Networking Self-Organising Maps And Similarity Weight Associations

Using a Self-Organising Map (SOM), the structure of a data set can be explored when analysing patterns between data that are multivariate, nonlinear and unlabelled in nature. As a SOM alone cannot be used to explore patterns between different data sets, a similarity weighting scheme was previously introduced to associate different SOMs in a network fashion and approximate output patterns for given inputs. This approach uses a global weight association method on the combination of all SOMs specified for a network. However, there has been a difficulty in defining the association when changing the SOM network structure. Furthermore, it has always produced the same output weight distribution for different input data that have the same best matching unit. In an attempt to overcome the issues, we propose a new approach in this paper for locally associating a pair of SOMs as a basic network building block and approximating individually associated weight distribution. The experiments using ecological data demonstrate that the proposed approach effectively associates a pair of input and output SOMs for structural flexibility of the SOM network with better approximation of output weight distributions for individual input data.