Contrastive Training for Models of Information Cascades