Fig. 8

Feature weights over time averaged over 20 runs of NOL-HTR on a BA and b BTER networks. In the BA network, the lost reward feature is weighted highly, consistent with many queried nodes connecting to a hub node before it is queried. In the BTER network, which is made up of Erdos-Renyi graphs with sizes following a heavy tail, the size of the connected component a node connects to is important for predicting reward, indicating preference for the largest community. Insights like these can help in both guiding feature selection and helping understand the structure of a partially observed network