One-Stage Incomplete Multi-view Clustering via Late Fusion

Abstract

Incomplete multi-view clustering (IMVC) is an unsupervised approach, among which IMVC via contrastive learning has received attention due to its excellent performance. The previous methods have the following problems: 1) Overreliance on additional projection heads when solving the dimensional collapse problem in which latent features are only valid in lower-dimensional subspaces during clustering. However, many parameters in the projection heads are unnecessary. 2) The recovered view contain inconsistent private information and useless private information will mislead the learning of common semantics due to consistent learning and reconstruction learning on the same feature. To address the above issues, we propose a novel incomplete multi-view contrastive clustering framework. This framework directly optimizes the latent feature subspace, utilizes the learned feature vectors and their subvectors for reconstruction learning and consistency learning, thereby effectively avoiding dimensional collapse without relying on projection heads. Since reconstruction loss and contrastive loss are performed on different features, the adverse effect of useless private information is reduced. For the incomplete data, the missing information is recovered by the cross-view prediction mechanism and the inconsistent information from different views is discarded by the minimum conditional entropy to further avoid the influence of private information. Extensive experimental results of the method on 5 public datasets show that the method achieves state-of-the-art clustering results.

Bi-Nuclear Tensor Schatten-p Norm Minimization for Multi-View Subspace Clustering

“…Tensors U and V are with dimensions as N ×k×M , where k is the upper bounding of the rank(Z). The proposed BTMSC model is shown as follows: (1) , Z (2)…”

“…We introduce auxiliary variables M and N into Eq. ( 10) to make variables U and V separable, leading to the following optimization model: (1) , Z (2)…”

“…D UE to the popularity of multi-view data and the difficulty of acquiring label information, multi-view clustering has been widely used in unsupervised knowledge discovery to automatically find underlying correlations [1], [2]. For example, one person may be identified by several different features, such as face poses, pedestrian walking postures and palmprint images.…”

“…Fortunately, the factorization form of the nonconvex Schatten-p norm is proved to be tractable [18]. (2) the high computational cost of each iteration has greatly limited their practical applicability since each iteration involves singular value decomposition (SVD) of large-scale matrices. To improve efficiency, Sun and Wang et al [19], [20] applied anchor to represent the actual latent data distribution, which is essentially a matrix factorization-based method.…”

Anchor Pseudo-Supervise Large-Scale Incomplete Multi-View Clustering

“…For the second group, instead of sharing a consensus representation, kernel-based IMVC methods [27]- [31] aim to learn a consensus kernel or consensus partition from incomplete views. This is achieved by uncovering the nonlinear information present in incomplete views.…”