Research On Attribute Reduction Based On Bi-directional Distance Correlation And Radial Basis Network

Along with the daily rapid increase of data quantity in current information society, people need effective data mining to extract knowledge from data. The quality of data mining is greatly decided by quality of the data. Data pretreatment, which can supply high quality data for mining algorithm, is the precondition of mining valuable knowledge. Attribute reduction is one of the hot spots of present data pretreatment research.In allusion to the limitation that most attribute reduction algorithms are only suitable for the data with classified output, this article applies the wrapper thought of heuristic attribute filtration and algorithm learning conformation , and proposes a attribute reduction based on bidirectional distance & correlation (BDDC) measure and radial basis neural network(RBNN).First, measure and rank the attributes' importance by BDDC measure, than select the attribute subsets by the BDDC rank and improved addition and subtraction combination strategy; finally the radial basis neural network is used to validate the subsets. The feasibility and the validity of the algorithm have been confirmed by the experiments on two Chinese city competitive ability databases and the urban climate database finally.This algorithm is more rational and more advanced than the input output correction reduction algorithm in following aspects: The BDDC importance measure proposed in the algorithm takes both the longitudinal I/O connection and the horizontal I/O difference into account, and designs different measure functions for classified data and approaching data; The longitudinal correction not only acknowledges that the longitudinal input variety causes output variety, but introduces the variety size and direction; The horizontal correction has considered the boundary influence of different input variety which means direct effect of single input attribute value on target output caused by the attribute inequality, This algorithm applies the partially approaching radial basis network with competed output or linear output as confirmation tool and proposes a clustering ,OLS and gradient mixed algorithm as the learning algorithm of RBNN . This method uses the improved addition and subtraction combination strategy (IASCS) to withdraw the attribute subset, under dual surveillance of BDDC importance arrangement and the network performance, which will enhance the efficiency.