查看“Sunwei”的源代码


{| cellspacing="1" cellpadding="1" border="0"
|-
| 
M.Sc. Student @ IIP&nbsp;Group<br/> Department of Computer Science and Technology<br/> Nanjing University

Email: weisun_@outlook.com

|-
| &nbsp;
| &nbsp;
|}

<span style="color:#3498db"><span style="font-size:larger">'''Supervisor'''</span></span>

----

*Professor Jun-Yuan Xie 

<span style="font-size:larger"><span style="color:#3498db">'''Biography'''</span></span>

----

*I&nbsp;received my B.Sc. degree in&nbsp;of Soochow University in June 2017. In the same year, I was admitted to study for a Master degree in Nanjing University&nbsp;without entrance examination.&nbsp; Currently I am a second&nbsp;year M.Sc. student of Department of Computer Science and Technology&nbsp;in Nanjing University&nbsp;and a member of IIP&nbsp;Group, led by professor Jun-Yuan Xie and&nbsp; Chong-Jun Wang. 

<span style="font-size:larger"><span style="color:#3498db">'''Research Interest'''</span></span>

----

*I am interested in Machine Learning and Multi-label Learning.&nbsp; 
*Specifically, now my first work is to exploit label correlations on multi-label learning, i.e. On Multi-label Text Classification (MLTC), text features can be regarded as '''detailed description of documents''' and label sets can be '''a summarization of documents'''. '''Hybrid Topics''' from text features and label sets by LDA (a method of '''topic model''') can effectively&nbsp;represent the whole label correlations. 
*My second work is to apply deep learning to multi-label text classification. We utilize '''dilated convolution''' to obtain the '''semantic understanding''' of the text and design a hybrid '''attention mechansim''' for '''different labels''' (Specifically, each label should attend to most relevant textual information).&nbsp; First, we initialize trainable label embeddings. We obation word-level information base on Bi-LSTM and get semantic understanding of a document based on word-level information. Then, we design a hybrid attention for different labels based on label embeddings.&nbsp; Besides, we add label cooccurrence matrix into loss function to guide the whole network to learn and achieve good results. 

<span style="font-size:larger"><span style="color:#3498db">'''Resources'''</span></span>

----

*[http://manikvarma.org/downloads/XC/XMLRepository.html Extreme Classiﬁcation Repository]: for large-scale multi-label datasets and off-the-shelf eXtreme Multi-Label Learning (XML) solvers. 
*[http://mulan.sourceforge.net/datasets-mlc.html Mulan Multi-Label Learning Datasets]: regular/traditional multi-label learning datasets. 
*[https://github.com/XSilverBullet/Multi-label-Paper Related Works]: This page categorizes a list of works of my interest, mainly in Multi-Label Learning. 

<span style="font-size:larger"><span style="color:#3498db">'''Rewards or Honors'''</span></span>

----

*Second-Class Academic Scholarship, 2018-2019 
*First-Class Academic Scholarship, 2017-2018 
*Outstanding Graduate Student, 2017.06 
*CCF Excellent University Student, 2016.10 
*National Scholarship, 2015.11