如何执行经纬度数据点的聚类

Question

My preferred algorithm is DBSCAN in scikit-learn. I am not sure however if (and how) to incorporate the radius in addition to latitude and longitude that I use already. My second question in how to compute the centers of the new clusters. Any ideas?

Answer 1

DBSCAN uses an epsilon radius query. This is where you use latitude and longitue.

I don't know if scikit-learn allows you to use arbitrary distances though. I've seen a blog post on using OPTICS (the successor of DBSCAN) to cluster 23 million tweets using latitude and longitude; but it used ELKI not scikit-learn.

DBSCAN doesn't use centroids. So you don't need to compute them on the sphere at all. In fact, centroids do not make sense for DBSCAN . They may be outside the cluster, if it is not convex; and DBSCAN can find non-convex clusters. Consider a city with a lake in the center. The centroid may be right in the lake. Or a city in a bay - the centroid will be inside the bay then. The centroid of the bay area (san francisco, Oakland, ...) probably is near treasure island... etc.

Answer 2

one way to calculate centroid,is to make a sum of each each cluster's longitudes(sum of each point's longitude) and calculate the mean value,this will give you a rough longitude for your centroid.Do the same thing for latitude. Here is a good example that was fair in my point of viw