Research On Correlation-based Multi-view Learning Method And Application Of Colliery Security Situation

Multi-view learning is a method to improve the generalization performance of classifiers by using the correlation information between different feature sets.With the advent of the era of big data,data in the real world has grown tremendously.In the traditional single-view learning methods,high feature dimensions usually lead to overfitting and data redundancy.In order to solve the problems in single-view algorithms,this dissertation proposes a supervised data view construction algorithm(DVC)based on mutual information and canonical correlation analysis.On this basis,this dissertation uses the correlation information to construct the correlation view,and proposes a data view fusion algorithm(DVF)based on the correlation view.Experimental results show that the DVF algorithm can further improve the performance of the classifier.The main work of this dissertation is described as follows:1.This dissertation uses mutual information to compute the correlation information within the views and extracts the correlation information among views by canonical correlation analysis.Based on the above evaluation metrics of views,a data view construction algorithm is proposed.Firstly,calculating the sum of mutual information between features and label.Secondly,canonical correlation analysis is used to calculate the correlation coefficient between views.Finally,the subtraction of the two metrics is used as the target function of the data view construction algorithm.Experiments on eight UCI data sets and gas concentration data sets in the field of colliery security situation show that using the data view construction algorithm can improve the performance of the classifier.2.This dissertation employs the view evaluation metrics to construct correlation view and proposes a multi-view fusion algorithm based on correlation view.Firstly,after the data views are constructed,the correlation information in the view and among views are calculated for each view.Secondly,the subtraction of the above two evaluation metrics is considered as the correlation view value of this view.Finally,after each view completes the training,fusion weight is assigned to each view according to the correlation view,and the results of all views are fused in the way of weighted fusion.Experiments show that based on the data view construction algorithm,the performance of the classifier is further improved by using the data view fusion algorithm.