Method for computer-based analysis of a data set from observations

摘要

The invention relates to a method for computer-aided analysis of a data set from observations (OB), wherein the data record for each observation (OB) contains a data vector which contains the values of a plurality of input variables (x1, x2,..., XN) and the value a target size (y). In the method according to the invention, based on the data set, a neural network structure (NNS) is learned from a plurality of differently initialized neural networks (NN1, NN2,..., NNm), the neural networks (NN1, NN2,..., NNm) each an input layer (I1, 12, ..., Im), one or more hidden layers (H1, H2, ..., Hm) and an output layer (01, 02, ..., Om), the input layer ( I1, 12, ..., Im) of a respective neural network (NN1, NN2, ..., NNm) comprises at least part of the input variables (x1, x2,..., XN) and the output layer (01, 02, ..., Om) of a respective neural network (NN1, NN2, ..., NNm) comprises the target variable (y), wherein the neural network structure (NNS) the mean value of the target quantities (y) of the output layers (01, 02,. .., Om) of the neural networks (NN1, NN2, ..., NNm). Sensitivity values (SV) are determined and stored by means of the learned neural network structure (NNS), each observation value (SV) being assigned an observation (OB) and an input variable (x1, x2,..., XN) and the respective sensitivity value (SV ) contains the derivation of the target variable (y) of the associated observation (OB) according to the assigned input variable (x1, x2, ..., xN).