Sparse Twin Support Vector Clustering Using Pinball Loss.

Area of Science:

• Machine Learning
• Data Mining
• Computational Biology

Background:

• Clustering algorithms analyze unlabelled data, with Twin Support Vector Clustering (TWSVC) being a recent advancement.
• Traditional TWSVC uses hinge loss, which is sensitive to noise and resampling, limiting its effectiveness on real-world datasets.
• There is a need for more robust clustering techniques that can handle noisy data and provide stable results.

Purpose of the Study:

• To propose a novel Sparse Pinball loss Twin Support Vector Clustering (SPTSVC) algorithm.
• To enhance the noise-insensitivity and resampling stability of TWSVC.
• To demonstrate the efficacy and applicability of SPTSVC on synthetic, benchmark, and biomedical datasets.

Main Methods:

• The proposed SPTSVC algorithm utilizes the ϵ-insensitive pinball loss function.
• This formulation aims to achieve a sparse solution, improving model efficiency and testing time.
• The algorithm's performance is evaluated through numerical experiments on various datasets.

Main Results:

• SPTSVC demonstrates improved noise-insensitivity and resampling stability compared to traditional hinge loss-based methods.
• Experiments show the efficacy of SPTSVC on both synthetic and real-world benchmark datasets.
• The proposed method is validated on biomedical datasets, including epilepsy and breast cancer data, showing its practical applicability.

Conclusions:

• The novel Sparse Pinball loss Twin Support Vector Clustering (SPTSVC) offers a more robust and stable clustering approach.
• SPTSVC's use of pinball loss enhances performance on noisy datasets and improves computational efficiency.
• The successful application on biomedical data highlights SPTSVC's potential in various scientific domains.