This is an official PyTorch implementation of paper entitled "QoSEraser: A Data Erasable Framework for Web Service QoS Prediction".
To select appropriate cloud services for users, the Quality-of-Servic (QoS) based collaborative prediction models are widely used. Despite the success of collaborative prediction models in selecting appropriate cloud services for users, existing models do not take into account the user’s authority to manage their own generated data as stipulated in GDPR. Besides, due to the security concerns such as data poisoning attacks, the unlearning is urgently needed. Existing QoS prediction methods did not optimize for unlearning, suffering from low model availability when handling unlearning requests by full retraining.
To solve the problem, we propose QoSEraser, a novel efficient machine unlearning framework for QoS prediction tasks. The central concept of the QoSEraser involves (1) dividing the training data into multiple shards to train submodels, and obtaining node embeddings by utilizing contextual information to derive graph embeddings. These embeddings are then employed in balanced clustering slicing, ensuring the preservation of the QoS record between users and services, our framework not only improves time performance but also enhances the accuracy of QoS prediction; (2) a concatenate aggregation method and stacking & attention-based aggregation method are used to synthesize information from submodels more efficiently. Experiments on large-scale datasets show that QoSEraser achieves efficient forgetting learning and outperforms state-of-the-art unlearning approaches in performance.