Research On The Problems Of Classification And Prediction Based On Machine Learning Methods

With the continuous development of computer and Internet technology,large amounts of data are prevalent in different fields of daily life.People can extract valuable information from big data.With the advent of big data,most academic and industrial fields have faced new challenges.Data analysis and application efficiency of big data has become a hot issue.Machine learning,an important branch of artificial intelligence,has been at the forefront of big data research and processing.The basic premise of most machine learning algorithms is to establish an optimization model.An objective function is developed and optimized using computational or optimal methods,and the optimal model is found through training.Therefore,optimization method development holds a dominant position in the research and implementation of machine learning algorithms.In this work,BP neural networks,support vector machine(SVM),several intelligent optimization algorithms,and their associated algorithms,are introduced and applied to different research fields to analyze their feasibility and practicability.The detailed work is as follows:(1)To better balance the global search ability and the local development ability of particles,and to solve the problems of premature convergence and poor local optimization ability in the GSA,the PSO-GSA algorithm based on a time-varying inertial weight strategy(TVIW-PSO-GSA)was proposed.The algorithm is to combine the group information exchange function of the PSO with the local search function of the GSA.23 benchmark functions were selected to evaluate the search performance of the TVIW-PSO-GSA algorithm.The experimental results show that the TVIW-PSO-GSA algorithm has the highest convergence accuracy,best stability,fastest convergence speed and best performance when compared to the PSO-GSA,GSA,GA and PSO algorithms in this study.(2)To solve the SVM parameter selection problem,we designed an improved SVM method by using the TVIW-PSO-GSA algorithm to optimize the penalty parameter C and the Kernel function parameter ? of the original SVM(TVIW-PSO-GSA-SVM).We also verify the feasibility and effectiveness of the improved method in practical application,e.g.,when applied to air quality level classification predication and UCI data set classification problems,and compare our newly-developed method to other methods.The experimental results show that the TVIW-PSO-GSA-SVM method has higher accuracy than the PSO-GSA-SVM,GSASVM,GA-SVM and PSO-SVM methods.(3)We consider the data lag of traditional influenza surveillance systems.First,we consider a flu prediction model that was established by google flu trends(GFT)data.The genetic algorithm-based BP(GA-BP)neural network was applied to the influenza prediction and a non-linear influenza prediction model based on GA-BP was established.Second,by analyzing influenza-like illness(ILI)data of ten regions in the United States,we’ve found that the number of ILI outbreaks has distinct seasonality in the ten regions;Therefore,we have divided the information on influenza outbreaks from these ten regions of the United States into epidemic and non-epidemic periods.A seasonal influenza prediction model was established from this separation.Finally,by comparing the prediction results between the models,and the prediction effects of the GA-BP and original least squares(OLS)methods,we’ve found that the prediction effect of non-linear modeling,based on GA-BP,is better than a linear model in most regions.We have found that interregional interaction has a certain impact on the spread of influenza.Compared with the original unseasonal influenza prediction model,the prediction results of the seasonal influenza prediction model were more accurate,more effective,and better able to reflect the true level of influenza transmission.(4)Three influenza prediction models,based on Twitter and CDC data,were established by analyzing influenza transmission factors.An improved PSO algorithm was proposed to optimize the parameters of support vector regression(IPSO-SVR),which was applied to each influenza prediction model to predict the regional percentage ILI(%ILI).By comparing the prediction results of each model,we can found that Twitter and historical influenza data are shown to partial complement each other,that is,Twitter data ensures the accuracy of real-time prediction,and historical data can better predict the trend of future influenza.The IPSO-SVR prediction results of the Model 3 was better than the influenza prediction results of a BP neural network based on improved artificial tree algorithm(IAT-BPNN).The IPSO-SVR method is not only suitable for influenza prediction in the ten regions defined by HHS,but also provides a new method for optimizing SVR parameters.(5)The estimation of the direction of arrival(DOA)plays an important role in MEMS array signal processing.In this paper,the TVIW-PSO-GSA-BP and TVIW-PSO-GSA-MUSIC methods were proposed for the estimation of the DOA from hydrophone vectors.The estimation results of the TVIW-PSO-GSA-BP and TVIW-PSO-GSA-MUSIC methods were compared to the estimation results of other methods via a simulation experiment and lake trial in the Fenhe Second reservoir.The experiment results show that the DOA evaluation index results of the TVIW-PSO-GSA-BP model were superior to the BP,PSO-BP,and GSA-BP methods,and compared to the MUSIC,PSO-MUSIC and GSA-MUSIC methods,the TVIW-PSO-GSAMUSIC model has a higher accuracy,which verifies that the validity of the two DOA estimation methods on the DOA estimation of the MEMS vector hydrophones.