Semi-Supervised Learning with Heterophily.

摘要: We derive a family of linear inference algorithms that generalize existing graph-based label propagation by allowing them to propagate generalized assumptions about "attraction" or "compatibility" between classes neighboring nodes (in particular those involve heterophily where "opposites attract"). thus call this formulation Semi-Supervised Learning with Heterophily (SSLH) and show how it generalizes improves upon recently proposed approach called Linearized Belief Propagation (LinBP). Importantly, our framework allows us reduce the problem estimating relative compatibility from partially labeled graph simple optimization problem. The result is very fast algorithm -- despite its simplicity surprisingly effective: we can classify unlabeled within same in time as LinBP but superior accuracy not knowing compatibilities.