Election results prediction using the social network neighbourhood information is a complex network theory problem which includes advanced complex network modelling and complex network analysis techniques. Presuming that social networks display a high degree of homophily, which in binary elections leads to polarization and a split into two giant clusters, we translated this problem into a domain of cluster size inference, for these two clusters.
First, we have analysed how the social neighbourhood is related to a person's voting preference. Next, we have shown that the prediction of this preference can be done from the neighbourhood information in a straightforward manner. Finally, we have shown how, by modelling social networks with the expanded stochastic Kronecker graph model and using our novel Them vs. us estimator, network's vote share can successfully be estimated, even from a small and highly biased node sample.
Author: Ivan Bestvina
Mentor: Mile Šikić, PhD
Supplement project code demonstrating the election results estimation from the biased network samples is presented in the performance analysis notebook. Code here is slow, but easily readable, following the algorithms described in the thesis directly. Its purpose is to provide a deeper insight into the methods used, and to show how the performance graphs are generated from the simulations.
To execute simulations on a large scale, use the higly optimized code found in simulations folder:
python run_sims.py net_id
where net_id is the id of the network you want to simulate, specified in the nets_to_sample.csv. For a better understanding of the network and simulation parameters, consult the thesis.