A user browsing model to predict search engine click ...

Click Models & Diversified Search Aleksandr Chuklin Pavel Serdyukov Maarten de Rijke

Click Models Many tasks that leverage web search users’ implicit feedback rely on a proper and unbiased interpretation of user clicks. “Efficient Multiple-Click Models in Web Search”, Guo et al.

Why Predicting Clicks?

•  Simulate user behavior when real clicks are not available (e.g. Hofmann et al. WSDM’13) •  Estimate relevance of the documents (e.g. Chapelle et al. WWW’09) •  Understanding user behavior and click biases (e.g. Dupret et al. SIGIR’08)

Clicks and Relevance

Can we say that the click probability is determined by the relevance of the document?

P (Ri = 1) = rd . ( algorithm for model inference. Cascade Hypothesis assumes that examinati the us Next, we introduce two[7] important hypothesis: Simple Click hypothesis andbottom cascade hypothesis, which areresult the foundatio linearly Models from top to of the search pag of most existing click models. documentExamination is examined only if all the Bing SERP pageashowHypothesis ([20,previous 7]) assumes docum that a do ts at different slots. examined ument is clicked and only if it is is examined and relevan and the firstifdocument always examine i

which can be formulated as:

udy the user behavior in federP (E = 1 | E = 0) = 0 i+1 i C = 1 ⇐⇒ E = 1, R = 1 ( i i i d an observation that user bes highly different from that in where Ei and Ri are independent. we c P (E1 = 1)More = precisely, 1 better characterize user behavrepresent the click-through rate as: propose a novel Bayesian model Based the cascade P (Ci = 1) = P (E P (Ci model = 1|Ei = [7] 1) co( i = 1) el (FCM), which introduces two on the hypothesis, ! "# $ ! "# $ a user will continue the examination the fir ure the distinctive userthat behavposition bias documentuntil relavance first illustrates that users tend and then she abandons the whole search session: results and the visual attention where P (Ri = 1) = P (Ci = 1 | Ei = 1) indicates t xamination probability of other probability of click after examination. [20] assumes that t P (C =on rposition i = 1 | E i = 1) di otivates us to reconsider the exexamination probability depends solely i, a h document in federated search thus P (Ci = 1) = λi rdi , in which λi models the positi P (E = 1 | E = 1, C ) = 1 − C i+1 i i i del beyond the cascade hypothbias.

•  Examination model:

•  Cascade model:

The dependent click model (DCM) [14] generalizes cade model to allow multiple clicks: P (Ei+1 = 1 | Ei = 1, Ci = 1) = λi .

Probabilistic Graphical Models

“Efficient Multiple-Click Models in Web Search”, Guo et al. WSDM’09

Diversified Result Page

Diversified Result Page (Fresh Block)

Challenges for Click Prediction

•  Some result items may be more attractive than others •  Attractive results might appear in different positions (not necessarily top/bottom) •  Different users may have different intents

Solution: Explicitly Add Intent into Model

Fig. 2: The graphical model for UBM-IA. Gray squares correspond to observed va ables, blue circles to hidden variables. Arrows show dependency links.

The aim our work is that not towe studyhave how toafind intents corresponding to the que 1. ofAssume prior intent nstead, given that we know the query intent spectrum, we aim to investigate the eff distribution P(I) of this distribution on the users’ click-through behavior. So we assume that for ea 4 session2.  we Recalculate have a prior distribution of the intents P (I = i). Importantly, unlike click probability using Bayes rule: et al. [17] we do not assume that our intent distribution is fixed for the session. Wh predicting the next click, we modify the intent distribution P (I) using Bayes’ rule: X P (Ck |C1 , . . . , Ck 1 ) = P (Ck |C1 , . . . , Ck 1 , I) · P (I|C1 , . . . , Ck 1 ) | {z } | {z } I

probability from single intent model

posterior intent distribution

Dupret and Piwowarski [13] find that the single browsing model outperforms a mixt of browsing models when inferring intent distribution from clicks. We show that ng layout information and prior knowledge of intent distribution, we can significan outperform the single browsing model.

UBM click model •  C – click •  E – examination

Ek γ kd

Ek indicates whether the user looked at the document at rank k (hidden variables C C C Ck indicates whether the user clicked on the k-th document (observed variables). ... 1

k −1

k

a uI

rder to define a click model we need to denote dependencies between these var s. For example, for the UBM model we define k

P (Ek = 1 | C1 , . . . , Ck

1)

Ek = 0 ) C k = 0

=

P (Ck = 1 | Ek = 1) = auk ,

kd

(1

(2

(3

re kd is a function of two integer parameters: the current position k and the distanc UBM model by Dupret et al. SIGIR’08 e rank of previous click d = k P revClick = k max{j | 0  j < k & Cj = 1

Intent-Aware Version: UBM-IA Gk Ek

C1

...

Ck

C k −1 γ kd (G k , I )

a uI I

k

•  C – click •  G – document type (“layout”): fresh or web •  E – examination •  I – user intent: fresh or web

Based on UBM model by Dupret et al. SIGIR’08

Experiments

Experimental Setup

•  Result pages with fresh results sampled from Yandex query log in July 2012 •  30 days, ~15*106 sessions, ~3*106 queries •  Train click model on a day j, test on a day j + 1 •  Report average perplexity of predicting clicks

Table 1: Average perplexity gain for the combined UBM-IA model. Model Average Perplexity Gain Results

Confidence Interval (Bootstrap)

Table Table 1: 1: Average Average perplexity perplexity gain gain for for the the combined combined UBM-IA UBM-IA model. model. M-IA vs. UBM % gain for the combined [1.25%, 1.43%] Table 1: Average1.34 perplexity UBM-IA model. Model Average Model Average Perplexity Perplexity Gain Gain Confidence Confidence Interval Interval (Bootstrap) (Bootstrap) Model

UBM-IA vs. UBM UBM-IA vs. UBM UBM-IA vs. UBM

Average Perplexity Gain

1.34 % 1.34 % 1.34 %

Confidence Interval (Bootstrap)

[1.25%, 1.43%] [1.25%, 1.43%] [1.25%, 1.43%]

1.5 %

1.5 % 1.51.5 %%

1.1 1.5 % 1.5 % 1.5 %

%

1.1 % 1.11.1 %%

1.1 % % 1.1 % 1.1 0.7

%

0.7 % 0.70.7 %%

0.7 % 0.7 0.7 % %

0.3 % 0.30.3 %%

0.3 % 0.3 0.3 % %

-0.1 -0.1 % combined -0.1 %%

0.3 %

-0.1 %

layout

combined combined combined

intents only

layout layout layout

intents intentsonly only

intents only

(a) Fresh results withresults special (a) Fresh withsnippets special snippets

(a) (a) Fresh Fresh results results with with special special snippets snippets

combined

combined combined

combined

layout

%-0.1 intents-0.1 only -0.1 % %

layout only only layout intents intents

layout

intents only

(b)(b)Fresh results with ordinary Fresh results with ordinary snippetssnippets (b) (b) Fresh Fresh results results with with ordinary ordinary snippets snippets

Fig. 3: Perplexity gains for layout and intent models models compared to UBM. ig. 3: Perplexity gains for layout and intent compared UBM. Fig. UBM. Fig. 3: 3: Perplexity Perplexity gains gains for for layout layout and and intent intent models models compared compared to toto UBM.

then test these ideas separately and see what their contribution is. We call the resulting then test these ideas and see their contribution is. We the resulting then test these ideas separately separately andwhat see what what their contribution is. (5), We call thethe resulting click models UBM-layout andsee UBM-intents; they are defined using (9), (10) andresu these ideas separately and their contribution is. Wecall call

Results: Perplexity for Different Positions

5%

combined layout only intents only

4% 3% 2% 1% 0% -1 % 1

2

3

4

5

6

7

8

9

10

Comparison to Other Models 1.5 % 1.2 % 0.9 % 0.6 % 0.3 % 0% UBM-IA vs. UBM

EB_UBM vs. UBM

EB_UBM-IA vs. UBM

EB_UBM-IA vs. EB_UBM

•  UBM – model that does not consider vertical results •  UBM-IA – current work •  EB_UBM – exploration bias model by Chen et al. (WSDM’12)

Conclusion

Why Predicting Clicks (reminder)

•  Simulate user behavior when real clicks are not available •  Estimate relevance of the documents •  Understanding user behavior and click biases

Intent-Aware Click Models

•  Any existing click model can be modified to use intent and layout information for better click prediction •  The model allows to infer per-topic relevance from the click logs •  Per-topic analysis of user behavior and click biases (e.g. user patience)

Code available at

github.com/varepsilon/clickmodels

References 1.  Hofmann, K. et al. Reusing historical interaction data for faster online learning to rank for IR. WSDM (2013). 2.  Chapelle, O. and Zhang, Y. A dynamic bayesian network click model for web search ranking. WWW (2009). 3.  Dupret, G. and Piwowarski, B. A user browsing model to predict search engine click data from past observations. SIGIR (2008). 4.  Guo, F. et al. Efficient multiple-click models in web search. WSDM (2009). 5.  Chen, D. et al. Beyond ten blue links: enabling user click modeling in federated web search. WSDM (2012).