Can you use past clicks on search results to predict future clicks? Efficient Multiple-Click Models in Web Search is a Microsoft paper aiming at providing an answer to this question.
Two different models are evaluated. An Independent Click Model (ICM) where clicks are indipendent each other (i.e. position is not taken into account), and a Dependent Click Model (DCM) where position is taken into account. The models are built on a strong theoretical background based on log-likelihood maximization. Anyway, the final formulas derived are very simple to compute and requires linear time and space and can be computed incrementally.
a similar paper based on Bayesan network and HMM / EM is http://www2009.eprints.org/1/1/p1.pdf
ReplyDelete