Clustering 500,000 geospatial points in python

Question

I\'m currently faced with the problem of finding a way to cluster around 500,000 latitude/longitude pairs in python. So far I\'ve tried computing a distance matrix with nump

Answer 1

Older versions of DBSCAN in scikit learn would compute a complete distance matrix.

Unfortunately, computing a distance matrix needs O(n^2) memory, and that is probably where you run out of memory.

Newer versions (which version do you use?) of scikit learn should be able to work without a distance matrix; at least when using an index. At 500.000 objects, you do want to use index acceleration, as this reduces runtime from O(n^2) to O(n log n).

I don't know how well scikit learn supports geodetic distance in its indexes though. ELKI is the only tool I know that can use R*-tree indexes for accelerating geodetic distance; making it extremely fast for this task (in particular when bulk-loading the index). You should give it a try.

Have a look at the Scikit learn indexing documentation, and try setting algorithm='ball_tree'.

Answer 2

I don't have your data so I just generated 500k random numbers into three columns.

import numpy as np
import matplotlib.pyplot as plt
from scipy.cluster.vq import kmeans2, whiten

arr = np.random.randn(500000*3).reshape((500000, 3))
x, y = kmeans2(whiten(arr), 7, iter = 20)  #<--- I randomly picked 7 clusters
plt.scatter(arr[:,0], arr[:,1], c=y, alpha=0.33333);

out[1]:

I timed this and it took 1.96 seconds to run this Kmeans2 so I don't think it has to do with the size of your data. Put your data in a 500000 x 3 numpy array and try kmeans2.