Graph sharpening

摘要: In many graph-based semi-supervised learning algorithms, edge weights are assumed to be fixed and determined by the data points' (often symmetric) relationships in input space, without considering directionality. However, may more informative one direction (e.g. from labelled unlabelled) than reverse direction, some strong between oppositely points) unhelpful either direction. Undesirable edges reduce amount of influence an point can propagate its neighbours - outgoing have been ''blunted.'' We present approach ''sharpening'' which adjusted meet optimization criterion wherever they directed towards points. This principle applied a wide variety algorithms. this paper, we solution satisfying principle, order show that it improve performance on number publicly available bench-mark sets. When tested real-world problem, protein function classification with four vastly different molecular similarity graphs, sharpening improved ROC scores 16% average, at negligible computational cost.