Improving visualization of mixed-type data with a dynamic SOM

Self-Organizing Map (SOM) possesses an effective visualization capability for supporting analysts efficiently extract valuable information from a large amount of high-dimensional data. Growing SOMs were proposed to overcome the constraint of fixed-size map in conventional SOMs. Nevertheless, the lack of a robust solution to mixed-type data processing causes most growing SOMs to fail to appropriately manipulate numeric, ordinal and categorical values simultaneously. In this paper, we propose a Growing Mixed-type SOM (GMixSOM), combining distance hierarchy with a dynamic-structure scheme to tackle the problems occurring in growing SOMs. Experimental results indicate not only are foregoing drawbacks of growing SOMs improved but topological relationship between mixed-type data can be also revealed effectively via the proposed model.