Data driven for feature selection based on fusion with soft computing

This paper proposed an algorithm of feature selection used in fusion of soft computing and based on the chain of data-information-cognition. The algorithm is as follow: Firstly, the weights w_ij from input layer to hidden layer are obtained when the training accuracy of BP neural network (BPNN) is got. Where i denotes the i th feature and j denotes the j th node in hidden layer of BPNN. Secondly, zeta_i = Sigma_j=1 ^h|w_ij| is worked out, where zeta_i denotes the significance of i th feature and h is the number nodes of hidden layer which is optimized by genetic algorithm. The higher the magnitude of zeta_i is, the more the corresponding feature is important. Thirdly, the features are ranked into a new set according to decreasing magnitude of zeta_i, then selects the highest k ranked features from the new set and retrains the BPNN using these selected features only. At last, the algorithm is used into extracting features based on the data of oilsk81 and oilsk83 in some oilfield of China.