Online Sparse Representation Clustering for Evolving Data Streams.

Area of Science:

• Computer Science
• Data Mining
• Machine Learning

Background:

• Data stream clustering aims to uncover patterns in continuous data sequences.
• Existing density-based algorithms struggle with high-dimensional data, Euclidean distance limitations, and knowledge transfer between data windows.
• Adaptive exploitation of evolving subspace structures is crucial for effective data stream clustering.

Purpose of the Study:

• To propose an online sparse representation clustering (OSRC) method for high-dimensional data streams.
• To enhance pattern discovery by adaptively exploiting evolving subspace structures.
• To enable knowledge transfer across data windows for improved clustering performance.

Main Methods:

• Introduced a low-dimensional projection (LDP) into sparse representation to mitigate noise and redundancy in high-dimensional data.
• Utilized L1-norm optimization to select representative data objects and form a dictionary for sparse representation.
• Integrated the dictionary into sparse representation to adaptively exploit evolving subspace structures and facilitate knowledge transfer between landmark windows.

Main Results:

• The proposed OSRC method demonstrated effectiveness in learning an affinity matrix for high-dimensional data objects in evolving streams.
• Experimental results on synthetic and benchmark datasets showed superior performance compared to state-of-the-art data stream clustering methods.
• The method successfully addressed limitations in constructing microclusters and merging them using Euclidean distances.

Conclusions:

• The OSRC method provides an effective approach for data stream clustering, particularly for high-dimensional data.
• The integration of LDP and sparse representation with adaptive subspace structure exploitation significantly improves clustering accuracy.
• The ability to transfer knowledge across data windows enhances the robustness and adaptability of the clustering process.