Clustering spatial data with a hybrid EM approach

Abstract

In spatial clustering, in addition to the object similarity in the normal attribute space, similarity in the spatial space needs to be considered and objects assigned to the same cluster should usually be close to one another in the spatial space. The conventional expectation maximization (EM) algorithm is not suited for spatial clustering because it does not consider spatial information. Although neighborhood EM (NEM) algorithm incorporates a spatial penalty term to the criterion function, it involves much more iterations in every E-step. In this paper, we propose a Hybrid EM (HEM) approach that combines EM and NEM. Its computational complexity for every pass is between EM and NEM. Experiments also show that its clustering quality is better than EM and comparable to NEM.