Dynamic Programming for Instance Annotation in Multi-Instance Multi-Label Learning.

Area of Science:

• Machine Learning
• Computer Science
• Data Science

Background:

• Labeling data for classification tasks is labor-intensive and costly.
• Traditional methods require individual instance labeling, which is inefficient.
• Group-level labeling (bag labeling) offers a cost-effective alternative but introduces ambiguity in instance-level labels.

Purpose of the Study:

• To develop a discriminative probabilistic model for instance annotation within bags.
• To create an efficient inference framework using expectation maximization.
• To address the computational challenges of calculating instance label posterior probabilities.

Main Methods:

• Proposed a discriminative probabilistic model for instance annotation.
• Introduced an expectation maximization framework for inference based on maximum likelihood.
• Developed a dynamic programming method for efficient posterior probability computation (linear time complexity).

Main Results:

• The dynamic programming approach significantly reduces computational complexity compared to brute-force methods.
• The proposed framework demonstrated superior performance over state-of-the-art Multiple-Instance Learning (MIL) methods.
• Outperformance was observed in both instance label prediction and bag label prediction across diverse datasets.

Conclusions:

• The novel framework effectively handles the ambiguity inherent in instance annotation within bags.
• The dynamic programming method provides a computationally efficient solution for probabilistic inference.
• The approach offers a significant advancement for cost-effective and accurate data labeling in various domains.