Weighted merge context for clustering and quantizing spatial data with self-organizing neural networks

This publication presents a generalization of merge context, named weighted merge context (WMC), which is particularly useful for clustering and quantizing spatial data with self-organizing neural networks. In contrast to merge context, WMC does not depend on a predefined (sequential) ordering of the data; distance is evaluated by recursively taking neighboring observations into account. For this purpose, WMC utilizes a weight matrix that describes the neighborhood relationships between observations. This property distinguishes WMC from existing approaches like contextual neural gas (NG) or the GeoSOM, which force spatially close observations to be represented by similar prototypes, but neglected the similarity of the observations’ neighborhoods. For practical studies, WMC is combined with the NG algorithm to obtain weighted merging NG (WMNG). The properties of WMNG and its usefulness for clustering and quantizing spatial data are investigated on two different case studies which utilize an simulated binary grid and a real-world continuous data set.