HTG: A heterogeneous topology aware model to improve cold start in cloud service QoS prediction

Abstract

The massive deployment of cloud services makes it a significant challenge to guarantee end‐to‐end QoS, QoS prediction is used to address this challenge. Accurate QoS prediction values can help select quality services and sense the network situation. Most existing QoS prediction methods focus only on improving prediction accuracy in the warm‐start situation. However, they do not consider the frequent cold‐start phenomenon, hindering their applicability in real‐world scenarios. In addition, previous approaches have made great progress in leveraging contextual information, such as geographic regions, but have neglected the process of service invocation. This article proposes a Heterogeneous Topology aware model based on Graph neural network (HTG) for QoS prediction. In HTG, users/services and invocation processes are modeled as a heterogeneous communication subgraph. HTG captures subgraph neighborhood features of users and services based on a predefined node information aggregation approach. Multi‐task learning is introduced to enhance HTG's generalization ability and simultaneously predict multiple QoS attributes. For new users or services, HTG generates their representations by fusing their attributes with geographically adjacent neighbors' features. Then, these representations of new users/services are fed into the well‐trained HTG to obtain their related QoS prediction values, without relying on additional training. Extensive experimental results demonstrate that HTG not only outperforms the current state‐of‐the‐art methods in the warm‐start situation but also significantly improves the prediction accuracy in the cold‐start situation. In the warm‐start situation, HTG improves the state‐of‐the‐art method by an average of 11.47% and 9.08% for MAE in different matrix densities on the task of throughput and response time prediction, respectively. As for the cold‐start situation, these gains are 31.76% and 32.17%.