Incremental multi-class semi-supervised clustering regularized by Kalman filtering.

Area of Science:

• Machine Learning
• Computer Vision
• Data Science

Background:

• Sequential data analysis presents challenges due to evolving patterns.
• Semi-supervised learning offers a solution by utilizing limited labeled data.
• Kernel Spectral Clustering (KSC) is a powerful clustering technique.

Purpose of the Study:

• To develop an on-line semi-supervised learning algorithm for sequential data.
• To adapt a multi-class semi-supervised KSC (MSS-KSC) algorithm for real-time clustering.
• To improve label estimation for unlabeled data points using prototypes.

Main Methods:

• Formulation as a regularized kernel spectral clustering (KSC) approach.
• Adoption and adaptation of a multi-class semi-supervised KSC (MSS-KSC) algorithm for on-line data.
• Utilizing out-of-sample extension for updating data point memberships.
• Integration of Kalman filter tracking for object motion labeling and regularization.

Main Results:

• The algorithm effectively estimates labels for unlabeled data in a sequential manner.
• Demonstrated performance on synthetic datasets and real-life video segmentation.
• Successful application of video segmentation as a semi-supervised learning problem.
• Kalman filter integration enhances regularization and labeling accuracy.

Conclusions:

• The proposed on-line semi-supervised KSC algorithm is effective for clustering sequential data.
• The method shows promise for dynamic environments like video segmentation.
• Combining KSC with Kalman filtering provides a robust approach for evolving clusters.