The Research And Application Of GMM-SAFNN Model Combined With Knowledge

Fuzzy neural networks are broadly used data mining method in recent years.In contract to the black box model of artificial neural network,fuzzy neural network can utilize real world data and represent the model as a set of interpretable fuzzy rules,which can be very useful in making decisions on realistic problems.In this way,fuzzy neural network can be broadly in information retrieval and knowledge discovery.In this paper,we make study on the problems in the application of prediction and fuzzy rules extraction on real world tasks.Aimed at the shortages of FNN in actual problems,we put forward the method of GMM structure adaptive fuzzy neural network(GMM-SAFNN)which makes improvements to the traditional one in the following ways: Firstly,because of the accuracy of fuzzy neural network model can be influenced by initial parameters,a GMM-Fuzzy neural network method is proposed to solve this problem.By using Gaussian mixture model(GMM)in optimizing the initial parameters of membership functions instead of simply using random initial parameters,this method can improve the accuracy of prostate cancer diagnosis.Secondly,In order to solve the problem with low interpretable rules extracted by fuzzy neural network,a structure adaptive fuzzy neural network(SAFNN)method is proposed.By modifying the loss function,this method can control the combination of similar membership functions,adjust the structure of fuzzy neural networks adaptively and reduce the number of fuzzy rules in the process of model training.Moreover,this method can extract interpretable rules and guarantee the diagnosis accuracy.Thirdly,to simplify the calculation process and improve training efficiency of backpropagation algorithm,particle swarm optimization(PSO)algorithm is adopted to train the structure and parameters of the GMM-SAFNN model.Finally,the Knowledge Unit Theory is used to structure the fuzzy rules extracted from the diagnosis data and make the diagnosis model can be stored and used in medical decision support systems.The knowledge units extracted from the model can be stored and used as primary knowledge in diagnosis process.Experiment studies is conducted with the inspection data of prostate diseases provided by National Clinical Medicine Information Center to verify the accuracy and efficiency of the proposed method.The result shows the proposed method is efficient in prostate cancer diagnosis,which can provide with decision and method supports in the process of primary diagnosis and biopsy for health workers.The experiment also proved our method can be helpful in other data mining and knowledge discovery tasks.