Study Of Multi-label Class Imbalance Classification Based On Extreme Learning Machine

Multi label learning investigates the case of single object related to multiple labels,while class imbalanced learning mainly studies the impact of unbalancedly distrubuted samples on the algorithm.Both of them are the hot spots in the field of machine learning.Class imbalance is the common phenomenon existing in multi-label datasets.Though a large number of multi label learning algorithms have been proposed,there are only a few researches on the intrinsic characteristics of the dataset.To address the problem above,we choose extreme learning machine as the standard classifier,learn from threshold technology and cost-sensitive learning in traditional class imbalance,study the problem of class imbalance in multi label.Two new models which respectively address multi-label imbalance problem are presented.The main contributions of this thesis include the following two aspects:1)Extreme learning machine has the advantages of fast training speed and good generalization ability,it has been made a lot of achievements in regression,clustering,binary-class and multi-class fields.However,its research in multi-label learning is relatively little,and no research has studied the class imbalance problem in multi-label learning.This thesis presents a PSO-based Multi-Label Threshold Adaptation Extreme Learning Machine(MLTA-ELM).This algorithm fully combines the advantages of extreme learning machine and threshold adaptive selection strategy in class imbalanced learning.First,a single hidden layer feed forward neural network is built as extreme learning machine,and then the multi labels are predicted preliminarily realized with this model.Finally,the particle swarm optimization algorithm is taken as the threshold adaptive selection strategy to obtain the optimal threshold combination for label prediction.Lastly,this paper conducts experiments on 12 baseline multi-label datasets to verify the feasibility and effectiveness of the proposed algorithm,The experimental results indicate that the proposed method outperforms several state-of-the-art ones.2)The label-weighted extreme learning machine pioneered the application of cost-sensitive learning ideas in multi-label learning.However,in multi-label learning,the label-weighted extreme learning machine does not fully exploit the inherent imbalance characteristics of the data set.To address this problem,this paper presents an AdaBoost-based Multi-Label Weighted Extreme Learning Machine.This algorithm is based on AdaBoost framework,combines its weights distribution with the label-weighted extreme learning machine,and makes full use of the imbalance characteristics of the multi-label data set in the weight update process,so that the new algorithm not only has the low time cost,but also makes improvement in the performance and stability.Finally,this paper conducts experiments with the other 9 algorithms on 12 multi-label datasets,it is proved that the performance of the Ada-MLW-ELM algorithm is significantly superior to other algorithms.