Incomplete Label Multiple Instance Multiple Label Learning.

Area of Science:

• Machine Learning
• Artificial Intelligence
• Data Science

Background:

• Manual data labeling is a significant bottleneck in machine learning due to increasing data volumes.
• Reduced label settings like semi-supervised and active learning aim to mitigate labeling costs.
• Multiple-instance multiple-label learning with missing bag labels is an underexplored but important area.

Purpose of the Study:

• To introduce a novel discriminative probabilistic model for multiple-instance multiple-label learning with missing bag labels.
• To address the inference challenges associated with this learning setting.
• To evaluate the model's robustness and performance compared to existing methods.

Main Methods:

• Developed a novel discriminative probabilistic model for handling missing labels in multiple-instance multiple-label learning.
• Implemented an efficient expectation-maximization (EM) algorithm for model inference.
• Explored an alternative inference approach using label-wise marginal likelihood maximization.

Main Results:

• The proposed model demonstrates robustness on benchmark datasets.
• The approach shows a significantly smaller performance decrease with increased proportions of missing labels compared to state-of-the-art methods.
• The EM algorithm implementation provides an efficient inference solution.

Conclusions:

• The novel probabilistic model effectively addresses the challenge of missing bag labels in multiple-instance multiple-label learning.
• The proposed inference methods are efficient and robust.
• This work offers a promising solution for scenarios with incomplete supervision in complex labeling tasks.