Heejin Choi

PhD candidate at TTIC
Email: (my first name) + cs @ttic.edu

I am a PhD student at TTIC, a philanthropically endowed academic computer science institute located on the University of Chicago campus. I am advised by Nati Srebro. My interests are machine learning on large structured data, multi-task learning, and structured prediction.

Publications:

  1. Choi, Heejin, Ofer Meshi, and Nathan Srebro. "Fast and Scalable Structural SVM with Slack Rescaling." to be appeared in AISTATS 2016.

    While many advantages of slack rescaling exist over margin rescaling, slack rescaling is not tractable due to the fact the potential of each label does not decompose over the structures for large structures. In this paper, we present an efficient method to do inference the most violated label assuming that an efficient the oracle of margin rescaling is given. (under review at AISTATS 2016)


  2. Choi, Heejin, Yutaka Sasaki, and Nathan Srebro. "Normalized Hierarchical SVM." arXiv preprint arXiv:1508.02479 (2015).

    3. For a large hierarchical structures, for instance the structure of Wikipeda, it is common that the label structure is unbalanced, i.e. the distance of labels from the root to each labels are very different. Our proposed normalizes this inbalanced in the structure, and show that it is critical for the performance. Also, a new structural norm, shared Frobenius norm, is presented. (under revision)

  3. Choi, Heejin, and Nathan Srebro. . "Hierarchical Classification with Strutured SVMs" Extreme classification workshop on NIPS 2013.