Difference between revisions of "Z. Kou and W. Cohen. SDM 2007"
From Cohen Courses
Jump to navigationJump to searchPastStudents (talk | contribs) |
PastStudents (talk | contribs) |
||
| Line 8: | Line 8: | ||
== Summary == | == Summary == | ||
| − | This paper is an extension of Stacked Sequential Learning and shows how stacking can be used in non-sequential tasks, such as | + | This paper is an extension of Stacked Sequential Learning (SSL) and shows how stacking can be used in non-sequential tasks, such as |
text region detection and document classification. | text region detection and document classification. | ||
| Line 23: | Line 23: | ||
* Relational information comes from: the prediction of other candidates within the same panel and the predictions of four neighbor panels (up, down, right and left). | * Relational information comes from: the prediction of other candidates within the same panel and the predictions of four neighbor panels (up, down, right and left). | ||
* The predictions of the candidate region and neighbor regions return a vector of binary features indicating whether a character is found in those regions. | * The predictions of the candidate region and neighbor regions return a vector of binary features indicating whether a character is found in those regions. | ||
| + | |||
| + | '''webpage classification''' | ||
| + | * Task is to classify webpages on the WebKB dataset. | ||
| + | * Relational information comes from: the number of incoming and outgoing links in each category. | ||
| + | |||
| + | '''NER''' | ||
| + | * Relational information comes from nearby words. In this case Stacked Graphical Model is the same with SSL. | ||
Revision as of 22:45, 8 October 2010
Citation
Zhenzhen Kou and William W. Cohen. Stacked Graphical Models for Efficient Inference in Markov Random Fields in SDM-2007.
Online version
Summary
This paper is an extension of Stacked Sequential Learning (SSL) and shows how stacking can be used in non-sequential tasks, such as text region detection and document classification.
The key points of the paper are:
- arbitrary base learner
- very fast inference time, 40 to 80 times faster than Gibbs sampling.
- relational information is captured by augmenting data with the predictions of related instances.
- Authors showed its applications on the problems where the relation of instances is not sequential.
Example Stacked Graphical Models Usage
text region detection
- Task is to find the text regions in figures, where each figure may contain several panels.
- Relational information comes from: the prediction of other candidates within the same panel and the predictions of four neighbor panels (up, down, right and left).
- The predictions of the candidate region and neighbor regions return a vector of binary features indicating whether a character is found in those regions.
webpage classification
- Task is to classify webpages on the WebKB dataset.
- Relational information comes from: the number of incoming and outgoing links in each category.
NER
- Relational information comes from nearby words. In this case Stacked Graphical Model is the same with SSL.
