<p>The knowledge distillation (KD) methods have been demonstrated to be an effective way to improve model performance through teacher-student training strategies. They also show excellent performance in mobile application recommendation, but they are easily affected by app popularity, causing bias problems in recommendation. To improve the efficiency and eliminate bias of recommendation system, an unbiased stratified knowledge distillation-based mobile application recommendation method is proposed in this paper. Firstly, it trains the teacher model to encode the learned knowledge into soft labels. Secondly, apps are grouped based on app popularity, where the popularity of apps within each group is similar and the total popularity is equal. Then, the apps are ranked in descending order based on the teacher’s soft labels. Finally, the ranking knowledge is extracted from each group and used for the training of the student model. This training method comprehensively considers apps with different popularity levels, thereby achieving the debiasing effect. We conducted a series of comparative experiments using a dataset comprising 29,650 users, 3078 apps, and 132,733 user-app interactions. Our method achieved an average improvement of 4.67% in recommendation accuracy compared to traditional methods, while significantly reducing bias in recommendation results. These findings demonstrate the effectiveness and unbiasedness of our proposed method in mobile application recommendation.</p>