Abstract: Existing models ignore the problem that the user's trust dependence on friends in social relationships will change with interest groups. To this end, a social recommendation method, Social-TD, is proposed. The algorithm extracted user features in different interest groups from a single user feature by a slice layer, and used a graph neural network to learn the trust dependence of users in different interest groups on friends. In addition, the slice layer allowed only the user feature in the interest group of the current item to be considered when the item was modeled, which effectively reduced the noise information in the propagation process. Experiments on two public datasets show that the Social-TD algorithm performs better than others in predictive accuracy, which verifies the effectiveness of learning user trust dependence through interest groups.