Research On Multiclass Classification Algorithm Based On Evolutionary Multitasking

As an important research topic of machine learning,multiclass classification has wide applications ranging from computer vision to bioinformatics,and a variety of multiclass classification algorithms with promising performance have been proposed.Among them,the decomposition based algorithms have shown their competitiveness,since they transform the original complex multiclass classification problem into several easily solved binary classification sub-problems.Simultaneously,multitasking learning,as a sub-field of machine learning,particularly,transfer learning uses auxiliary data or knowledge from related/similar tasks to facilitate the learning in a new task.In multitasking learning,the common information contained in these related tasks can be exploited to improve the learning efficiency and generalization performance of each task-specific model.Different from the existing decomposition based algorithms,which tackle each sub-problem independently,in this thesis,an evolutionary multitasking method is suggested for multiclass classification,where the concept of multitasking is introduced to achieve the multiclass classifier with better quality.To be specific,each binary classification sub-problem is firstly viewed as a task.Then during the evolution,the tasks with low performance(termed“not-good” tasks)are aided by some well selected “assisting” tasks through using the evolutionary multitasking learning,which ensures the useful information in “assisting” tasks can be transferred into these “not-good” tasks and help them to achieve classifiers with higher accuracy.Based on this,this thesis proposes a multiclass classification algorithm based on evolutionary multitasking assistance and an efficient evolutionary multitasking multiclass classification algorithm based on clustering and grouping from the perspective of evolutionary multitasking.The main work of this thesis is summarized as follows:(1)This thesis proposes a multiclass classification algorithm based on evolutionary multitasking assistance,as mentioned above,the proposed algorithm utilizes the idea of multitasking learning to solve the multiclass classification problem.To be specific,we adopt the one-vs-one decomposition strategy to transform the original multiclass classification problem,the basic idea of proposed algorithm is to use the binary classification sub-problems(tasks)to assist the “not-good” binary classification sub-problems(tasks),so as the useful knowledge in the “good” tasks can be transferred to the “not-good” tasks,and finally obtain multiclass classifier with higher quality.Specifically,the algorithm mainly includes four phase: 1)each task evolving independently,2)selecting the “not-good” tasks,3)selecting the “assisting” tasks for each “not-good” task,4)evolving by using the “assisting”tasks to assist the “not-good” tasks.Experiments results show that the algorithm is at the same time average accuracy value and MAUC value better results can be obtained on both indicators.(2)Note that the more classes of multiclass classification problems,the more number of binary classification sub-problems after decomposition.Therefore,it is particularly important to propose an efficient multiclass classification algorithm based evolutionary multitasking.This thesis proposes an efficient evolutionary multitasking and multiclass classification algorithm based on clustering and grouping.The main idea of the algorithm is to cluster all the decomposed binary classification sub-problems(tasks)to speed up the “assisting” task selection stage.The tasks included in the group can ensure that the useful information would be delivered to those “not-good” tasks by using the assistant strategies.At the same time,the tasks included in the group are dynamically adjusted through the suggested update strategy.This algorithm can reduce the time of “assisting” task selection and speed up the assisting evolution process.The experimental results show that the algorithm can not only greatly reduce the training time but also obtain good performance.