Data Preprocessing for Learning to Rank - Learning to Rank for Information Retrieval - page 161

Fig. 13.1 The DCM model

In recent years, several new click models have been proposed in order to solve the
aforementioned problems. In this subsection, we will introduce three representative
models, namely the Dependent Click Model [ 19 ], Bayesian Browsing Model [ 24 ],
and Dynamic Bayesian Network Click Model [ 7 ].

13.2.1.1 Dependent Click Model

As aforementioned, the cascade model assumes that a user abandons examination of
web documents upon the first click. This unfortunately restricts the modeling power
to query sessions with at most one click, which leaves the gap open for real-world
applications where multiple clicks are possible, especially for informational queries.
In order to tackle the problem, in [ 19 ], a dependent click model (DCM) is pro-
posed, which generalizes the cascade model to multiple clicks by including a set of
position-dependent parameters to model probabilities that an average user returns to
the search result page and resumes the examination after a click. The DCM model
can be illustrated using Fig. 13.1 .
According to Fig. 13.1 , a position-dependent parameter λ i is used to reflect the
chance that the user would like to see more results after a click at position i . In case
of a skip (no click), the next document is examined with probability one. The λ i are
a set of user behavior parameters shared over multiple query sessions.
The examination and click probabilities in the DCM model can be specified in
the following iterative process (1

≤ i ≤ m , where m is the number of documents in

the search result):

e d 1 , 1 = 1 ,

c d i ,i = e d i ,i r d i , (13.1)

e d i + 1 ,i + 1 = λ i c d i ,i + (e d i ,i − c d i ,i ),