Computing Reviews
Today's Issue Hot Topics Search Browse Recommended My Account Log In
Review Help
Search
Imputation of missing links and attributes in longitudinal social surveys
Ouzienko V., Obradovic Z. Machine Learning95 (3):329-356,2014.Type:Article
Date Reviewed: Apr 17 2015

Social network surveys, in which people are asked to identify their friends, are important tools for social scientists for studying social settings and phenomena such as villages, urban settings, street gangs, and crime. Such surveys are often conducted in waves over time, in which case a set of surveys is referred to as a longitudinal social survey of the target population. A significant portion of social network surveys has some part of the input data missing, and reaching the subjects who are not responsive to the surveys can be prohibitively expensive and difficult. Therefore, “imputation,” that is, filling in missing data in longitudinal social surveys, has been a key focus of recent research.

This problem is closely related to the task of predicting unobserved links, which has also received much attention in computer science and physics recently. The possibility of having multiple nonrespondents causes the dyadic relationships (in which both entities involved in that relationship do not respond to the survey) to be very hard to infer. Further, besides simply predicting missing links, the problem can be generalized to predict the attributes (for example, closeness) of the links.

Traditional imputation approaches, such as reconstruction and preferential attachment, are not able to utilize the temporal nature of longitudinal social surveys. Each wave leads to a time-specific sociomatrix that is not independent from the other sociomatrices collected for the same group of people. Therefore, imputation techniques that do not use the temporal nature of such surveys may produce values that are not optimal, and possibly even inconsistent with other sociomatrices collected from the same group of people. This deficiency, among others, has created an opening for newer and more sophisticated imputation techniques to be studied.

In this paper, the authors propose a new method based on the exponential random graph model. This approach, called ITERGM, uses the temporal aspect of the surveys and can be applied to both link and attribute prediction. ITERGM is, in essence, an expectation maximization algorithm over two Markov chain Monte Carlo inferences. This is an iterative approach that learns the parameters of two models in each iteration and stops when the weights of the two models have converged or are within a certain threshold. Experimental results suggest that ITERGM is at least ten percent more accurate than the best of the previously known algorithms, and in many cases it is 30 percent to 40 percent more accurate than the previously known algorithms.

A natural follow-up question to this work is as follows: knowing the power of the state-of-the-art imputation algorithms, can the longitudinal social surveys be designed so as to minimize the cost and still achieve the social outcomes and conclusions that they set out to achieve?

Reviewer:  Amrinder Arora Review #: CR143363 (1507-0605)
Bookmark and Share
 
Probability And Statistics (G.3 )
 
 
Network Monitoring (C.2.3 ... )
 
 
Public Networks (C.2.3 ... )
 
 
Social And Behavioral Sciences (J.4 )
 
Would you recommend this review?
yes
no
Other reviews under "Probability And Statistics": Date
Probabilities from fuzzy observations
Yager R. (ed) Information Sciences 32(1): 1-31, 1984. Type: Article
Mar 1 1985
Randomness conservation inequalities; information and independence in mathematical theories
Levin L. Information and Control 61(1): 15-37, 1984. Type: Article
Sep 1 1985
The valuing of management information. Part I: the Bayesian approach
Carter M. Journal of Information Science 10(1): 1-9, 1985. Type: Article
May 1 1986
more...

E-Mail This Printer-Friendly
Send Your Comments
Contact Us
Reproduction in whole or in part without permission is prohibited.   Copyright 1999-2024 ThinkLoud®
Terms of Use
| Privacy Policy