Abstract—This paper studies the problem of constructing a suitable hierarchy for hierarchical classification. It presents a new method to fuse multiple similarity relatedness between concepts. The method is based on the kernel target alignment technology. We also develop a method to construct a hierarchy for image classification automatically. The hierarchy is constructed based on the previous fused similarity measure. Then, we utilize the structured support vector machine (SVM) for classification with a meaningful hierarchy. Experiments on tow real-world datasets show that hierarchical classification perform better than flat classification, and the structured SVM with the fused classes hierarchy provides a better image classification.
Index Terms—Hierarchies construction, hierarchical classification, taxonomies, structural learning.
The authors are with School of Computer Science and Technology, Tianjin University, Tianjin 300350, P. R.China (e-mail: szhao@tju.edu.cn, zouquan@nclab.net).
Cite: Shuo Zhao and Quan Zou, "Fusing Multiple Hierarchies for Semantic Hierarchical Classification," International Journal of Machine Learning and Computing vol.6, no. 1, pp. 47-51, 2016.