US 12,393,644 B2
High-order correlation preserved incomplete multi-view subspace clustering method and system
Xinzhong Zhu, Jinhua (CN); Huiying Xu, Jinhua (CN); Chang Tang, Jinhua (CN); and Jianmin Zhao, Jinhua (CN)
Assigned to ZHEJIANG NORMAL UNIVERSITY, Jinhua (CN)
Appl. No. 18/288,040
Filed by ZHEJIANG NORMAL UNIVERSITY, Jinhua (CN)
PCT Filed Apr. 24, 2022, PCT No. PCT/CN2022/088792
§ 371(c)(1), (2) Date Nov. 27, 2023,
PCT Pub. No. WO2022/228348, PCT Pub. Date Nov. 3, 2022.
Claims priority of application No. 202110446987.X (CN), filed on Apr. 25, 2021.
Prior Publication US 2024/0248960 A1, Jul. 25, 2024
Int. Cl. G06F 16/00 (2019.01); G06F 17/16 (2006.01); G06F 18/2321 (2023.01)
CPC G06F 18/2321 (2023.01) [G06F 17/16 (2013.01)] 8 Claims
OG exemplary drawing
 
1. A high-order correlation preserved incomplete multi-view subspace clustering method, comprising:
S1, inputting an original data matrix, and converting the inputted original data into an observed part and an incomplete part;
S2, obtaining a plurality of affinity matrices corresponding to incomplete multi-view data according to self-representation characteristics of the original data;
S3, mining a high-order correlation between the plurality of affinity matrices by means of tensor factorization;
S4, learning a unified affinity matrix from the plurality of affinity matrices to obtain a global affinity matrix;
S5, constructing a hypergraph on the basis of the global affinity matrix, and constraining an incomplete part of the incomplete multi-view data by using a hypergraph-induced Laplacian matrix;
S6, integrating the global affinity matrix, the tensor factorization and the hypergraph-induced Laplacian matrix constraint into a unified learning framework to obtain an objective function;
S7, solving the obtained objective function by means of an alternating iterative optimization strategy to obtain a solution result; and
S8, applying spectral clustering to the global affinity matrix according to the obtained solution result to obtain a clustering result,
wherein converting the inputted original data into the observed part and the incomplete part in S1 is represented as:

OG Complex Work Unit Math
wherein, Xv∈Rdv×n represents a feature matrix of a v-th view; n represents the number of samples; dv represents the number of features in the v-th view; Xo(v)∈Rdv×nv represents a feature matrix observed in the v-th view; nv represents the number of samples observed in the v-th view; Xu(v)∈Rdv×(n-nv) represents incomplete samples in the v-th view; Po(v)∈Rnv×n and Pu(v)∈R(n-nv)×n represent that two transformation matrices map observed samples and incomplete samples intoa same matrix, and are represented as:

OG Complex Work Unit Math
wherein, Po(v)ij represents an i-th row and a j-th column of a matrix Po(v); Pu(v)ij represents an i-th row and a j-th column of a matrix pu(v); xo(v)i represents an i-th column of a matrix xo(v); xj(v) represents a j-th column of a matrix x(v); xu(v)i represents an i-th column of a matrix xu(v).