Click Models for Mobile Search

Thanks for visiting.

This project focuses on building click models for mobile search.

It is based on the project by Aleksandr Chuklin (https://github.com/varepsilon/clickmodels).

Published under the BSD license.

How to install

git clone https://github.com/THUIR/click_model_for_mobile_search.git
cd mobile_click_model
python setup.py install

Models Implemented

• Mobile Click Model (MCM): Jiaxin Mao, Cheng Luo, Min Zhang, and Shaoping Ma. 2018. Constructing Click Models for Mobile Search. SIGIR (2018).
• Viewport Time Click Model (VTCM): Not yet published.
• Dynamic Bayesian Network (DBN) model: Chapelle, O. and Zhang, Y. 2009. A dynamic bayesian network click model for web search ranking. WWW (2009).
• User Browsing Model (UBM): Dupret, G. and Piwowarski, B. 2008. A user browsing model to predict search engine click data from past observations. SIGIR (2008).
• Exploration Bias User Browsing Model (EB-UBM): Chen, D. et al. 2012. Beyond ten blue links: enabling user click modeling in federated web search. WSDM (2012).
• Dependent Click Model (DCM): Guo, F. et al. 2009. Efficient multiple-click models in web search. WSDM (2009).
• A version of User Browsing Model considering different result layouts (UBM-layout): Aleksandr Chuklin, Pavel Serdyukov, and Maarten de Rijke. 2013. Using intent information to model user behavior in diversified search. ECIR (2013)

Format of the Input Data (Click Log)

A small example can be found under data/train. This is a tab-separated file, where each line has 10 elements. A record of a query session is shown as follows:

0032202F707D59045B51569FA5952795#1#1533349671#5#0	快递单号查询	0	0.0	["http://www.guoguo-app.com", "https://m.ickd.cn/", "http://m.ickd.cn/", "http://m.kuaidi100.com/", "http://m.kuaidi100.com/", "http://www.chawuliu.cn/", "http://www.17ckd.com/", "http://wap-hint/", "http://www.sogou.com/", "http://m.aikuaidi.cn", "http://wap.yto.net.cn/", "http://www.sf-express.com/mobile/cn/sc/index.html", "http://www.sogou.com/"]	["21194401", "-1", "-1", "-1", "30000909", "-1", "-1", "30010081", "50023801", "-1", "-1", "30000909", "50023901"]	[1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0]	[30000, 30000, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0]	[20333, 2055, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0]	[345, 157, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1]

1. 0032202F707D59045B51569FA5952795#1#1533349671#5#0 contains the hashid and beginning timestamp of the session (currently not used in MCM and VTCM).

2. 快递单号查询 is the query of the session.

3. A integer identifier which represent the user's region (currently not used in MCM and VTCM).

4. The probability $P(I = V)$ that the user has a vertical intent $V$ (currently not used in MCM and VTCM).

5. A list of the URLs of the documents that make up SERP (search engine result page).

6. A list with the presentation types (vertical types) of the documents.

7. A list of user clicks.

8. A list of raw viewport time.

9. A list of viewport time weighted by the result height and viewport height.

10. A list of result heights.

Usage

Click prediction

You can train and test a click model with given example data:

python test_click_models.py ../data/train ../data/test -m MCM-VPT -N 1 -M 1 -o ../mcm_mr --ignore_no_clicks --ignore_no_viewport --viewport_time

• -m: The click model that you would like to run, such as MCM-VPT (VTCM), MCM, and other popular click models.

• -N: The number of files that you would like to use for training in the train dictionary.

• -M: The number of files that you would like to use for test in the test dictionary.

• -o: The path to output dictionary.

• --ignore_no_clicks: Ignore the sessions that don't contain any click.

• --ignore_no_viewport: Ignore the sessions that don't contain viewport time information

• --viewport_time: Read viewport time information of train and test dataset.

• -V: The viewport time model used in VTCM (used with -m MCM-VPT).

  • 0: VTCM$_e$ with log-normal
  • 1: VTCM$_e$ with gamma
  • 2: VTCM$_e$ with Weibull
  • 3: VTCM$_c$ with log-normal
  • 4: VTCM$_c$ with gamma
  • 5: VTCM$_c$ with Weibull

Relevance estimation

You can get the relevance scores estimated by click models based on the outputs of click prediction:

python get_ranking_from_relevance_estimation.py ../mcm_mr/relevance_estimation.txt ../mcm_mr/ranking_relevance_estimation.txt -n 10 -m MCM

Information

Authors

Yukun Zheng, Tsinghua University

Jiaxin Mao, Tsinghua University

Citations

For MCM, please cite:

Jiaxin Mao, Cheng Luo, Min Zhang, Shaoping Ma. Constructing Click Models for Mobile Search. The 41st International ACM SIGIR Conference on Research and Development in Information Retrieval. (SIGIR 2018)

Contact

Please contact Yukun Zheng by the following email if you have any question:

zhengyk13 at gmail.com