Personalized recommendation aims to recommend candidate items to users based on their preferences by simulating their cognitive decision-making process. User-item interaction data typically follows a power-law distribution. However, existing works usually learn the representations of users and items in Euclidean space, resulting in a mismatch between the data volume space and the embedding space, which causes significant distortion in the representations. Moreover, the presence of cognitive biases, such as conformity, can also introduce distortion in representation learning. Therefore, we propose a Debiased Hyperbolic Recommendation model, called DHRec. Specifically, first, we choose to model the representations of user and item in hyperbolic space, which has exponential growth capabilities. Second, in addition to the user-item interaction graph, we also construct semantic graphs to capture the semantic neighbor information of users and items. Then, by adjusting the weights of neighbor nodes, we learn debiased representations of users and items, effectively alleviating the bias caused by conformity. Finally, we compute the predicted scores between user and candidate items in hyperbolic space. Extensive experiments on three datasets demonstrate that our model surpasses the strongest baseline, achieving a 11.04% and 10.09% improvement on Recall and NDCG, respectively.