Toward Highly-Efficient and Accurate Services QoS Prediction via Machine Unlearning

Abstract

Personalized Internet of Things (IoT) services prediction based on Quality-of-Service (QoS) is an indispensable technique for selecting appropriate services for each user. However, existing collaborative prediction models do not take into account the user’s authority to manage their own generated data. From the standpoint of users, the expectation is for models to eliminate the impact of their sensitive data to the greatest extent possible. Meanwhile, IoT service providers face the challenge of data contamination during service provision, which necessitates models to forget data quickly and accurately to restore performance. Furthermore, existing QoS prediction methods usually suffer from low model availability when handling unlearning requests by full retraining. This underscores the need to address security, availability, fidelity, privacy, and related issues, highlighting the urgency of unlearning. To solve the problem, we propose Context-Aware Data Driven Eraser (CADDEraser), a novel efficient machine unlearning framework for QoS prediction tasks. Firstly, we divide the training data into multiple shards to train submodels and obtain node embeddings by utilizing contextual information to derive graph embeddings. Then these embeddings are employed in a balanced clustering partition, ensuring the preservation of the QoS record between users and services. Finally, we use a concatenate aggregation method and stacking & attention-based aggregation methods to synthesize information from sub-models more efficiently. Experiments on large-scale datasets show that our CADDEraser framework not only improves efficiency but also enhances the accuracy of QoS prediction, achieving efficient unlearning and outperforms state-of-the-art unlearning approaches. Source codes are available at https://github.com/ZengYuXiang7/CADDEraser.

Keywords

• Internet of Things
• machine unlearning
• context-aware
• QoS prediction