Learning from the Dark: Boosting Graph Convolutional Neural Networks with Diverse Negative Samples