可能性匹配知识迁移原型聚类算法

Possibility-matching based knowledge transfer prototype clustering algorithm

针对迁移原型聚类的优化问题,本文以模糊知识匹配迁移原型聚类为基础,介绍了聚类场景中从源域到目标域的迁移学习机制,明确了源域聚类中心辅助目标域得到更好的聚类效果。但目前此类迁移机制依然面临如下的挑战:1)如何克服已有迁移原型聚类方法中不同类别间的知识强制性匹配带来的负作用。2)当源域与目标域相似度较低时,如何避免模糊强制性匹配的不合理性以及过于依赖源域知识的缺陷被放大。为此,研究了一种新的迁移原型聚类机制,即可能性匹配知识迁移原型机制,并基于此实现了2个具体的迁移聚类算法。借鉴可能性匹配的思想,该算法可以自动选择和偏重有用的源域知识,克服了源域和目标域之间的强制性匹配限制,具有较好的可调节性。研究结果表明:在不同迁移场景下模拟数据集和真实NG20groups数据集上的实验研究表明,提出的算法较已有的相关算法展现了更好的性能。

Aiming at the optimization problem of migration prototype clustering,this paper introduces a migration learning mechanism from the source domain to the target domain in the clustering scene,considering fuzzy knowledge matching migration prototype clustering,and clarifies that the source domain clustering center assists the target domain to obtain better clustering effect.However,this method still faces the following challenges:1)how to overcome the negative effect brought by knowledge matching among different classes in existing transfer prototype clustering methods.2)when the similarity between the source domain and target domain is low,how to avoid the irrationality of fuzzy mandatory matching and the magnification of the defect of overdependence on knowledge from the source domain.Therefore,a new transfer prototype clustering mechanism called possibility matching-based knowledge transfer prototype clustering algorithm is proposed,and two transfer prototype clustering algorithms are further presented.Referring to the idea of possibility matching,the proposed algorithm can automatically select and focus on useful source domain knowledge,overcome the constraint of mandatory matching between the source domain and target domain,and has better adjustability.Experimental results on synthetic datasets and real NG20 text datasets in different transfer scenarios show that the proposed algorithms outperform the existing related algorithms.