3
$\begingroup$

The hierarchical dirichlet process (Teh 2005) allows you to discover unlimited topics to describe a document. An alternative process, the Indian Buffet process (Griffiths 2011) is another nonparametric method that allows you to find hidden features that are combined together to explain the data.

My question is, can the IBP be used for topic modeling, and if so how would the topics differ from using the HDP?

My intuition is that the only difference is that IBP topics would overlap with each other, and HDP topics explicitly don't. But I'm not sure this is accurate.

Improve this question
$\endgroup$

1 Answer 1

Reset to default
2
$\begingroup$

Your intuition is not correct. IBP allows observations (customers) to share features (dishes), similarly HDP follows the metaphor of the Chinese Restaurant Franchise. In the CRF tables are shared among restaurants (features are shared) and customers (observations) select a table at one restaurant and according to that table (feature) the customer selects another table and so on.

Both HDP and IBP can be used for topic modeling, the IBP decouples into the Hierarchical Beta process see: http://people.ee.duke.edu/~lcarin/thibaux-jordan-aistats07.pdf

The problem with IBP is complicated Gibbs Sampling Procedures and its inability to scale well with large datasets.

Improve this answer
$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.