Abstract: Sequence recommendations are designed to predict a user's next interaction by modeling the user's historical interaction sequence. At present, the improved series recommendation model based on Transformer utilizes dot product self-attention to measure the relationships in the series and has excellent performance. However, due to the inclusion of some noise in the user interaction data, the model is easy to fit the noisy data during training, resulting in overfitting. In addition, the traditional self-attention mechanism also naturally has the problem of high memory consumption and secondary computing complexity. In order to solve the above two problems, a probabilistic self-attention sequence recommendation model FPA-Rec (Filter ProbAttention Sequence Recommendation) is proposed. After the input sequence goes through the embedding matrix mapping, it first passes through a layer of learning filter structure to suppress the noise signal in the sequence and get a more pure sequence input. Then probabilistic self-attention is used instead of traditional self-attention to solve the problems of high memory consumption and secondary computing complexity in the traditional self-attention mechanism. Experimental verification results show that FPA-Rec model achieves the optimal recommendation effect. On the two data sets represented by Beauty and Yelp, the evaluation index Hit@10 increased by 1.2% and 1.7% compared with the optimal model, the evaluation index NDCG@10 increased by 2.4% and 2.0% compared with the optimal model.