Research On Multi-Target Regression Method Based On Multi-Category Correlation Analysis

Traditional regression task uses a set of input variables to predict a single continuous variable,while multi-target regression aims to use a set of input variables to predict multiple continuous target variables at the same time.Multi-target regression is widely used in data mining,computer vision and medical image analysis in the era of big data.At the same time,the joint processing of handling multiple categories of potential information and input-output relationships plays a key role in the final regression prediction effect.The challenges of multi-target regression algorithm mainly include the following three aspects: 1)Each target should have a unique feature set,and build a specific input space to predict different output targets.2)The correlation among outputs should be explored and utilized,and the targets with strong correlation shares more features than the targets with weak correlation or uncorrelation.3)The correlation among inputs should be explored and utilized,and the strong correlation instances have similar target space.This paper focuses on these three challenges and proposes two multi-target regression algorithms based on multi-category correlation analysis.The main research work of this thesis is as follows:1.Multi-target regression algorithm combining multiple categories of correlation information is proposed.In order to effectively analyze and utilize the potential correlation of each category,firstly,considering the inter-target correlations and the inter-instance correlations at the same time,the convex optimization objective function is constructed,and the coefficient matrix is solved by the accelerated proximal gradient method.The coefficient matrix is used to extract the target-friendly feature by applying sparse constraints to the original feature space.Then,the centroid samples are obtained by cluster analysis of the original samples,and the distance between the original samples and the centroid samples is taken as the target-correlative feature.Finally,above features are extended to the original feature space simultaneously.2.Multi-target regression algorithm combining clustering ensemble and correlation clain is proposed.In order to effectively analyze and utilize the potential correlation of each category,firstly,K-means clustering is used to divide the instances into boxes,the instance similarity matrix is constructed considering the inter-instance correlations,and the instance similarity matrix is updated by characterizing the inter-target correlations.Then,the centroid samples are obtained by spectral clustering based on graph,and the target-specific features are constructed by using the distance measure between the original samples and the centroid samples.The clustering ensemble is used to generate more effective and robust clustering results.Finally,the target space is regarded as feature space,the local structure of the target space is simulated,and the correlation analysis is carried out through Laplace score to form chain order prediction logic.The proposed methods can effectively analyze and utilize the multi-category potential correlation.The comparative experiments are carried out on 18 multi-target regression datasets with six multi-target regression methods in recent six years.The experimental results fully show the advantages of the two methods proposed in this thesis.