Research On Service Recommendation Methods Based On Deep Learning

With the widespread deployment of Cloud Services,Micro Services,Mobile Services,IoT services and other services on the Internet the era “Everything as a Service” is coming.Various service resource deployed on Internet has heavily raise user's difficulties when finding available services.Hence,how to recommend services to users that meet their differentiated needs has become one of the most challenging research topics in service computing.In Service-Oriented Architecture(SOA)recommenders,the user's personalized demands(i.e.preferences)is visualized as Quality-of-Service(QoS).Specifically,QoS is a collection of many non-functional attributes(e.g.response time,throughput,trust,etc.)description of Web services.Thus,Predicting the QoS becomes the crucial procedure in service recommendation.Among existing methods,Collaborative Filtering(CF)and Deep Learning(DL)are two classical paradigms in QoS prediction.The CF-based method uses the users' historical behavior records to predict QoS,in which similarity calculation is key point.Growing numbers of recent works have incorporated contextual information(e.g.,location,time,trust etc.)into similarity calculation to improve prediction accuracy.However,CF-based methods suffer from these shortcomings:(1)the similarity calculation method employed by traditional CF-based methods can only learn the low-dimensional and linear relation between users and services,and(2)the data sparsity problem in the real-world significantly impacts their recommendation performance.To tackle the above issues,some efforts have been devoted to explore the impact of DL in service recommendation.Not only can DL-based method learn high-dimensional non-linear relationships between users and services,but also can alleviate the data sparsity problem in real-world to some extent.Thus,this paper proposes two deep learning-based service recommendation methods for time-space-aware QoS prediction.The main contributions are as follows:(1)This paper proposes Time-aware Recurrent Tensor Factorization(RTF)method.This method integrates Recurrent Neural Network and Tensor Factorization for memorizing long-short-term dependency patterns between users and services.Firstly,we granulated three-dimensional user-service-time interaction tensor into three fixed-size embedding dense vectors.Secondly,two sub-methods,Personalized Gated Recurrent Unit(PGRU)and Generalized Tensor Factorization(GTF),simultaneously work on shared embedding dense vectors to memorize the long-short-term dependency patterns.Thirdly,we design a Hybrid Loss function that integrates the L1 Loss and the L2 Loss,which greatly enhances the model's fitting ability on multiple evaluation metrics.Finally,experimental results indicate that RTF obviously outperforms the six state-of-the-art service recommendation methods.(2)This paper proposes Location-aware Deep Collaborative Filtering(LDCF)method.This method integrates Multi-Layer-Perceptron(MLP)and location similarity Adaptive Corrector(AC)to learn high-dimensional non-linear relationship between users and services.Firstly,we project location features into dense high-dimensional embedding vectors.Secondly,we introduce the Huber loss function and makes the model robust and prevail on all evaluation metrics.Finally,experimental results show that LDCF significantly outperform 9 state-of-the-art service recommendation methods.(3)To verify the validity and practicability of these methods we proposed,extensive experiments are conducted on real-world WS-Dream dataset(i.e.,Dataset # 1 and Dataset # 2).To simulate the actual situation,this paper sparses the original dataset under various densities,and finally predicts two QoS attributes including response-time and throughput.The results indicate that the proposed methods' recommendation performance is significantly better than many state-of-the-art service recommendation methods.