Hessian-regularized co-training for social activity recognition.

Area of Science:

• Machine Learning
• Computer Vision
• Data Science

Background:

• Co-training is a multi-view learning method that trains classifiers on distinct data views.
• Traditional co-training can suffer from unstable performance due to erroneous labels from mediocre classifiers, especially with limited initial labeled data.

Purpose of the Study:

• To introduce Hessian-regularized co-training to enhance classifier stability and generalizability.
• To address the limitations of traditional co-training algorithms in scenarios with scarce labeled data.

Main Methods:

• Proposed Hessian-regularized co-training, integrating Hessian information from each data view.
• Penalized regression functions along the potential data manifold to exploit local data structure.
• Conducted experiments on the unstructured social activity attribute (USAA) dataset for social activity recognition.

Main Results:

• Hessian regularization significantly boosts classifier generalizability, particularly with few labeled and many unlabeled examples.
• The proposed method demonstrated superior performance compared to traditional co-training and LapCo algorithms.
• Validated effectiveness in the context of social activity recognition.

Conclusions:

• Hessian-regularized co-training offers a robust solution to improve multi-view learning stability and accuracy.
• The method is particularly beneficial for datasets with limited labeled instances.
• Effective application shown in social activity recognition tasks.