Generalized Radial Basis Function Networks Trained with Instance Based Learning for Data Mining of Symbolic Data

摘要: The application of the Radial Basis Function neural networks in domains involving prediction and classification symbolic data requires a reconsideration careful definition concept distance between patterns. This addition to providing information about proximity patterns should also obey some mathematical criteria order be applicable. Traditional distances are inadequate access differences work proposes utilization statistically extracted measure for Generalized (GRBF) networks. main properties these retained new metric space. Especially, their regularization potential can realized with this type distance. However, examples training set applications not all same importance reliability. Therefore, construction effective decision boundaries consider numerous exceptions general motifs that frequently encountered mining applications. paper supports heuristic Instance Based Learning (IBL) approaches uncover within uneven structure set. is exploited estimation an adequate subset serving as RBF centers parameter settings those centers. IBL learning steps applicable both traditional statistical spaces improve significantly performance cases. obtained results two-level method better than nearest neighbour schemes many problems.