Lab Home | Phone | Search
Center for Nonlinear Studies  Center for Nonlinear Studies
 Colloquia Archive 
 Postdoc Seminars Archive 
 Quantum Lunch 
 Quantum Lunch Archive 
 CMS Colloquia 
 Q-Mat Seminars 
 Q-Mat Seminars Archive 
 P/T Colloquia 
 Kac Lectures 
 Kac Fellows 
 Dist. Quant. Lecture 
 Ulam Scholar 
 CNLS Fellowship Application 
 Student Program 
 Past Visitors 
 History of CNLS 
 Maps, Directions 
 CNLS Office 
Thursday, July 19, 2012
1:00 PM - 2:00 PM
CNLS Conference Room (TA-3, Bldg 1690)


Leveraging Social Science for Prediction

Rich Colbaugh
Sandia National Laboratory

There is significant interest in developing methods for predicting human behavior, for instance to enable the outcomes of unfolding events to be forecast or the nature of ongoing but “hidden” ac-tivities to be inferred, and machine learning (ML) has proven to be a useful approach to such problems. In this talk I propose that the performance of ML algorithms can often be improved by incorporating social science concepts and models into their development and implementation, and offer three examples illustrating ways this can be accomplished. First I consider the problem of predicting whether nascent social diffusion events will ultimately propagate widely or will instead quickly dissipate. “Complex contagion” models from sociology are used to identify novel features of the diffusion phenomenon that are predictive of the diffusion’s fate, and these fea-tures are used in an ML algorithm to generate the desired forecasts. Next the task of anticipating and defending future actions of opponents in adversarial settings is addressed, and it is shown that simple game-theoretic models from economics can be combined with ML to enable the de-sign of effective defenses. Finally I examine the problem of inferring the (unobserved) nature of relationships in social networks, and demonstrate that the social psychological theory of structur-al balance can be exploited to enhance ML solutions for this task. In all cases the social science-based learning algorithms are shown to outperform “gold-standard” methods in empirical tests.

Host: Joshua Neil, 665,2121,