分论坛 > 南京 > 新闻动态
CCF YOCSEF南京即将举办“Theoretical analysis for Boosting and AUC optimization; Learning methods for safely using unlabelled data”报告会
2015-05-04 阅读量:710 小字

CCF YOCSEF南京

 

20150602日(星期二)10:00

东南大学九龙湖校区计算机楼413

会议安排

 

09:30      签到

10:00      报告会开始

特邀讲者:高尉,李宇峰 南京大学

演讲题目:Theoretical analysis for Boosting and AUC optimization; Learning methods for safely using unlabelled data

执行主席:耿新 CCF YOCSEF南京 AC委员,主席

 

参加人员:IT领域专业人士、研究生、媒体、其他有兴趣者

 

Theoretical analysis for Boosting and AUC optimization

报告提要: Learning theory plays an important role in the design and analysis of learning algorithms. Generalization and consistency are two central issues in machine learning theory. In this talk, I will introduce our progresses on Boosting and AUC optimization. First, we provide new generalization error bounds which defend the margin-based explanation against Breiman’s doubts, and by incorporating factors such as average margin and variance, we present generalization error bounds that is heavily related to the whole margin distribution. Second, AUC (Area Under ROC Curve) has been an important criterion, and many learning approaches have been developed by minimizing pairwise surrogate losses. We present sufficient and necessary condition for AUC consistency. Based on this finding, we develop the One-Pass AUC algorithm (OPAUC) which only needs to maintain the first and second-order statistics of training data in memory, resulting a storage requirement independent from the size of training data. To efficiently handle high-dimensional data, we develop a randomized algorithm that approximates the covariance matrices by low-rank matrices. We verify, both theoretically and empirically, the effectiveness of proposed algorithm.

特邀讲者:高尉

Wei Gao is currently an assistant researcher at computer science and technology department, Nanjing University. He received his Ph.D. degree from computer science, Nanjing University in 2014. His research focuses on machine learning, especially in theoretical analysis for boosting, multi-label learning, AUC optimization, etc. His research works were published in AIJ, COLT, ICML, IJCAI, AAAI, etc. He received the outstanding doctoral dissertation award from China Computer Federation (CCF 2014). He was served as reviewer of several journals and conferences such as MLJ, TNN, TKDE, IJCAI, NIPS, AAAI, etc. 

 

 

Learning methods for safely using unlabelled data

报告提要: When the amount of labelled data is limited, it is usually expected that learning methods exploiting additional unlabelled data will help improve learning performance. In many situations, however, it is reported that learning methods using unlabelled data may even decrease the learning performance. This phenomenon affects the deployment of unlabelled data learning methods in real-world situations. It is thus desirable to develop “safe” unlabelled data learning methods that often improve performance, while in the worst cases do not decrease the learning performance. In this talk, I introduce some of our recent progresses in this direction. Firstly, considering that decreased performance may be caused by not-good-enough optimization solution, we present a tight convex relaxation method and derive a scalable algorithm. Secondly, considering that decreased performance may be caused by the uncertainty of model assumption, we present a learning method based on the worst-case accuracy improvement so as to avoid the harm of uncertain model assumption. Finally, considering that decreased performance may be caused by the difficulty in optimizing complicate performance measures, we also present to optimize a combined performance gain for a set of common used performance measures (e.g., AUC, F1, Top-k precision) and show that these performance measures can be globally optimized in an efficient manner.

特邀讲者:李宇峰

Yu-Feng Li is currently an assistant researcher at computer science and technology department, Nanjing University. He received his B. Sc and Ph.D. degree from computer science, Nanjing University in 2006 and 2013, respectively. His research focuses on machine learning and data mining. Particularly, he is interested in semi-supervised learning, multi-instance learning, multi-label learning and statistical learning algorithms. His research works were published in JMLR, IEEE Trans. PAMI, AIJ, ICML, NIPS, AAAI, etc. He received several awards, including outstanding doctoral dissertation award from China Computer Federation (CCF), outstanding doctoral dissertation award from Jiangsu Province and Microsoft Fellowship Award. He was served as a program committee member of several conferences such as IJCAI, ICML, NIPS and KDD. 

 

交通方式:地铁3号线东南大学九龙湖校区站

 

 

CCF聚焦