Selective harvesting over networks

摘要: Active search on graphs focuses collecting certain labeled nodes (targets) given global knowledge of the network topology and its edge weights (encoding pairwise similarities) under a query budget constraint. However, in most current networks, nodes, topology, size, are all initially unknown. In this work we introduce selective harvesting, variant active where next node to be queried must chosen among neighbors set; available training data for deciding which is restricted subgraph induced by set (and their attributes) (without any or attributes). Therefore, harvesting sequential decision problem, decide at each step. A classifier trained scenario can suffer from what call tunnel vision effect: without recourse independent sampling, urge only promising forces classifiers gather increasingly biased data, show significantly hurts performance methods standard classifiers. We demonstrate that it possible collect much larger targets using multiple classifiers, not combining predictions as weighted ensemble, but switching between used step, way ease effect. discover collects more (a) diversifying (b) broadening choices future. This highlights an exploration, exploitation, diversification trade-off our problem goes beyond exploration exploitation duality found classic problems. Based these observations propose D $$^3$$ TS, method based multi-armed bandits non-stationary stochastic processes enforces diversity, outperforms competing five real datasets evaluation exhibits comparable other two.