I need to do relevance feedback that works with objects for which a low-level vector representation is always available and for which in some cases structured+free text metadata are available (e.g., text documents, images, 3d-models, audio). Are there any methods out there that are preferable to Rochio in this scenario? Should I consider active learning?

Interaction with the system can be initiated in three ways:

  1. A query string
  2. An object already in the database
  3. A user-uploaded object not in the database

I want to be able to use the same family of feedback mechanism in all cases (i.e. different parameterization of the same method is acceptable for different object modalities and query types).

asked Mar 02 '11 at 12:52

Oscar%20T%C3%A4ckstr%C3%B6m's gravatar image

Oscar Täckström
2024133450

edited Mar 02 '11 at 13:10

If you have everything in a vector space why not semi-supervised learning a classifier that predicts positive versus negative feedback using the feedback as the supervised training set and the originally returned items as the unsupervised training set?

(Mar 02 '11 at 13:28) Alexandre Passos ♦

The feedback will work by letting the user select relevant/non-relevant objects in the results set and then the query will be expanded and/or reweighted to better reflect what the user is looking for. I could probably use semi-supervised learning as you suggest, but I am worried that it a) would be computationally very expensive, and b) would not allow to find objects that where not in the original result set.

The idea I have currently is to let the user select objects from a set of simultaneously maximally diverse and maximally relevant objects and then use Rochio to upate the query based on the selections.

(Mar 02 '11 at 13:43) Oscar Täckström
Be the first one to answer this question!
toggle preview

Subscription:

Once you sign in you will be able to subscribe for any updates here

Tags:

×4

Asked: Mar 02 '11 at 12:52

Seen: 450 times

Last updated: Mar 02 '11 at 13:44

powered by OSQA

User submitted content is under Creative Commons: Attribution - Share Alike; Other things copyright (C) 2010, MetaOptimize LLC.