Improving attribute classification with imperfect pairwise constraints

li, Zequn, Li, Honglei and Shao, Ling (2019) Improving attribute classification with imperfect pairwise constraints. Proceedings of the International Conference on Electronic Business (ICEB), 2019. pp. 253-262. ISSN 1683-0040

Preview

Text ICEB_2019_paper_72_full.pdf - Published Version Download (521kB) | Preview

Abstract

Semantic attributes extracted from images could help to improve many interesting applications, including image classification, recommendation systems and online advertising. However, learning of such attributes requires a large well-labelled dataset which is usually difficult and expensive to collect and sometimes requires human domain experts to annotate. Partially labelled data, on the contrary, are relatively easy to obtain from social media websites or be annotated by less experienced people. However, a partially labelled dataset usually contains a lot of noisy data which are challenging for previous methods. In this paper, we propose a semi-supervised Random Forest algorithm that can handle a small well-labelled attribute dataset and large scale pairwise data at the same time for classifying grouped attributes. Results on two typical attribute datasets show that the proposed method outperforms the state-of-the-art attribute learner.

Actions (login required)

View Item

Downloads