Spatial Adaptive Agglomerative Aggregation (SA3) clustering

[u3ks]

Hi all,

We implemented an algorithm to delineate contiguous areas within cities that have identical characteristics and configurations of buildings and streets, but we thought it might be useful for other applications since the procedure is quite generic.

The idea is that is a kind of spatially restricted HDBSCAN, so there is only one parameter to specify - the minimum number of observations to form a cluster. The procedure basically consists of two steps: first, carrying out a full spatially, restricted sklearn.cluster.AgglometariveClustering clustering; and second, extracting clusters from the resulting linkage matrix, using density-clustering extraction algorithms - Excess of Mass or Leaf. This results in multiscale (clusters have varying ranges of internal similarity), contiguous clusters with noise (some observations are not attached to any clusters).

I try to explain more how it works, examples and advantages and disadvantages in the sa3.ipynb notebook.