High-Order Correlation Recovery in Incomplete Multi-View Clustering
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing incomplete multi-view clustering algorithms fail to effectively utilize high-order correlations between samples and views, leading to information loss and limited clustering performance due to unbalanced and incomplete data.
Innovation Solution
A high-order correlation preserved incomplete multi-view subspace clustering method that utilizes tensor factorization and hypergraph-induced Laplacian regularization to capture and integrate high-order correlations, reconstructing incomplete samples and preserving underlying subspace structures through an alternating iterative optimization strategy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If conventional multi-view clustering methods are used on incomplete multi-view data, then the clustering process can be performed, but the clustering performance is limited due to information loss from ignoring high-order correlations
Solution Approach 1:
The patent transforms the traditional pairwise correlation model into a high-order correlation model by introducing tensor factorization. This dimensional escalation from second-order to higher-order correlations enables the method to capture complex interrelationships among multiple views and samples simultaneously, thereby reducing information loss and improving clustering reliability on incomplete multi-view data.
Solution Approach 2:
The patent introduces tensor factorization as an intermediary mechanism to model and exploit high-order correlations. This intermediary approach enables the integration of information across multiple views and samples in a unified framework, allowing the method to recover missing data more effectively and improve clustering performance without directly modifying the underlying data structure.
2Reliability
If only paired sample correlation and paired view correlation are utilized, then the algorithm complexity remains manageable, but the clustering performance is limited due to unbalanced information contribution from different views
Solution Approach 1:
The patent transitions from pairwise correlation modeling to high-order correlation modeling through tensor factorization. This dimensional escalation enables the simultaneous capture of interactions among multiple views and samples, allowing the algorithm to overcome unbalanced information contribution while maintaining a structured computational framework that manages complexity through low-rank constraints.
3Measurement precision
If high-order correlation modeling is implemented through tensor factorization, then the recovery of incomplete samples and subspace structures is improved, but the computational complexity increases
Solution Approach 1:
The patent employs low-rank constraints as a parameter change strategy to control the complexity of tensor factorization. By restricting the rank of the factorized tensors, the method achieves a balance between capturing high-order correlations for precise sample recovery and maintaining computational tractability, thereby improving recovery precision without excessive computational burden.
Data Source
AI summary
A high-order correlation preserved incomplete multi-view subspace clustering method and system. The method comprises: S11, inputting an original data matrix, and converting original data into an observed part and an incomplete part; S12, obtaining a plurality of affinity matrices according to self-representation characteristics of the original data; S13, mining a high-order correlation between the plurality of affinity matrices by tensor factorization; S14, learning a unified affinity matrix from the plurality of affinity matrices, so as to obtain a global affinity matrix; S15, constructing a hypergraph on the basis of the global affinity matrix, and constraining an incomplete part of incomplete multi-view data by using a hypergraph-induced Laplacian matrix; S16, integrating the global affinity matrix, the tensor factorization and the hypergraph-induced Laplacian matrix constraint into a unified learning framework; S17, solving the obtained objective function by an alternating iterative optimization strategy; and S18, applying spectral clustering to the global affinity matrix.