geospatial data

[gurdeep101]

is it possible to use this for geospatial data by precomputing a distance matrix for lat-long values based on haversine distance?

[antklen]

Yes, definitely it is possible. If you have precomputed distance matrix, you can make clustering with this code:

DISTANCE_THRESHOLD = 1  # set your distance threshold
model = MaxDiameterClustering(max_distance=DISTANCE_THRESHOLD, sparse_dist=False, precomputed_dist=True)
labels = model.fit_predict(dist_matrix)

Don't forget to set sparse_dist=False if your distance matrix is ordinary numpy array (not sparse matrix).

I haven't tried this clustering with geospatial data, but it looks like very suitable use case. Having maximum distance between points in cluster makes a lot of sense for geo data.

Feel free to ask if you have any questions.

[gurdeep101]

thanks. I'll try this out with the above suggestions and compare with the R library. Will let you know.

[gurdeep101]

ok so by and large we're getting results for LeaderClustering with the library that are similar to the R implementation. I did this by defining a custom function "my_hav" to calculate the haversine distance and passing it to the model parameter - see sample code below. Turns out didn't need to precompute a distance matrix.

The only request would be if it is possible to add functionality to return cluster centroids for each cluster as well. This helps especially if one needs to plot the points and draw a circle around them

`
import numpy as np
import diameter_clustering as dc
import haversine as hs

coords = [lat, long]. to_numpy()

def my_hav(c1, c2):
return hs.haversine(c1, c2)

max_rad = 1
model = dc.LeaderClustering(max_radius = max_rad, metric = my_hav, deterministic = True)
labels = model.fit_predict(coords)
`

[antklen]

It is interesting that you don't need to precompute distance matrix, I missed this opportunity :)

May be I will be able to add cluster centroids for each cluster next week.
As for now, after model was fitted it has leaders_ attribute. It contains array with 1 for points which are cluster centroids and with 0 for all other points. Unfortunately there is no information which cluster each centroid belongs to. But I think it can help with drawing centroids and circles around them.

[supriyabachal]

hi! @gurdeep101 how many points did you use while clustering and max_rad=1 is in miles or kms or radians?