"Click Model Learning Note" A Survey on the Click Modeling in Web Search_lianghao li_ppt

is a review of the nature of PPT.

Main content:
To summarize the research results in the academic field by predicting the ad clicks in the search.
Search Ads main Display bit is: 1. Search results page on top side; 2. Search results to the right.
Research significance: The number of clicks on ads directly affects revenue

Problem abstraction: For a query Q, and an ad ad, predict the user clicks on them.

Specific content:

1. The simplest click Model: Predict by the number of clicks, the calculation formula is
P = #count of clicks/#count of Impressions (presentation)
Cons: Click is affected by user browsing behavior, there is a cold start problem for long tail query and AD.

2. Click Model: A unified framework--it's actually a list of all the factors.
U--User
Q--Query
A--Ad
R--Position of AD
C--click, 1 if a is clicked by U
L--The impression list
S--The click sequence
The task of clicking on the model is to predict the future click of the ad through the user's click Log, formally speaking, in the future presentation, calculate the value of P (c=1| q,a,u,r,l,s)

3. Click on the assumptions of the different levels of the model
(1) Unbiased hypothesis:p (c | q,a,u,r,l,s) = P (c|q,a)
(2) Position bias Hypothesis:p (c | q,a,u,r,l,s) = P (c|q,a,r)
(3) Depend on click Pattern:p (c | q,a,u,r,l,s) = P (c|q,a,r,s)
(4) Depend on ad externality:p (c | q,a,u,r,l,s) = P (c|q,a,r,l)
(5) Depend on user intent:p (c | q,a,u,r,l,s) = P (c|q,a,u,r)

Expand to say:

3-1:unbiased hypothesis
Click on only the query and the ad itself, regardless of any factors

3-2:position bias hypothesis
Examination hypothesis: Take the user's examine process into account, the "User Click" event is decomposed into two events, namely "User examine" and "user click". On this basis, the probability decomposition:
P (c|q,a,r) = P (e=1|r) * p (c=1|q,a,e=1)
where P (e=1|r) represents the probability of a user viewing (examine) in position R. It can be obtained by tracking the position (heat) of the person's eye on the screen, or by placing the same ad in a different location to calculate the click-through rate, but it seems to cost a lot.

3-3:depend On Click Pattern
The feeling of this PPT is that the more complex the model, the doubt really useful in practice?
Cascade hypothesis (WSDM08): is a further deepening of examination hypothesis, that is, on the basis of examination hypothesis, further assume that the user is sequential examine AD, and put this check order in the conditional probability.
Multiple-click model (WSDM09): In cascade hypothesis based on the integration of the user's multiple clicks, the implicit assumption is that in a list of results, users often have to complete the requirements through multiple clicks to meet. In particular, for an ad, the user clicks and non-click probability linear interpolation, the overall approach is similar to cascade hypothesis.
DBN (wsdm09): Apply DBN to model user examine and clicks
The above three methods, the experimental results, seemingly dbn the best, however, see the logistic as baseline, also not bad

3-4:depend on ad externality
This assumes that the relationship between AD in the ad list is also modeled, seemingly more detached from the actual application.
Temporal click Model (SIGIR09): The key assumption is that if an ad is shown together with a higher-quality AD, the ad's click-through rate will drop. Use graph model to describe this relationship.
Relational Click Predication (WSDM12): The key hypothesis, the displayed ad list, the similarity between ad will affect the ad Ctr. approach, the ad list is treated as a whole, instead of being treated separately for each ad, described in CRF.

3-5:depend on User Intent
Task Centric Click Model (KDD11): The key assumption is that the user is gradually refining his needs (through increasingly precise query) and tends to click on documents that are not present in the previous query (new document). To do with graph model.