There are a lot of points, each point has many properties, between points on the physical connectivity should be fixed and determined, the pipeline, for example,
But the problem is: with the properties of the neural network learning point to speculate that the connectivity, because neural network learning parameter is random, every time the inferred results are becoming, that is to say, for a while connecting the two points, the next training becomes the impassability,
Excuse me is there any good ways to solve this problem,

PS: the point of the attribute's change over time, is there anyone according to the time attribute to link prediction on the point of related articles,