This paper introduces BPR (Bayesian Personalized Ranking), a method for personalized ranking using implicit feedback. The key contribution is the BPR-OPT optimization criterion derived from Bayesian analysis, which maximizes the posterior probability of the desired user preference order. The learning algorithm LEARNBPR is proposed, based on stochastic gradient descent with bootstrap sampling. The method is applied to two state-of-the-art recommender models: matrix factorization and adaptive kNN. Experiments show that BPR outperforms standard learning techniques for these models in personalized ranking tasks. The results highlight the importance of optimizing models for the correct criterion. BPR-OPT is shown to be equivalent to maximizing the area under the ROC curve (AUC) for ranking. The method is evaluated on two datasets: Rossmann (an online shop) and Netflix (DVD rental). The results demonstrate that BPR-based models achieve higher AUC scores than other methods, including non-personalized ranking approaches. The paper also discusses the relationship between BPR and other methods like WR-MF and MMMF, showing that BPR is more suitable for personalized ranking. The results confirm that optimizing for the correct criterion is crucial for achieving high prediction quality in personalized ranking tasks.This paper introduces BPR (Bayesian Personalized Ranking), a method for personalized ranking using implicit feedback. The key contribution is the BPR-OPT optimization criterion derived from Bayesian analysis, which maximizes the posterior probability of the desired user preference order. The learning algorithm LEARNBPR is proposed, based on stochastic gradient descent with bootstrap sampling. The method is applied to two state-of-the-art recommender models: matrix factorization and adaptive kNN. Experiments show that BPR outperforms standard learning techniques for these models in personalized ranking tasks. The results highlight the importance of optimizing models for the correct criterion. BPR-OPT is shown to be equivalent to maximizing the area under the ROC curve (AUC) for ranking. The method is evaluated on two datasets: Rossmann (an online shop) and Netflix (DVD rental). The results demonstrate that BPR-based models achieve higher AUC scores than other methods, including non-personalized ranking approaches. The paper also discusses the relationship between BPR and other methods like WR-MF and MMMF, showing that BPR is more suitable for personalized ranking. The results confirm that optimizing for the correct criterion is crucial for achieving high prediction quality in personalized ranking tasks.