Multiview Multi-Instance Multilabel Active Learning.

Area of Science:

• Machine Learning
• Computer Vision
• Data Science

Background:

• Multiview Multi-instance Multilabel (M3L) learning models complex objects with multiple feature representations and non-exclusive labels.
• Traditional M3L methods require extensive labeled data, which is costly and often impractical for training predictive models.

Purpose of the Study:

• To develop an active learning-based M3L approach (M3AL) to minimize labeling expenses.
• To enhance the performance of M3L models by reducing the need for large labeled datasets.

Main Methods:

• M3AL utilizes multiview self-representation learning to extract shared and individual information from bags across different views.
• A novel query strategy is introduced, leveraging bag-instance distributions and view-specific information to select the most informative bag-label pairs for annotation.

Main Results:

• Experimental results demonstrate that M3AL significantly reduces the cost of querying informative data points.
• M3AL achieves superior performance compared to existing M3L methods when operating under similar labeling budget constraints.

Conclusions:

• The proposed M3AL approach effectively addresses the high labeling cost challenge in M3L.
• M3AL offers a promising solution for efficient and effective modeling of complex objects in machine learning applications.