From the abstract: "Increasingly, many important domains in the world can be viewed as networks of linked nodes: people connected by social network 'friendships,' webpages connected by hyperlinks, and even geo-political areas connected by proximity and common interests. To leverage these links for prediction and analysis tasks, Machine Learning researchers have developed multiple techniques for link-based classification (LBC). While LBC can substantially improve prediction accuracy in some domains, current limitations greatly restrict its applicability when used to evaluate heterogeneous domains (e.g., when the collection of 'nodes' under study are actually drawn from multiple populations). Additionally, traditional LBC predicts only categorical outputs, while link-based regression and the prediction of continuous outputs have been left largely unexplored. One such application that requires continuous outputs involves elections. Predicting the voting outcome of national or regional elections is a challenging yet important problem, and has great implications for regional and international security. [...] This study used a collaborative filtering approach to implicitly leverage the correlation present between 'nearby' regions. They did not, however, consider formulating the regions as a network. This project presents the first extension of LBC algorithms to multiple predictive 'models' and continuous outputs (thus yielding heterogeneous collective regression, HCR). To demonstrate the effectiveness of this approach, we apply it to the voting outcome prediction task [...] Overall, we demonstrate that, for the voting prediction task, HCR can be highly effective, robust to multiple choices of regression parameters and linking strategies, and computationally practical. This success opens the door to the application of HCR to other analysis tasks for link-based data."
Trident Scholar Report No. 472
Defense Technical Information Center (DTIC): http://www.dtic.mil/dtic/