Representation learning of knowledge bases aims to embed both entities and relations into a continuous vector space. Most existing models such as TransE, DistMult, ANALOGY and ProjE consider only binary relations involved in knowledge bases, while multi-fold relations are converted to triplets and treated as instances of binary relations, resulting in a loss of structural information. M-TransH is a recently proposed direct modeling framework for multi-fold relations but ignores the relation-level information that certain facts belong to the same relation. This paper proposes a Group-constrained Embedding method which embeds entity nodes and fact nodes from entity space into relation space, restricting the embedded fact nodes related to the same relation to groups with Zero Constraint, Radius Constraint or Cosine Constraint. Using this method, a new model is provided, i.e. Gm-TransH. We evaluate our model on link prediction and instance classification tasks, experimental results show that Gm-TransH outperforms the previous multi-fold relation embedding methods significantly and achieves excellent performance.