mask & core_samples_mask]
    plt.plot(xy[:, 0], xy[:, 1], 'o', markerfacecolor=tuple(col),
             markeredgecolor='k', markersize=14)

    xy = X[class_member_mask & ~core_samples_mask]
    plt.plot(xy[:, 0], xy[:, 1], 'o', markerfacecolor=tuple(col),
             markeredgecolor='k', markersize=6)

plt.title('Estimated number of clusters: %d' % n_clusters_)
plt.show()

Option 2: Use the binary executable

Compile and run the program:

mkdir build
cd build
cmake ..
cd executable
make -j # this will take a while
./dbscan -eps 0.1 -minpts 10 -o clusters.txt <data-file>

The <data-file> can be any CSV-like point data file, where each line contains a data point -- see an example here. The data file can be either with or without header. The cluster output clusters.txt will contain a cluster ID on each line (other than the first-line header), giving a cluster assignment in the same ordering as the input file. A noise point will have a cluster ID of -1.

Option 3: Include directly in your own C++ program

Create your own caller header and source file by instantiating the DBSCAN template function in "dbscan/algo.h".

dbscan.h:

template<int dim>
int DBSCAN(int n, double* PF, double epsilon, int minPts, bool* coreFlagOut, int* coreFlag, int* cluster);

// equivalent to
// int DBSCAN(intT n, floatT PF[n][dim], double epsilon, intT minPts, bool coreFlagOut[n], intT coreFlag[n], intT cluster[n])
// if C++ syntax was a little more flexible

template<>
int DBSCAN<3>(int n, double* PF, double epsilon, int minPts, bool* coreFlagOut, int* coreFlag, int* cluster);

dbscan.cpp:

#include "dbscan/algo.h"
#include "dbscan.h"

Calling the instantiated function:

int n = ...; // number of data points
double data[n][3] = ...; // data points
int labels[n]; // label ids get saved here
bool core_samples[n]; // a flag determining whether or not the sample is a core sample is saved here
{
  int ignore[n];
  DBSCAN<3>(n, (void*)data, 70, 100, core_samples, ignore, labels);
}

Doing this will only compile the function for the number of dimensions that you want, which saves on compilation time.

You can also include the "dbscan/capi.h" and define your own DBSCAN_MIN_DIMS and DBSCAN_MAX_DIMS macros the same way the Python extension uses it. The function exported has the following signature.

extern "C" int DBSCAN(int dim, int n, double* PF, double epsilon, int minPts, bool* coreFlag, int* cluster);

Right now, the only two files that are guaranteed to remain in the C/C++ API are "dbscan/algo.h" and "dbscan/capi.h" and the functions named DBSCAN within.

Citation

If you use our work in a publication, we would appreciate citations:

@inproceedings{wang2020theoretically,
  author = {Wang, Yiqiu and Gu, Yan and Shun, Julian},
  title = {Theoretically-Efficient and Practical Parallel DBSCAN},
  year = {2020},
  isbn = {9781450367356},
  publisher = {Association for Computing Machinery},
  address = {New York, NY, USA},
  url = {https://doi.org/10.1145/3318464.3380582},
  doi = {10.1145/3318464.3380582},
  booktitle = {Proceedings of the 2020 ACM SIGMOD International Conference on Management of Data},
  pages = {2555–2571},
  numpages = {17},
  keywords = {parallel algorithms, spatial clustering, DBScan},
  location = {Portland, OR, USA},
  series = {SIGMOD ’20}
}