Managed to find an answer to my own question : negative sampling was poorly done (mainly random links) which led the model to always have the same percentage of trust for positive and negative links. I coded my own neg sampling function and made it generate links only inside the said "gamme". Now i have around 0.88 AUC.