Robust Dimension Reduction for Clustering With Local Adaptive Learning.

Area of Science:

• Data Mining
• Pattern Recognition
• Machine Learning

Background:

• Clustering is vital in data mining, with K-means (KM) being popular but challenged by high-dimensional data and the curse of dimensionality.
• Traditional KM struggles with redundant features and noise in high-dimensional datasets, limiting its effectiveness.
• Existing subspace learning and KM methods have limitations in capturing discriminative information, considering geometric properties, and noise robustness.

Purpose of the Study:

• To propose a novel clustering model addressing the limitations of traditional K-means in high-dimensional data.
• To enhance clustering by adaptively exploring discriminative information through local adaptive subspace learning and KM.
• To improve robustness against noise by employing a robust l2,1-norm loss function.

Main Methods:

• Unifying local adaptive subspace learning with K-means clustering to explore discriminative information.
• Extending the model with a robust l2,1-norm loss function for weighted iterative calculation of cluster centroids.
• Analyzing the relationships between the proposed algorithm and existing clustering studies.

Main Results:

• The proposed model effectively explores discriminative information in low-dimensional subspaces.
• The integration of local adaptive subspace learning enhances the capture of intrinsic geometric information.
• Experiments demonstrate the proposed model's superior performance over state-of-the-art clustering approaches on benchmark datasets.

Conclusions:

• The novel clustering model successfully addresses the challenges posed by high-dimensional data and the curse of dimensionality.
• The robust l2,1-norm loss function improves the stability and accuracy of clustering by mitigating noise effects.
• The proposed method offers a significant advancement in clustering high-dimensional data, outperforming existing techniques.