Robust Online Multilabel Learning Under Dynamic Changes in Data Distribution With Labels.

Area of Science:

• Machine Learning
• Data Mining
• Artificial Intelligence

Background:

• Multilabel learning deals with data where instances can belong to multiple categories.
• Dynamically changing data streams pose challenges for traditional machine learning models.
• Concept drift, specifically changes in data distribution with labels (CDDL), is a critical issue in streaming multilabel learning.

Purpose of the Study:

• To propose a robust online multilabel learning method for dynamic data streams.
• To address the challenge of concept drift, particularly CDDL.
• To improve accuracy, training speed, and robustness in multilabel learning scenarios.

Main Methods:

• A novel objective function based on label ranking for higher accuracy.
• A closed-form solution for fast training and model updates, avoiding gradient descent.
• A sequential update rule preserving label ranking information.
• A fixed threshold for label bipartition to ensure robustness against data distribution changes.

Main Results:

• The proposed method demonstrates high robustness to CDDL in both sequential updates and multilabel thresholding.
• Significant improvements in evaluation measures like Hamming loss, F1-measure, Precision, and Recall were observed.
• The method achieved short training times across most evaluated benchmark datasets.

Conclusions:

• The developed online multilabel learning method is robust to concept drift, specifically CDDL.
• The approach offers a balance of high accuracy, computational efficiency, and adaptability to evolving data streams.
• This work provides a valuable tool for real-world applications involving dynamic multilabel data.