Improved Spectral Clustering Algorithm And Its Application In Risk-model

Spectral clustering has been one of the most popular clustering algorithms in recent years,it has been a hot area of machine learning and data mining research as well which could help us figure out the connections between things more clearly.As a new and efficient clustering analysis algorithm,spectral clustering has the advantage of dealing with data sets in any shape comparing with traditional clustering algorithms.Thus it could solve a much wider variety of problems.An important research approach of spectral clustering in the future is to use it to extract information from huge amounts of data.As the development of artificial intelligence,improving the accuracy of the algorithm and reducing the time complexity is urgent.Based on information entropy theory and the idea of Multi-View clustering,this paper improves the spectral clustering algorithm,and enriches the knowledge system of it.It provides a new way to solve various problems in clustering analysis.This thesis mainly focuses on the following aspects.First of all,different scale parameters ? have different effect on the clustering result of spectral clustering algorithm.The clustering results will also change with the alteration of the similarity metric function.How to adaptively select scale parameters is the first innovation of this article.Applying information entropy theory to the selection of scale parameters,I found that minimizing the information entropy to optimize scale parameter is a good way to present the data distribution characteristic.Improving the Normalized Cuts objective function based on the idea of Multi-view clustering and knowledge of mutual information theory is the second innovation of this thesis.According to the Multi-view clustering theory,each sample has a one-to-one correspondence with its own category and we could control the redundancy of features through mutual information.Thus it is feasible to minimize the objective function of the Normalized Cuts while minimizing the error between the sample and the class matrix as well as the redundancy between the data features.The results of this study showed that the improved spectral clustering algorithm get better clustering results.And then,I assume that AEMVFS algorithm could be used to anti-fraud strategy in consumer finance area.Constructing anti-fraud model is an important approach for the financial company to control risk.It could effectively reduce the bad debts rate and unnecessary fund loss.It reduces the operating cost as well.In this paper,I tested the improved spectral clustering algorithm with the real transaction data of internet financial companies.It showed that spectral clustering algorithm behaved well to identify ‘bad' customers which provides another perspective for corporate decision-makers.At last,I suppose the focus of future researches could be put on the robustness and interpretability of spectral clustering algorithm.We should combine theory with practical application scenarios.