.png){kind=icon}
.png){kind=icon}
Return
Dual graph-regularized sparse robust adaptive non-negative matrix factorization
.png){kind=icon}
PRE
AI
DOI：10.1016/j.eswa.2025.127594.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
This paper introduces a sparse non-negative matrix factorization approach that integrates robust estimators with dual graph learning. This method boasts robustness and effectively segregates samples from different classes while clustering those within the same class. Specifically, a robust estimator is employed to ensure that normal samples dominate the modeling process, assigning lower weights to outliers to mitigate their influence. Recognizing the potential for traditional squared L-2-norm to amplify outliers, we adopt a q-order (1 <= q <= 2) of L-2-norm to refine error measurement, enhancing model performance. Subsequently, a similarity matrix is constructed leveraging stable adaptive spectral clustering and cosine similarity, with dual graph learning techniques, independent of unknown parameters, applied to deeply explore the local, global, and manifold structures of the data, particularly adept at handling complex nonlinear data structures. Finally, to further boost model noise resilience, computational efficiency, and interpretability, while overcoming the NP-hardness of L-0-norm solutions and the non-smoothness of L-1-norm, we incorporate a computable L-2,(p)-norm (0 <= 1) as a sparse regularization term, refining the model's construction. The algorithm designed to solve our model is not only rigorously proven in theory to ensure its convergence, but also, through experimental results in practical applications, further demonstrated its ability to not only converge but also exhibit remarkable fast convergence speed. Our approach demonstrates superior performance in experiments, offering a novel and effective tool for data analysis and processing. Experimental results reveal that our method can cluster similar samples and separate dissimilar ones, yielding a compact data distribution with strong noise resilience. The evolution process of our method on the Drivface and ORL datasets illustrates the roles and advantages of each component. Statistical analysis further underscores the merits of our approach.
Keywords:
Dual-graph regularized
L-2,L-p-norm regularization terms
Stable adaptive spectral clustering
Non-negative matrix factorization
Non-negative matrix factorization
Journal
IF:
7.5
Papers:
2.9W
Citations:
10.2W