Knowledge Discovery Method Based On Fuzzy Theory And Its Application

With the rapid development of information technology,the data-driven development model based on knowledge discovery is accelerating the transformation from digital and networked to intelligent in all fields of economy and society.As a kind of natural data-knowledge convert-er,fuzzy theory can simulate human reasoning to some extent,and it can visually describe the potential useful knowledge in the complex data by its unique fuzzy rule with semantics and inter-pretability.In this paper,we firstly investigate the application of fuzzy rule-based system.The classical Wang-Mendel method in fuzzy rule-based system is improved to propose a semantic description method of multi-ethnic facial features.This method is also applied in recognizing ethnic attribute of human face.We then summarize the weakness of traditional fuzzy set theory in practical application.Aiming at clustering and classification tasks in the field of knowledge discovery,we integrate axiomatic fuzzy set theory with spectral clustering and neural network with random weights in order to develop corresponding solutions.The main topics of this paper include the following aspects.(1)A semantic description method based on improved Wang-Mendel algorithm is pro-posed to characterize the facial features of Chinese ethnic groups with larger population.We firstly utilize the face landmarking technique to extract facial feature points automatically.Ge-ometric features are defined with these detected landmarks,including distances,perimeters and areas.Then the Wang-Mendel method is improved to generate linguistic rule from facial geomet-ric feature data,which implements semantic description for multi-ethnic facial characteristics.Finally,a case study of learning ethnicity from face with proposed method is investigated in Chinese ethnic face database.The experiment results indicate that the linguistic rule base ob-tained by our method is competitive in ethnicity recognition compared with method Naive Bayes,C4.5,Decision Table,Random Forest,Adaboost and Logistic Regression in terms of accuracy and results' interpretability.(2)The methodology of spectral clustering has become one of the most widely used clus-tering technique in recent years due to its excellent capability of clustering non-convex dataset.However,the previous spectral clustering methods are very sensitive to parameter setting when constructing similarity matrix,which seriously jeopardizes the robustness of these algorithms to noise data.Moreover,in some practical application scenario,such as credit assessment or med-ical diagnosis,the knowledge discovery process requires that the model's output results should be understandable.In order to make a rational trade-off between clustering accuracy and results'interpretability,we propose a clustering method that combines spectral clustering with axiomat-ic fuzzy sets theory.This method integrates the ability of axiomatic fuzzy set theory in concept extraction and representation with the advantage of spectral clustering in taking no account of dataset's distribution shape.Compared with traditional spectral clustering algorithm and other types of clustering algorithms in 17 UCI datasets,the results show that our method can effectively identify the clusters implied in raw data,and give fuzzy description for these clusters.(3)Combining the characteristics of axiomatic fuzzy set theory and neural network with random weights,a classification method with interpretable result is proposed in this paper.This method embeds the coherent membership function based on axiomatic fuzzy set theory into the hidden nodes of neural network with random weights.In neural network with random weights,the input weights are determined randomly.Inspired by this idea,we randomly generate the relationship among features,simple concepts and complex concepts in our solution.The complex concepts in hidden nodes are built by randomly selected simple concepts via logic operation in axiomatic fuzzy set theory.In the traditional classification methods based on axiomatic fuzzy set theory,the complex concepts that are suitable to describe target class can be obtained by tuning the parameters of constrain conditions,which is very time consuming.But the output weights of neural network is utilized to evaluate the fitness of complex concepts for target category in our method.Compared with other classification methods based on neural network,our method can utilize the complex concepts in hidden nodes to construct understandable fuzzy description for target category,which overcomes the so-called "Black Box" problem existing in the models based on neural network.In the experimental analysis,the proposed method is compared with 5 neural network-based classifiers including Ensemble,EvRBFN,NNEP,LVQ and iRProp on 10 UCI datasets.The experimental results show that our method not only has better classification capability,but also gives visual description for classification results.