Interpretable models for prediction on networks with cohesion
Abstract
Short Work Statement:P I Levina and her team shall develop novel methods to predict behavior of individuals connected by a network. Data on such networks can be obtained from social media or other sources such as intelligence or electronic surveillance. The proposed framework is intended to balance good prediction performance with interpretability of models, allowing us to assess the relationship between each predictor and the outcome of interest. This will contribute to the effort of understanding causal relationships between predictor variables and network outcomes, leading to greater understanding of dynamics of social networks and potential actionable interventions. The specific aims include developing such a framework for predicting quantitative outcomes through classical linear models (Aim 1), categorical and time to event outcomes through general loss functions (Aim 2), and extension to dynamic networks evolving overtime (Aim 3). All methods will be implemented in software tools in R, adding C and/or Python as needed, in a scalable fashion, and evaluated through extensive numerical analysis of both simulated and real networks.Objective:O ur objective is to develop a novel framework which uses simple interpretable models to do prediction on networks in many different contexts, and takes advantage of the empirically known phenomenon of network cohesion to improve predictive performance. This framework will be developed for quantitative, qualitative, and time-to-event outcomes of interest, and extended to dynamic networks evolving over time.Approach:W e will achieve the objective by combining classical predictive models that give predictors a direct causalinterpretation with regularization penalties that enforce network cohesion and therefore improve predictionperformance without losing interpretability of the model. Unlike in currently available approaches, the meaning of the relationship between predictors and response is not changed by enforcing cohesion, but prediction errors are reduced.ONR Mission/RelevanceB y developing new capabilities for the problem of predicting events occurring over social networks, this project will help to transform data collected from electronic surveillance into actionable predictions that can be used for preventative interventions or for rapid response. This framework allows to incorporate diverse multiple sources of data on units in a social network, such as their age, race, religion, education, travel history, and so on, and combine it with the allimportantdata on the actions of their social circle in automated yet informative ways. The resulting prediction tools, inaddition to making accurate predictions of future actions, will also allow for greater understanding of the mechanisms contributing to these actions, by providing easily accessible interpretations of the role of various factors in a particular action of interest. This can lead to enhanced capabilities for both long-term strategic planning and short-term tactical response, and thus contribute to increased national security.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Sep 30, 2016
- Source ID
- N000141612910
Entities
People
- Elizaveta Levina
Organizations
- Board of Regents of the University of Michigan
- Office of Naval Research
- United States Navy