Variational Label Enhancement.

Area of Science:

• Machine Learning
• Artificial Intelligence
• Data Science

Background:

• Traditional multi-label learning rigidly partitions labels into relevant/irrelevant categories.
• Real-world data often exhibits relative label relevance, requiring a more nuanced approach.
• Existing methods lack the ability to capture fine-grained label distributions.

Purpose of the Study:

• To propose LEVI, a novel method for label enhancement in multi-label learning.
• To recover the underlying fine-grained label distributions from training data.
• To improve the performance of multi-label predictive models.

Main Methods:

• Developed a generative model for label distributions.
• Employed variational inference to approximate posterior densities.
• Induced a multi-label predictive model using recovered distributions and a specialized objective function.

Main Results:

• Successfully recovered label distributions on fourteen datasets.
• Demonstrated superior performance of the multi-label predictive model on fourteen datasets.
• Validated the advantage of LEVI over state-of-the-art approaches.

Conclusions:

• LEVI effectively recovers fine-grained label distributions, addressing limitations of rigid label assignments.
• The proposed method enhances multi-label learning by leveraging recovered distributions.
• LEVI offers a significant advancement for multi-label classification tasks.