OSNN: An Online Semisupervised Neural Network for Nonstationary Data Streams.

Area of Science:

• Machine Learning
• Artificial Intelligence
• Data Science

Background:

• Data streams from nonstationary environments present challenges due to evolving target functions and data distributions (concept drifts).
• Existing methods for concept drift adaptation in classification primarily rely on labeled data, which is often prohibitively expensive for life-long data streams.
• Exploiting abundant unlabeled data is crucial for efficient adaptation in dynamic environments.

Purpose of the Study:

• To develop a novel algorithm that leverages unlabeled data for effective concept drift adaptation in classification problems.
• To introduce an online semisupervised radial basis function neural network (OSNN) capable of handling nonstationary data streams.
• To enhance robustness to concept drifts through dynamic learning and manifold-based training.

Main Methods:

• An online semisupervised radial basis function neural network (OSNN) was developed.
• The OSNN utilizes a novel semisupervised learning vector quantization (SLVQ) for training network centers and learning dynamic data representations.
• Manifold learning on dynamic graphs was employed to adjust network weights, enabling adaptation to concept drifts.

Main Results:

• The proposed OSNN effectively utilizes unlabeled data to uncover underlying data stream structures.
• The algorithm demonstrates robustness to concept drifts through its dynamic topology learning capabilities.
• Experimental results validate the effectiveness of OSNN in semisupervised classification within nonstationary environments.

Conclusions:

• The OSNN algorithm offers a powerful solution for learning from data streams with concept drifts by integrating unlabeled data.
• Semisupervised learning combined with manifold techniques provides a robust framework for adapting to changing data distributions.
• This approach significantly reduces the reliance on expensive data labeling while maintaining high classification performance.