0.25
0.5
0.75
1.25
1.5
1.75
2
Sparse modeling: some unifying theory and “topic-imaging”
Published on 2011-05-064810 Views
Information technology has enabled collection of massive amounts of data in science, engineering, social science, finance and beyond. Extracting useful information from massive and high-dimensional d
Presentation
Some Unifying Theory and Topic Imaging00:00
David Blackwell (1919-2010)02:42:10
IT revolution --> data revolution19:10:00
Spectrum of Media Reporting58:52:26
Improves News Media Analysis, Improves News Media, Improves How the World Works69:32:27
Our Approach to Media Analysis: Topic-Imaging91:23:20
Case Study: Topic Image of Business Section of NYT104:36:08
Case Study (cont)120:50:00
Solving a modern data problem153:10:14
Today’s Talk168:30:04
Occam’s Razor180:05:22
Occam’s Razor via Model Selection in Linear Regression194:10:13
Sparse Modeling in the 70’s: Model Selection209:26:40
Model Selection for Topic-Imaging Problem250:04:06
Lasso: L1-norm as a Penalty286:39:36
Lasso: Computation and Evaluation304:14:52
Lasso: Theoretical Work331:41:49
Regularized M-estimation including Lasso384:56:26
Example 1: Lasso (sparse linear model)405:50:06
Example 2: Structured (inverse) Cov. Estimation430:27:57
Example 3: Low-rank matrix approximation436:28:34
Unified Analysis460:31:04
Why can we estimate parameters?490:31:44
In high-dim and when r corresponds to true structure (e.g. sparsity), why estimation is still possible509:17:01
Main Result for Regularized M-estimation562:20:21
Examples of decomposable regularizers577:19:12
Recovering Existing Result in Bickel et al 08583:00:41
Obtaining New Result (Robustness of Lasso)598:45:23
Summary of unified analysis614:04:30
Partial Summary622:37:35
Topic Imaging: Subject-Specific Summarization of Document Corpus633:14:59
Our approach: predictive sparse methods + human experiment640:39:41
Sample Result from Our Approach (Document-S^3)652:50:03
Document Corpus672:12:09
Pre-processing682:29:15
Matrix Set-up and Labeling688:33:11
List of Generation Methods693:56:53
Flow Chart of Our Automatic Summarization700:09:31
How to Select from 120 Possible Lists for One Topic?702:56:11
Prediction Is Not the Goal721:37:06
A Human Experiment740:41:37
Running Human Experiment755:02:23
Human Experiment Results in a Glance776:39:41
Qualitative Summary of Human Exp. Results790:27:19
Summary of talk803:22:05
Future Directions: Richer Data/Subejct Applications826:46:57
Future Directions: Research Topics833:22:31
Acknowledgements848:32:54
Stat-news project (El Ghaoui and Yu)856:05:40