Abstract: Web service recommendation based on QoS awareness is one of the solutions to help users find high-quality services. In order to accurately predict the QoS value of candidate services, it is usually necessary to collect historical QoS data of users. However, most of the existing methods ignore the outliers in the historical data, which will lead to lower prediction accuracy. In order to solve this problem, a service quality prediction method based on outlier detection matrix decomposition is proposed. Firstly, K-means based isolated forest algorithm is used to cluster historical QoS data, and then outliers in historical data are eliminated. Then it is used to predict the unknown value in the improved matrix decomposition model. Finally, Cauchy loss is used to evaluate the difference between the observed value and the predicted value. The experiment uses WSDream data set to test, and the results show that the MAE and RMSE indexes of the response time of the proposed outlier detection model are improved by 19.11% and 39.59% on average, and the MAE and RMSE indexes of the throughput are improved by 9.82% and 29.89% on average, which proves that the proposed model effectively improves the prediction accuracy.