Context-specific preference learning of one-dimensional quantitative geospatial attributes using a neuro-fuzzy approach

With the recent explosion of information availability in geospatial datasets, query complexity has increased. Multiple users access the same data collections with highly diversified needs. Information retrieval goals can vary significantly due to the large number of potential scenarios/applications, a common problem in geospatial data collections. The current approaches are deterministic and do not allow the incorporation of user preferences in the query process. The approach developed in this thesis adjusts query returns using a preference-based similarity modeling and therefore expresses more accurately user anticipation of results.; In this thesis we present a machine learning approach to express user preferences within one-dimensional, quantitative attributes. Training is performed in multiple stages and is based on a training dataset provided by the user. Depending on the provided preference complexity our algorithm adjusts the learning process. Several families of functions are used progressively, from simple planar to complex sigmoidal functions. The design of the algorithm allows previously interpolated functions to act as approximations for more complex ones that follow, thereby decreasing training time and increasing robustness.; A customized neural network, a Multi-Scale Radial Basis Function (MSRBF) network, is also developed specifically to express the characteristics of the problem. We model potential errors that result from the interpolation of the fuzzy functions; we do not want our neural network to expand to portions of the input space without significant evidence. Therefore, our network design forces the network to operate in a localized manner and only where necessary. At the last training stage fuzzy functions are combined with the MSRBF into one solution and if found appropriate, the fuzzy functions go through a self-organizing process, where they adjust further to the overwhelming preference.; The proposed neuro-fuzzy system outperforms the currently used distance-based nearest neighbor methods. It does so by design because it recognizes and supports distance dependent user preferences, while simultaneously offering advanced modeling capabilities. Our system also exhibits high robustness as statistical simulations demonstrate. This is partially due to the ability of the algorithm to adjust its complexity as the user preference complexity increases.