Click Models

http://www.cnblogs.com/bentuwuying/p/6227645.html

http://www.cnblogs.com/bentuwuying/p/6241504.html


http://www.sciencedirect.com/science/article/pii/S2468232216300245
Vertical-aware Click Model
Sequential Click Model

Most click models follow the examination hypothesis [6]: a document being clicked (Ci=1) should satisfy (→) two conditions: it is examined (Ei=1) and it is relevant (Ri=1) (most click models assume P(Ri=1)=ru, which is the probability of the perceived relevance), and these two conditions are independent of each other.

equation(1)

Ci=1→Ei=1,Ri=1

Turn MathJaxoff

equation(2)

Ei=0→Ci=0

Turn MathJaxoff

equation(3)

Ri=0→Ci=0

Turn MathJaxoff

Following this assumption, the probability of a document being clicked is determined as follows:

equation(4)

P(Ci=1)=P(Ei=1)P(Ri=1)

Craswell et al. [6] proposed the cascade model, which assumes that while a user examines the results from top to bottom sequentially, he/she immediately decides whether to click on a result. The cascade model is mostly suitable for single-click sessions. A number of succeeding models were proposed to improve both its applicability and performance.

equation(5)

P(E1)=1

Turn MathJaxoff

equation(6)

P(Ei+1=1|Ei=1,Ci)=1−Ci

Turn MathJaxoff

Here the examination of the (i+1)-th result indicates the i-th result has been examined but not clicked. Although the cascade model performs well in predicting the click-through rates, this model is only suited for a single-click scenario.

Based on the cascade hypothesis, the Dependency Click Model (DCM) [9] extends the cascade model in order to model user interactions within multi-click sessions. DCM assumes that a user may have a certain probability of examining the next document after clicking the current document, and this probability is influenced by the ranking position of the result. The DCM model is characterized as follows:

equation(7)

P(Ei+1=1|Ei=1,Ci=0)=1

Turn MathJaxoff

equation(8)

P(Ei+1=1|Ei=1,Ci=1)=λi

Turn MathJaxoff

where λi represents the preservation probability¹ of the position i.