Introduction
In social networks, a small subset of users usually plays an important role in influencing the social dynamics, and users may have different representative degrees from different aspects (topics). We use an example from a coauthor network to clearly demonstrate the motivation of this work.
In this work, we study the problem of sampling representative users to statistically represent the original social network. We formally define the representative user finding problem and theoretically prove its NP-hardness by linking it to the Dominating Set Problem. We design a quality function Q to evaluate the selected set of representative user and develop two instantiation models for the Q functions. To approximate the optimal solution, we further present eddicient algorithms to solve the two models with provable approximation guarantees. Experimental results show that the proposed models for finding representative users significantly outperform (+3.6%-12.9% in terms of F1-measure) several alternative methods using authority or structure information only. The proposed algorithms are also effective in terms of time complexity.
This work has been accepted in AAAI15. If you use data sets provided or refer to the model, please cite:
@INPROCEEDINGS{Tang:15AAAI,
AUTHOR = "Jie Tang and Chenhui Zhang and Keke Cai and Li Zhang and Zhong Su",
TITLE = "Sampling Representative Users from Social Network",
BOOKTITLE = "AAAI'15",
YEAR = {2015}
}
Dataset
[download]
We preform excperiments on two author datasets: DB and DM. Each conists of authors and coauthor relationships extracted from major conferences in a specific domain. Specifically, we extract the PC dataset from two domains, Database and Data Mining. For the Database domain, we extract all authors from papers published for the three major conferences (SIGMOD, VLDB, ICDE) during 2007-2009. In total, we collect 8,027 authors and 23,770 coauthor relationships. For the Data Mining domain, we similarly extract authors from papers published for the three major conferences (SIGKDD, ICDM, CIKM) during 2007-2009. In total, we collect 6,394 authors and 12,454 coauthor relationships.
The author datasets can be downloaded here.
Model & Code
[download]
We first give the formal definition of the representative users finding problem below:
We develop two instantiation sampling models for the problem.
Model 1: Statistical Stratified Sample (S3)
For each attribute group, we assign a score ml to indicate the "importance" of this group. The groups with higher ml value should be better represented. It is referable to have some "bias" to select user who can represent large attribute groups and give the following quality function:
Model 2: Strategic Sampling for Diversity (SSD)
In this model our goal is the diversification of delegates. The "diversity" means not being centralized in this case. The objective is that all these values can be distributed properly (for example, where they are all equal or the same proportion to the size of each group). We give the quality function as follow.
Experiments
We take program committee (PC) members of the conferences in each dataset during 2007-2009 as the representative users. Our goal is to predict who are representative users (PC members of the three major conferences) for the two research domains, respectively. Finally, we collect 291 representative users (by removing authors who are not in the author list of papers we collected) in the Database dataset and 373 representative users in the Data Mining dataset.
We evaluate the performance of our representative users finding approaches in terms of Precision, Recall, F1-Measure.
We further compare the efficiency of the proposed methods with the baseline methods. Our methods are found to be very efficient, for the SSD model only requires 0.54s for Database, 0.74s for Data Mining and 0.67s for Weibo in program domain, and is faster than the S3 model.
