Research On New Elastic Network Algorithm For Cluster Analysis

Cluster analysis is one of the important method for humans to find hidden information from massive information.It's also occupies an important position in data mining technology.With the development of information technology,the amount of data tends to be large and high-dimensional,the structure of data tends to be complicated and diversified.Because of that,the cluster analysis technology is facing more and more problems and challenges.Traditional cluster analysis techniques can be divided into five types: hierarchical clustering,partition-based clustering,grid-based clustering,density-based clustering,and model-based clustering.After years of development,scholars proposed many other algorithm,such as entropy-based clustering,spectral clustering,uncertain clustering.However,most cluster analysis algorithms lack universality.When processing data sets with complex and diverse data structures,there are often encounter local minimum problems.In recent years,with the rise of neural network technology,scholars have found that this technology has a strong ability to deal with uncertain information.Neural network is very robust.Among them,the elastic net algorithm that belongs to unsupervised learning has good geometric properties.It's can be solved for specific objective function,which is consistent with the definition of the clustering problem.So the study for clustering is basied on elastic network.The main work content and results of this article are as follows:1.The original elastic net algorithm is only used to solve the TSP problem.Therefor,the elastic net of clustering based on maximum entropy,ENCM,is proposed at first.Change the objective function of the elastic network algorithm according the definition of clustering.Use the maximum entropy to determine the probability distribution of data set which without prior knowledge.Simulate physical systems.Under the framework of elastic net,use deterministic annealing technology to control the network activity.Use the steepest descent method to track the minimum value.2.Characteristics of data sets such as number,dimensions,etc.have a great influence on the clustering process.To reduce the interference of noise,a clustering algorithm,called weighting of elastic network clustering algorithm(WENC),is proposed.According to the purpose of clustering,following the principle of less significance less weighting to design a weight calculation method.According to the characteristics of elastic net,design an weighting method.This method optimize the clustering quality through reduce the interference of noise.Applied the proposed algorithm to the random data set and the UCI real data set for experimental analysis.The experiments show that both algorithms can solve the clustering problem well,the WENC is superior to ENCM from running time and clustering quality.WENC does not require manual guidance training.This algorithm can self-learn to solve clustering and gets high-quality results.Experimental results on both synthetic data sets and UCI real data sets have shown that WENC algorithm improves the quality of clustering.