基于多任务协同的粒子群聚类优化算法

Co-Evolutionary Multi-Tasking Particle Swarm Clustering Optimization Algorithm

不同聚类内部指标对不同数据集的划分评价不完全一致,为了同时对数据集使用不同指标进行聚类优化以及提高聚类性能,本文提出使用多任务的方法进行聚类。多任务学习可以将对不同指标的优化视作不同的任务,本文基于多任务协同的粒子群聚类优化算法进行聚类,算法相比单任务的聚类优化算法可以同时实现对多个聚类指标的优化,分别找到各个任务下的最优解,最后用专家知识找出最适合该数据集的聚类结果。实验测试了算法分别在人工和真实数据中的聚类优化能力,并与单任务方法对比,结果展现出性能的提升。多任务进化算法在聚类时的跨任务的种群交流可以提高算法的收敛性能,从而得到更好的聚类划分。

Different clustering internal indexes are not completely consistent in the classification and evaluation of different data sets.In order to use different indexes for clustering optimization and improve clustering performance for data sets at the same time,a method for clustering based on multi-task learning is proposed by this article.Multi-task learning can regard the optimization of different indexes as different tasks.This article is based on the co-evolutionary multi-tasking for clustering.Compared with the single-task clustering optimization algorithm,the algorithm can optimize multiple clustering indexes at once.Multi-task learning can regard the optimization of different indicators as different tasks,find the optimal solution for each task,and finally use expert knowledge to find the clustering result that is most suitable for the data set.The clustering optimization capabilities of the two algorithms in synthetic and real data sets respectively are tested by the article.The single-task method is compared with method based on multitasking.The results showed an improvement in performance.The communication across multiple populations during clustering can improve the convergence characteristics of the algorithm,and thus obtain better clustering.