LG: A clustering framework supported by point proximity relations
作者：
Highlights：
• We proposed an efficient and simple clustering analysis framework that consists of two stages, i.e., the clustering center locating and the label assigning stage, and both of them are closely related to the proximity relation matrix, which alleviates the problem of poor clustering results, e. g., non-global optimum resulting from random initialization of centers and low accuracy for non-hyperspherical distributed data sets caused by the inadequate utilization of the data points’ relationship.
• The nuclear model is introduced in the LEGO strategy to efficiently determine the cluster centers, which strengthens the physical meaning of the algorithms. Besides, the LEGO is insensitive to the initialization of the cluster centers.
• The GPA strategy innovatively adopts the idea that the cluster center actively selects data points as the same cluster, and it is the opposite of the traditional method in which the cluster centers passively accept data points. Therefore, GPA can improve the compactness of the clustering results. Additionally, the GPA strategy is effective to non-spherically distributed data, since the bone chain formed by the guide points can reveal and approximate the structure with arbitrary shape.
摘要
•We proposed an efficient and simple clustering analysis framework that consists of two stages, i.e., the clustering center locating and the label assigning stage, and both of them are closely related to the proximity relation matrix, which alleviates the problem of poor clustering results, e. g., non-global optimum resulting from random initialization of centers and low accuracy for non-hyperspherical distributed data sets caused by the inadequate utilization of the data points’ relationship.•The nuclear model is introduced in the LEGO strategy to efficiently determine the cluster centers, which strengthens the physical meaning of the algorithms. Besides, the LEGO is insensitive to the initialization of the cluster centers.•The GPA strategy innovatively adopts the idea that the cluster center actively selects data points as the same cluster, and it is the opposite of the traditional method in which the cluster centers passively accept data points. Therefore, GPA can improve the compactness of the clustering results. Additionally, the GPA strategy is effective to non-spherically distributed data, since the bone chain formed by the guide points can reveal and approximate the structure with arbitrary shape.
论文关键词：Clustering,Proximity relation,Local energy,Guide point,Face clustering
论文评审过程：Received 24 January 2019, Revised 29 January 2020, Accepted 8 February 2020, Available online 15 February 2020, Version of Record 20 February 2020.
论文官网地址：https://doi.org/10.1016/j.patcog.2020.107265
