Revisiting Multi-dimensional Classification from a Dimension-wise Perspective

While the class imbalance issue has been extensively investigated within the multi-class paradigm, its study in the multi-dimensional classification (MDC) context has been limited due to the imbalance shift phenomenon. A sample's classification as a minor or major class instance becomes ambiguous when it belongs to a minor class in one labeling dimension (LD) and a major class in another. Previous MDC methodologies predominantly emphasized instance-wise criteria, neglecting prediction capabilities from a dimension aspect, i.e., the average classification performance across LDs.

To solve the problems, a research team led by De-Chuan Zhan from LAMDA, Nanjing University published their new research on Multi-dimensional Classification in Frontiers of Computer Science co-published by Higher Education Press and Springer Nature.

The team asserts the significance of dimension-wise metrics in real-world MDC applications and introduce two such metrics. Furthermore, the team observes imbalanced class distributions within each LD and propose a novel Imbalance-Aware fusion Model (IMAM) for addressing the MDC problem.

In the research, IMAM first decomposes the task into multiple multi-class classification problems, creating imbalance-aware deep models for each LD separately. This straightforward method performs well across LDs without sacrificing performance in instance-wise criteria. Subsequently, IMAM employs LD-wise models as multiple teachers and transfer their knowledge across all LDs to a unified student model.

Extensive experiments are conducted on various MDC datasets. The results indicate that the proposed IMAM is superior to others in a big gap.

DOI: 10.1007/s11704-023-3272-9

https://journal.hep.com.cn/fcs/EN/10.1007/s11704-023-3272-9

Research Article, Issue Date: 14 Mar 2024
Yi Shi, Han-Jia Ye, Dong-Liang Man, Xiao-Xu Han, De-Chuan Zhan, Yuan Jiang. Revisiting Multi-dimensional Classification from a Dimension-wise Perspective. Front. Comput. Sci., https://doi.org/10.1007/s11704-023-3272-9

Attached files

Figure 1.The imbalance shift from one LD to another. The color map counts the number of instances over two LDs on Zappos. The numerical values annotated on the colored blocks in the figure represent values post logarithmic transformation. Many major class instances become minor ones when the LD changes. In other words, the major/minor class property of an instance is difficult to be kept across LDs.
Figure 2. An illustration of our proposed IMAM approach based on an MDC problem with two LDs. In the decomposition step (a), we construct imbalance-aware deep models for each LD. In the fusion step (b), we use the models in the former step as fixed teachers and fuse their knowledge into a compact student. Both embedding (green) and classifier (red) distillations help in matching knowledge between models. Subscript “T” denote the component of teacher. “CE” means the cross-entropy.
Table 1. Performance of four traditional MDC methods, four deep MDC methods, and our IMAM on four real-world datasets. Four evaluation metrics from instance and label aspects are calculated.