TY - JOUR
T1 - Fast Clustering With Anchor Guidance
AU - Nie, Feiping
AU - Xue, Jingjing
AU - Yu, Weizhong
AU - Li, Xuelong
N1 - Publisher Copyright:
© 1979-2012 IEEE.
PY - 2024/4/1
Y1 - 2024/4/1
N2 - Clustering aims to partition a set of objects into different groups through the internal nature of these objects. Most existing methods face intractable hyper-parameter problems triggered by various regularization terms, which degenerates the applicability of models. Moreover, traditional graph clustering methods always encounter the expensive time overhead. To this end, we propose a Fast Clustering model with Anchor Guidance (FCAG). The proposed model not only avoids trivial solutions without extra regularization terms, but is also suitable to deal with large-scale problems by utilizing the prior knowledge of the bipartite graph. Moreover, the proposed FCAG can cope with out-of-sample extension problems. Three optimization methods Projected Gradient Descent (PGD) method, Iteratively Re-Weighted (IRW) algorithm and Coordinate Descent (CD) algorithm are proposed to solve FCAG. Extensive experiments verify the superiority of the optimization method CD. Besides, compared with other bipartite graph models, FCAG has the better performance with the less time cost. In addition, we prove through theory and experiment that when the learning rate of PGD tends to infinite, PGD is equivalent to IRW.
AB - Clustering aims to partition a set of objects into different groups through the internal nature of these objects. Most existing methods face intractable hyper-parameter problems triggered by various regularization terms, which degenerates the applicability of models. Moreover, traditional graph clustering methods always encounter the expensive time overhead. To this end, we propose a Fast Clustering model with Anchor Guidance (FCAG). The proposed model not only avoids trivial solutions without extra regularization terms, but is also suitable to deal with large-scale problems by utilizing the prior knowledge of the bipartite graph. Moreover, the proposed FCAG can cope with out-of-sample extension problems. Three optimization methods Projected Gradient Descent (PGD) method, Iteratively Re-Weighted (IRW) algorithm and Coordinate Descent (CD) algorithm are proposed to solve FCAG. Extensive experiments verify the superiority of the optimization method CD. Besides, compared with other bipartite graph models, FCAG has the better performance with the less time cost. In addition, we prove through theory and experiment that when the learning rate of PGD tends to infinite, PGD is equivalent to IRW.
KW - Bipartite graph
KW - fast clustering
KW - optimization
KW - trivial solution
UR - http://www.scopus.com/inward/record.url?scp=85172987925&partnerID=8YFLogxK
U2 - 10.1109/TPAMI.2023.3318603
DO - 10.1109/TPAMI.2023.3318603
M3 - 文章
C2 - 37747866
AN - SCOPUS:85172987925
SN - 0162-8828
VL - 46
SP - 1898
EP - 1912
JO - IEEE Transactions on Pattern Analysis and Machine Intelligence
JF - IEEE Transactions on Pattern Analysis and Machine Intelligence
IS - 4
ER -