Difference between revisions of "Rosen-Zvi et al UAI 2004"

From Cohen Courses
Jump to navigationJump to search
 
(7 intermediate revisions by the same user not shown)
Line 11: Line 11:
 
== Summary ==
 
== Summary ==
  
−
This paper proposes the generative author-topic model, which extends LDA to model document contents and authors at the same time. The interest of each author is represented with a multinomial distribution over topics, while each topic is a multinomial distribution over words. Model estimation is performed with Gibbs sampling. Experiments show the topic-author and topic-word results discovered on the NIPS and [[UsesDataset::CiteSeer]] datasets.
+
This paper proposes the generative author-topic model, which extends [[LDA]] to model document contents and authors at the same time. The interest of each author is represented with a multinomial distribution over topics, while each topic is a multinomial distribution over words. Model estimation is performed with [[Gibbs sampling]]. Experiments show the topic-author and topic-word results discovered on the [[UsesDataset::NIPS]] and [[UsesDataset::CiteSeer]] datasets.
  
−
The experimented on 3 real world social graphs [[UsesDataset::Conference-Author dataset]], [[UsesDataset::Author-Paper dataset]] and [[UsesDataset::IMDB dataset]]
+
== Evaluation ==
 +
 
 +
This paper presents three aspects in the evaluation:
 +
  - qualitatively show topics and author interests modeled from the two corpora
 +
  - compare the [[Perplexity]] (predictive power) of the author-topic model with [[LDA]]
 +
  - reveal potential applications of the author-topic model in author similarity computation and reviewer recommendation
 +
 
 +
== Discussion ==
 +
My views on this paper are:
 +
* This is a natural (but smart) extension of the popular [[LDA]] model. It inspired a bunch of following works including the author-recipient-topic model.
 +
* Modeling author/user interests in the form of topic distributions is very useful in practice. For example, we can recommend news or papers to users based on their interests. Also, expert/expertise search can be improved by exploiting user interest information.
 +
* Compared with scientific papers, social media has much more information such as links, user groups and interaction patterns. The author-topic model can be extended to incorporate these information and model user interests/preferences in social media services.
 +
 
 +
== Related papers ==
 +
Here are two papers related with this work.
 +
* [http://www.cs.umass.edu/~mccallum/papers/art04tr.pdf The Author-Recipient-Topic Model for Topic and Role Discovery in Social Networks] is directly based on [[LDA]] and the author-topic model. It models links (interactions) in social networks, where each link is conditioned on the sender and recipient.
 +
* [http://www.datalab.uci.edu/author-topic/398.pdf Citation Author Topic Model in Expert Search] exploits citations to model author interests and perform expert search.

Latest revision as of 22:22, 26 September 2012

This a Paper discussed in Social Media Analysis 10-802 in Fall 2012.

Citation

The Author-Topic Model for Authors and Documents. Michal Rosen-Zvi, Thomas Griffiths, Mark Steyvers, Padhraic Smyth. In Proceedings of UAI 2004, pages 487-494.

Online version

The Author-Topic Model for Authors and Documents

Summary

This paper proposes the generative author-topic model, which extends LDA to model document contents and authors at the same time. The interest of each author is represented with a multinomial distribution over topics, while each topic is a multinomial distribution over words. Model estimation is performed with Gibbs sampling. Experiments show the topic-author and topic-word results discovered on the NIPS and CiteSeer datasets.

Evaluation

This paper presents three aspects in the evaluation: 

 - qualitatively show topics and author interests modeled from the two corpora
 - compare the Perplexity (predictive power) of the author-topic model with LDA
 - reveal potential applications of the author-topic model in author similarity computation and reviewer recommendation

Discussion

My views on this paper are:

  • This is a natural (but smart) extension of the popular LDA model. It inspired a bunch of following works including the author-recipient-topic model.
  • Modeling author/user interests in the form of topic distributions is very useful in practice. For example, we can recommend news or papers to users based on their interests. Also, expert/expertise search can be improved by exploiting user interest information.
  • Compared with scientific papers, social media has much more information such as links, user groups and interaction patterns. The author-topic model can be extended to incorporate these information and model user interests/preferences in social media services.

Related papers

Here are two papers related with this work.

Retrieved from "http://curtis.ml.cmu.edu/w/courses/index.php?title=Rosen-Zvi_et_al_UAI_2004&oldid=12046"