Abstract: Two-sided fairness in recommender systems considers both user and item fairness. However, most of the existing two-sided fairness methods only realize one of individual fairness or group fairness on the user or item side, ignoring the complexity of realizing these two types of fairness on both sides at the same time. In addition, cross-group fairness is still not considered in two-sided fairness, an issue that has not yet received sufficient attention in existing research. To address these two issues, a new approach is proposed to balance cross-group and individual fairness on both the user side and the project side. The method identifies users in unfavorable groups by introducing sensitivity-perceived loss and achieves group fairness through a collaborative loss-balancing strategy. Meanwhile, in order to avoid the degradation of individual fairness when enhancing group fairness, individual fairness is redefined from the perspective of ranking, so that the similarity relationship of users or items in the input space (feature space) is maintained in the recommendation results, thus realizing the effective coexistence of group fairness and individual fairness. In addition, predictive score normalization is used to adjust recommendation scores to treat positive examples in different cross-groups fairly. Through experimental validation on two public datasets, the results show that the method can significantly mitigate cross-bilateral unfairness and improve individual fairness, and outperforms existing baseline methods on several assessment metrics.