Abstract:
Incomplete multi-view clustering (IMVC) has become one of the most prominent
area of research in the recent past. The objective of IMVC is to integrate a set of
pre-specified incomplete views in order to improve clustering performance. Among
various excellent solutions already proposed in literature, multiple kernel k-means
with incomplete kernels (MKKM-IK) [1] has been one of the benchmark research
works, which formulates the incomplete multi-view clustering problem as a joint optimization
problem framework whereby the imputation and clustering paradigms are
integrated effortlessly. Both the processes are performed alternately in an iterative
fashion to make used of the advantages of clustering in the subsequent imputation
process and vice-versa. However, the computationally intensive and associated storage
requirements demanded more efficient methods to be devised. These include the
incomplete multi-view clustering with late fusion and the efficient and effective way
proposed by Liu et al [2]. However, all of the above mentioned algorithms initialize
the consensus clustering matrix, considering the unified kernel as a strict convex
combination of the incomplete base kernels. This bold assumption suppresses the
selectivity and representation capability of the unified kernel.
In order to find a solution to the above problem, we propose a novel method called
Optimal Neighborhood Kernel approach towards Incomplete Multi-View Clustering
(OK-IMVC) which takes into account the representability of the unified or the optimal
kernel. The consensus clustering matrix is continually updated via kernel
k-means on the optimal neighborhood kernel, which is in turn computed based on
the clustering results at the previous iteration. Specifically, our algorithm jointly
learns a consensus clustering matrix, imputes each incomplete base matrix,learns the
optimal neighborhood kernel and optimizes the corresponding alignment matrices.
Further, we conduct comprehensive experiments to study the proposed OK-IMVC
in terms of Normalized Mutual Information (NMI) index, purity score and running
time. As indicated, our proposed method significantly and consistently outperforms
some of the state-of-the-art algorithms with much less running time and memory.
