随着云计算技术的普及，云服务数量指数级增长，用户不再满足于功能性需求，服务质量(Quality of Service，QoS)成为比较服务优劣的关键性能指标。如何在动态、复杂的云环境中实时、准确地预测服务质量并为用户推荐高质量服务成为热点问题。考虑到云服务器的负载、网络状态、用户接入云环境的偏好等随着时间变化，本文提出了基于多源特征和多任务学习的时序QoS预测方法(T-MST)，它可以实时、准确地同时预测多种QoS属性。首先，T-MST对用户、服务进行特征表示，通过Time2Vec刻画时序特征，再结合多种QoS属性的历史记录生成多源特征表示。其次，基于滑动窗口采用LSTM感知窗口内的时序关系，借助注意力机制细化窗口内不同时刻的关键性，从而构造待预测时刻的隐藏状态。最后，T-MST采用多任务预测层实现多种QoS属性的同时预测，它们共享上游模型，仅在预测层采用不同的感知模块以提升模型的鲁棒性和计算效率。本文基于真实世界的数据集进行了全面的实验验证，结果表明T-MST在吞吐量和响应时间的时序预测任务上平均绝对误差(Mean Absolute Error，MAE)分别平均提升了 37。53%和 20。38%，优于现有的时序QoS预测方法;而且T-MST的计算效率更高，能够有效应对实时QoS预测的需求。
Temporal QoS prediction based on multi-source features and multitask learning in cloud service recommendation
With the popularization of cloud computing technology,the number of cloud services is increasing exponentially,and users are no longer satisfied with functional requirements.Quality of Service(QoS)has be-come a key performance indicator for comparing the services.How to predict QoS in a dynamic and complex cloud environment in real-time and accurately,and recommend high-quality services to users has become a hot issue.Considering that the load of cloud servers,network status,and user preferences for accessing the cloud environment vary over time,this paper proposes a Temporal aware model based on Multi-Source and Multi-Task(T-MST),which can synchronously and accurately predict multiple QoS attributes.Firstly,T-MST performs feature representation on users and services,characterizes temporal features through Time2Vec,and generates multi-source feature representations by combining historical records of multiple QoS attributes.Sec-ondly,based on the sliding window,LSTM is used to perceive the temporal relationships within the window,and attention mechanism is used to refine the criticality of different time slots within the window,thereby con-structing a hidden state for the predicted time.Finally,T-MST uses a multitask prediction layer to achieve si-multaneous prediction of multiple QoS attributes,sharing upstream models and only using different perception modules in the prediction layer to improve model robustness and computational efficiency.This paper con-ducts comprehensive experimental verification based on real-world datasets,and the results show that T-MST has an average improvement of 37.53%and 20.37%in MAE in throughput and response time temporal pre-diction tasks,respectively,which is superior to existing temporal QoS prediction methods.Moreover,T-MST has higher computational efficiency and can effectively meet the demand for real-time QoS prediction.
国家科技期刊平台
NSTL
万方数据
