In service-oriented architectures, collaborative filtering is a key technique for service recommendation based on QoS prediction. Matrix factorisation has emerged as one of the main approaches for collaborative filtering as it can handle sparse matrices and produces good prediction accuracy. However, this process is resource-intensive and training must take place in the cloud, which can lead to a number of issues for user privacy and being able to update the model with new QoS information. Due to the time-varying nature of QoS it is essential to update the QoS prediction model to ensure that it is using the most recent values to maintain prediction accuracy. The request time, which is the time for a middleware to submit a user’s information and receive QoS metrics for a candidate services is also important due to the limited time during dynamic service adaptations to choose suitable replacement services. In this paper we propose a stacked autoencoder with dropout on a deep edge architecture and show how this can be used to reduce training and request time compared to traditional matrix factorisation algorithms, while maintaining predictive accuracy. To evaluate the accuracy of the algorithms we compare the actual and predicted QoS values using standard error metrics such as MAE and RMSE. In addition, we propose an alternative evaluation technique using the predictions as part of a service composition and measuring the impact that the predictions have on the response time and throughput of the final composition. This more clearly shows the direct impact that these algorithms will have in practice.