User Adaptive Image Ranking for Search Engines Maryam Mahdaviani Nando de Freitas Laboratory for Computational Intelligence University of British Columbia.

User Adaptive Image Ranking for Search Engines

Maryam Mahdaviani Nando de FreitasLaboratory for Computational Intelligence

University of British Columbia

• Screen shot of apple/red apple/red apple fruit

• Screen shot of tiger

Image Retrieval systems mainly use linguistic

features (e.g. words) and not visual cues

Word Polysemy is a common problem in

IR system

legend

before after

Page 2 Page 3 Page 11

How do we do it? Instance Preference Learning by Gaussian Processes

• We want to learn a better ranking from m pair-wise relations: for

• We use the standard hierarchical Bayes probit model [Hebrich et al, NIPS 06; Wei Chu et al, ICML 05]

mk ,...,1kk uv

How do we do it? Instance Preference Learning by Gaussian Processes

• It then follows that :

• The posterior can be easily computed either using MCMC, Laplace’s method, mean field or Expectation Propagation.

before after

legend

Can also do Active Preference Learning

• The system prompts user with intelligent questions to increase the confidence in ranking

• The user can stop questioning once she is annoyed

• The system re-ranks the images based on the preferences

• We calculate for each unlabeled pair; pick the maximum and query the user accordingly [Wei Chu et al, NIPS 05]

before after

legend

?

Water is hardlegend

Conclusion and Future Directions

• We applied state-of-the-art preference learning algorithm for image ranking

• In future we should work on:

Improving the HCI

Improving the vision

Conducting using study

Expand the idea to other search Learning from many sources

Thank You!

Questions? Feedback?

Acknowledgment:

The code for this work has been built on Wei Chu’s supervised preference learning package, which is available online