Liu K and Tang L. CIKM'11

This is a scientific paper authored by Liu Kun and Lei Tang, and appeared in CIKM'11. Below is the paper summary written by Tuan Anh.

Citation

@inproceedings{Liu:2011:LBT:2063576.2063838,

author = {Liu, Kun and Tang, Lei},
title = {Large-scale behavioral targeting with a social twist},
booktitle = {Proceedings of the 20th ACM international conference on Information and knowledge management},
series = {CIKM '11},
year = {2011},
isbn = {978-1-4503-0717-8},
location = {Glasgow, Scotland, UK},
pages = {1815--1824},
numpages = {10},
url = {http://doi.acm.org/10.1145/2063576.2063838},
doi = {10.1145/2063576.2063838},
acmid = {2063838},
publisher = {ACM},
address = {New York, NY, USA},
keywords = {advertising, behavioral targeting, large-scale data mining, social targeting, social-network analysis}

}

Online Version

Large-Scale Behavioral Targeting with a Social Twist.

Summary

This paper reports experiments on assessing the effectiveness of supervised and unsupervised learning methods in leveraging social data in behavioral targeting problem. The experiments were conducted on a large scale dataset that contains information about users' activities and their social networks from ``a large IT company" (mostly Yahoo) in 14 weeks. User activities include users' web-browsing behavior such as the page visiting, ad clicking, searching, etc. User network is built based on instant messages exchanged among users. The resulting dataset includes of 180 million users and their behaviors in 60 consumer domains. To deal with the large scale of the dataset, all the experiments are implemented in Hadoop MapReduce framework.

The problem is formulated as a prediction task: given users'activities and their network first 10 weeks, predict users' activities in last 4 weeks. The key findings are as follows.

• Knowledge about social network is helpful but mainly in the categories where users have strong homophily
• Social network features should be used incorporating with other standard features

Dicussion

This paper investigate the use of social network in improving online advertising. However the social features used in this papers are mostly structure and static features, e.g., neighborhood and community. These features do not represent the temporal dynamic of the social network and that's why they do not help much in predicting user activity (the performance improvement is less than 5%).

In this paper, the predictive models are built for each category of user activity independently. The prediction performance may be improved if we investigate the correlation between the categories, e.g., a user clicking on tour ads will more likely to click on cheap flight ads.

Related papers

There have been a number of works on leveraging user network in predicting user behaviors. Below are some of representative works as I know.

• Hao Ma. et. al. SoRec: social recommendation using probabilistic matrix factorization. CIKM'08
• Sanjay Purushotham et. al. Collaborative Topic Regression with Social Matrix Factorization for Recommendation Systems
• Freddy Chua et. al. Generative Models for Item Adoptions Using Social Correlation. IEEE TKDE, 05 July 2012