Role-aware Conformity Influence Modeling and Analysis in Social Networks

We study how the conformity tendency of a person changes with her role, as defined by her structural properties in a social network.

Influence: We assign a value to an edge to represent the influence propgated from one user to another.

Conformity: We assign a value to a node to represent the conformity tendency of one user to be influend.

Formalizing Conformity

Conformity theory [Bernheim 1994]: Everyone in a group expresses her own individuality. Yet, even individualists pursue somewhat for status (esteem or popularity) and change their choices toward the social norm.

We formalize conformity theory by defining a utility function. We use a binary value to represent whether a user vi adopts an action (yi = 1) or not (yi = 0). Given the decision yi , we model the utility vi obtained from her decision from two aspects. One is the individual’s intrinsic utility in the absence of all other neighbors, the other is the esteem acquired through conforming:

where represents the intrinsic preferred selection of user vi , λi represents the conformity tendency of vi , and N(i) denotes the neighbors of vi at the time when vi makes the decision. d(:; :) is a metric that gives a utility of 1 when two decisions are the same, and 0 otherwise.

The basic idea is that users’ role distribution is determined by not only attributes but also actions. We use users’ attributes to determine her role distribution, which is then used as a prior to guide the sampling process for users’ actions. Specifically, the model consists of two parts. The first part models the generation of individual attributes. For an individual attribute, we first draw a role r from a multinomial distribution, and then draw the value of the attribute from a normal distribution with respect to r . The second part models the total utility of generating all the actions. Specifically, we extend the utility function in Eq. (1) to further incorporate the role and topic distributions of a user. Instead of binary actions, each user is now allowed to take a set of actions, denoted by W = {w}. When in role r , the utility function of vi taking an action w is defined as:

Note that we utilize a unified r for all the users in role r , to reduce the number of parameters.

Experiments

We design a task of word usage prediction to evaluate our role-conformity model:

• TF-IDF and PLSA only consider the intrinsic preference, and ignores the situation where a user’s topic distribution may change and become closer to her neighbors’ topic distribution over time.
• CIM (Citation influence model) directly learns the conformity tendency of each user, which becomes very difficult to be estimated accurately when very few historical actions of the user and/or her neighbors are available.
• Our model RCM clusters similar users into roles, and then learn the conformity tendency of each role.

We study the correlation between role and conformity influence, where each role is represented by a vector of the mean values of different network attributes. We discover that people with higher degree and lower clustering coefficient are more likely to conform to others. The phenonmenon can be explained as when a person collaborates with more authors and the coauthors are more structurally diverse, she may become more open-minded and tend to accept new ideas from others. However, when the social circle of the user is restricted to a few coauthors forming a dense collaboration network, the person will be more conservative and tend not to accept other ideas.

[Download]

We collect the data sets as follows.We first select eight domains from computer science, including database and data mining (DB&DM), human computer interaction (HCI), system and high performance computing (HP), software engineering (SE), computational theory (CT), artificial intelligence and machine learning (AI&ML), computer networks (CN), as well as computer vision and multimedia (CV&MM). For each domain, we then collect all the papers from the well-known journals and conferences in the domain and the citation relationships among them. There are in total 231,728 papers, 269,508 authors and 347,735 citation relationships, where each author has on average 3.44 papers and each paper has on average 1.68 citation relationships.

========Data Format=========
We collect papers and corresponding citations from eight domains. For each domain, we generate the training set containing the papers published in or before 2009, and the test set contains the papers publihsed in or before 2009. The information of each paper is separated by a blank line in both the training and test set. The format of each paper is defined as follows:
paper id
published year
published conference or journal
terms in title (seperated by blank space)
author id1|author name1%author id2|author name2 ....
citation id1|citation id2....
Sample data:
2860908
2010
International Journal of Computer Vision
improv bag feature large scale image search
1283518|herve jegou%432449|matthijs douze%606909|cordelia schmid
293662 907781 907739 908562
========Code=========
The code contains the implementation of our model RCM and two baselines PLSA and CIM.
Please execute run.sh to run the code.
Input parameters:
-K int: number of topics
-R int: number of roles
-maxnum int: maximal number of iterations
-eps double: stop condition: change difference of log-likelihood
-printstep int: number of iterative steps to print parameters
-twords int: number of top ranked words to print

The data and code can be download from [here].

Jing Zhang, Jie Tang, Honglei Zhuang, Cane Wing-Ki Leung, and Juanzi Li. Role-aware Conformity Influence Modeling and Analysis in Social Networks. In Proceedings of the 28th AAAI Conference on Artificial Intelligence (AAAI'14). [PDF]