Multi-label classification via label correlation and first order feature dependance in a data stream. (June 2019)

• Record Type: Journal Article
• Authors: Nguyen, Tien Thanh; Nguyen, Thi Thu Thuy; Luong, Anh Vu; Nguyen, Quoc Viet Hung; Liew, Alan Wee-Chung; Stantic, Bela
• Is Part Of: Pattern recognition. Volume 90(2019:Jun.)
• Page Start: 35
• Page End: 51
• Publication Date: 2019-06
• Subjects: Multi-label classification -- Multi-label learning -- Online learning -- Data stream -- Concept drift -- Label correlation -- Feature dependence; Pattern perception -- Periodicals; Perception des structures -- Périodiques; Patroonherkenning; 006.4
• DOI: 10.1016/j.patcog.2019.01.007
• Languages: English
• ISSNs: 0031-3203

Abstract

Highlights: A Bayesian-based multi-label online learning method for multi-label data stream classification is proposed. Our method not only learns the label correlation from each arrived sample but also dynamically determines the number of predicted labels based on Hoeffding inequality and the label cardinality. Our method can also handle missing values and concept drifts in the data stream effectively. Extensive comparative experiments with the state-of-the-art algorithms validate the superior performance of our method. Abstract: Many batch learning algorithms have been introduced for offline multi-label classification (MLC) over the years. However, the increasing data volume in many applications such as social networks, sensor networks, and traffic monitoring has posed many challenges to batch MLC learning. For example, it is often expensive to re-train the model with the newly arrived samples, or it is impractical to learn on the large volume of data at once. The research on incremental learning is therefore applicable to a large volume of data and especially for data stream. In this study, we develop a Bayesian-based method for learning from multi-label data streams by taking into consideration the correlation between pairs of labels and the relationship between label and feature. In our model, not only the label correlation is learned with each arrived sample with ground truth labels but also the number of predicted labels are adjusted based on Hoeffding inequality and the label cardinality. We also extend the model to handle missing values, a problem common in many real-world data. To handle concept drift, we propose a decay mechanism focusing on the age of the arrived samples to incrementally adapt to the change of data. The experimental results show that our method is highly competitive compared to several well-known benchmark algorithms under both the stationary and concept drift settings.