Affinity and Penalty Jointly Constrained Spectral Clustering With All-Compatibility, Flexibility, and Robustness

Area of Science:

• Machine Learning
• Data Mining
• Computer Science

Background:

• Existing semisupervised spectral clustering methods struggle with multiple supervision types and exhibit unstable performance.
• Limitations include inability to handle diverse supervisory information and inconsistent effectiveness.

Purpose of the Study:

• To propose two novel spectral clustering frameworks: type-I affinity and penalty jointly constrained spectral clustering (TI-APJCSC) and type-II (TII-APJCSC).
• To address the drawbacks of existing semisupervised spectral clustering approaches.

Main Methods:

• Development of TI-APJCSC and TII-APJCSC frameworks utilizing normalized affinity and penalty jointly constrained strategies.
• Implementation of algorithms designed for flexibility and robustness across various semisupervised scenarios.

Main Results:

• Both TI-APJCSC and TII-APJCSC demonstrate superior effectiveness compared to existing methods.
• Algorithms exhibit full compatibility with class labels, pairwise constraints, and grouping information.
• Framework normalization allows self-adaptation to diverse semisupervised settings and shows robustness to constraint numbers and affinity parameters.

Conclusions:

• TI-APJCSC and TII-APJCSC offer substantial improvements in effectiveness, compatibility, flexibility, and robustness for semisupervised spectral clustering.
• These novel algorithms are practical for medium- and small-scale semisupervised datasets, as validated by experimental studies on synthetic and real-world data.