Research On Multi-label Selective Ensemble Based On Variable Precision Neighborhood Rough Set

In the real world,each multi-label sample can belong to multiple classes at the same time.Multi-label learning is a kind of hotspot problem in machine learning.This kind of problems are common in text classification,biology,scene classification and so on.The difficulty of multi-label learning is how to deal with high-dimensional attribute space and high-dimensional label space to improve the accuracy of multi-label learning algorithms.Proposed by Pawlak in 1982,rough set is a mathematical tool to describe the incompleteness and uncertainty.It can analyze the data effectively and find the implicit knowledge,and has been applied to attribute reduction and classification learning.The variable precision neighborhood rough set is the extension of Pawlak's classic rough set,which can calculate the numerical data flexibly.By combining the accurate and diverse base learners,ensemble learning can achieve better generalization performance generally than a single base learner.However,the arithmetic speed decreases gradually and the memory space increases with the increase of the base learners.On the other hand,integrating partial base learners selectively can not only improve the arithmetic speed and save the memory space,but also further improve the generalization performance.This thesis explores attribute reduction and selective ensemble in multi-label learning based on variable precision neighborhood rough set.The research results and innovations are as follows:(1)The multi-label variable precision neighborhood rough set model is proposed,and then an attribute reduction algorithm is proposed.After that,the attribute subspaces are analyzed in detail.Different attribute subspaces can be obtained based on different accuracies and neighborhoods.The influence of accuracies and neighborhoods on attribute subspaces are discussed in detail.Different learners can be generated in different attribute subspaces.After that,the different learners are integrated and the corresponding ensemble performance is analyzed in detail.(2)Two multi-label selective ensemble algorithms are proposed based on clustering.The focus is to calculate the distance between samples in these algorithms.Based on similarity,different distance calculation methods are given.Different similarities can be used in these algorithms to get different results.Finally,the effects of selective ensemble are analyzed based on different algorithms in detail.(3)A multi-label selective ensemble algorithm is proposed based on sort.The focus is to find a suitable sort measure in this algorithm.Average Precision,Coverage,Hamming Loss,One Error,Ranking Loss can be used as the sort measures.After that,the effects of selective ensemble are analyzed based on the five sort measures in detail.