Abstract:

The Self-Organizing Map is used for appraisal purposes in National Land Survey of Finland. Special problems with real estate data analysis are that the property units are unique, sales data is available maybe once in 30 years, comparable sales are not easy to find and the market prices may change moderately strong over time. The scaling of the components is used to produce varying views to the observation space for varying purposes. Subjective views are produced objectively by generating more neurons for strongly scaled components and one or two neurons for weakly scaled components in a specific class in the self-organization process. Sensitivity of components is analysed by changing one or two component values and by computing the value of third component for an object or for a group of objects (portfolio). Smoothing of the function representation is done both by using several slightly varying maps and by using several sligthly varying objects in a portfolio.