Here , denotes the Euclidean inner product,W is a window size,N( |v) a negative sampling distributionforvertexv andσ(y):=(1+e y) 1 thesigmoidfunction. Usually one takesℓP(y) = logσ(y)andℓN(y) = logσ( y).

Werecallthisresult from [4] (see Theorem 7 thereinformoredetails).

Thesecond stage then introduces unlabeled data withdisjoint classesin a semi-supervised scheme to diversify these priors and achieve generalization. We assess our method on both synthetic data and real collected point clouds.

In order to illustratethe strength of our framework, we develop 16 new methods that fit this.