CN114757391A - Service quality prediction method based on network data space design - Google Patents

Service quality prediction method based on network data space design Download PDF

Info

Publication number
CN114757391A
CN114757391A CN202210264626.8A CN202210264626A CN114757391A CN 114757391 A CN114757391 A CN 114757391A CN 202210264626 A CN202210264626 A CN 202210264626A CN 114757391 A CN114757391 A CN 114757391A
Authority
CN
China
Prior art keywords
network
service
training sample
user
obtaining
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Granted
Application number
CN202210264626.8A
Other languages
Chinese (zh)
Other versions
CN114757391B (en
Inventor
鄢萌
王子梁
张小洪
吴云松
杨丹
付春雷
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Chongqing University
Original Assignee
Chongqing University
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Chongqing University filed Critical Chongqing University
Priority to CN202210264626.8A priority Critical patent/CN114757391B/en
Publication of CN114757391A publication Critical patent/CN114757391A/en
Application granted granted Critical
Publication of CN114757391B publication Critical patent/CN114757391B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
    • G06Q10/00—Administration; Management
    • G06Q10/04—Forecasting or optimisation specially adapted for administrative or management purposes, e.g. linear programming or "cutting stock problem"
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
    • G06Q10/00—Administration; Management
    • G06Q10/06—Resources, workflows, human or project management; Enterprise or organisation planning; Enterprise or organisation modelling
    • G06Q10/063—Operations research, analysis or management
    • G06Q10/0639—Performance analysis of employees; Performance analysis of enterprise or organisation operations
    • G06Q10/06395—Quality analysis or management

Landscapes

  • Engineering & Computer Science (AREA)
  • Business, Economics & Management (AREA)
  • Human Resources & Organizations (AREA)
  • Physics & Mathematics (AREA)
  • Theoretical Computer Science (AREA)
  • Economics (AREA)
  • Strategic Management (AREA)
  • General Physics & Mathematics (AREA)
  • Development Economics (AREA)
  • Entrepreneurship & Innovation (AREA)
  • Health & Medical Sciences (AREA)
  • Computational Linguistics (AREA)
  • Quality & Reliability (AREA)
  • General Business, Economics & Management (AREA)
  • Operations Research (AREA)
  • Marketing (AREA)
  • Game Theory and Decision Science (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Artificial Intelligence (AREA)
  • Biomedical Technology (AREA)
  • Biophysics (AREA)
  • Tourism & Hospitality (AREA)
  • Data Mining & Analysis (AREA)
  • Evolutionary Computation (AREA)
  • General Health & Medical Sciences (AREA)
  • Molecular Biology (AREA)
  • Computing Systems (AREA)
  • General Engineering & Computer Science (AREA)
  • Mathematical Physics (AREA)
  • Software Systems (AREA)
  • Educational Administration (AREA)
  • Management, Administration, Business Operations System, And Electronic Commerce (AREA)

Abstract

The invention relates to a service quality prediction method based on network data space design, which consists of potential network space learning and QoS prediction. The first part aims at learning three potential network spaces (i.e., a user potential network space, a service potential network space, and an interaction potential network space). The second part samples each feature in three potential network spaces first, and then fuses different features through three different convolutional networks to realize QoS prediction. By the two innovative components, the LNSL generates potential network space for each user and service at different positions, and high-precision QoS prediction is realized.

Description

Service quality prediction method based on network data space design
Technical Field
The invention relates to a network space learning method, in particular to a service quality prediction method based on network data space design.
Background
With the advent of the era where everything is a service, it becomes crucial how to recommend high quality services to users. The main idea of service recommendation is to predict the QoS (e.g. response time and throughput) values of services not invoked by the user and then recommend the best quality service to the user. In addition, some micro-service resource optimization methods require a high accuracy predictive QoS module. However, cyberspace information, which has a significant impact on QoS, is often difficult to capture due to cost and privacy concerns. For example, network space conditions (e.g., network speed, latency) of users and network providers offering services can greatly affect the quality of the service.
The main limitation of these predecessor QoS prediction methods is the lack of a method to mine the network space information for different users or service loss. In the existing QoS prediction method, collaborative filtering based on location information to find similar users and services is one of the core ideas. Wherein the known information is the location information of the service and the calling subscriber. Previous research has generally concluded that users and services located in close geographic locations tend to have the potential to achieve or provide a more similar quality of service. One important issue is that users or services that are close to each other do not necessarily have similar network states. As shown in fig. 1, both the user and the service are provided by a home carrier to provide network services. Two services or users in proximity to each other are provided by two different service providers. CF-based methods may lead to erroneous conclusions. Two users or services are in close physical proximity, but their computers may be located in different networks. They can be very remote in terms of network distance. That is, neighbors in physical location do not necessarily belong to the same network.
The existing QoS prediction methods can be divided into two categories, namely a QoS prediction method based on collaborative filtering and a QoS prediction method based on deep learning. The most common method of predicting QoS is a collaborative filtering model, which learns the QoS of a target service through similar users or services. The Qos prediction method based on cf can be divided into two categories, namely memory-based and model-based. The memory-based approach treats QoS (i.e., RT, TP) or attributes (i.e., distance) as a similarity measure between users or services. The key step is similarity calculation of the target object. Such methods include user-based (e.g., UPCC), service-based (e.g., IPCC), and user-service integration (e.g., UIPCC). To further improve the prediction accuracy, a number of model-based methods are applied, such as matrix decomposition (MF). This is an effective recommendation system technique, and has been widely used for QoS prediction in recent years. Due to the lack of information and data sparsity, the CF method often fails to achieve satisfactory prediction accuracy.
The deep neural network becomes an effective QoS prediction method. It captures the non-linear relationship by considering the characteristics of the object. The location information is relatively easy to obtain. A great deal of research has been done to improve the accuracy of QoS predictions using location information. Also, slice information is widely used in QoS prediction. The complex model improves the prediction accuracy to a certain extent, but needs many effective characteristics as prediction conditions. The existing deep neural network method cannot avoid mistakenly dividing users or services of different network spaces into similar groups due to the fact that the prediction accuracy is increased by utilizing the position information.
LNSL is inspired by other fields, such as variational-autocoder (VAE) -based models and latent-factor-based prediction methods. The main difference between the two methods is that the former learns different potential feature spaces by classifying features to fill in missing network information without pre-training.
Disclosure of Invention
Aiming at the problems in the prior art, the technical problems to be solved by the invention are as follows: how to provide a method with high QoS value prediction accuracy.
In order to solve the technical problems, the invention adopts the following technical scheme: a service quality prediction method based on network data space design comprises the following steps:
s1: constructing a training set, wherein each training sample in the training set comprises a user characteristic, a service characteristic, an interactive characteristic and a real QoS value, wherein:
user characteristics U are noted
Figure BDA0003551164200000021
Wherein
Figure BDA0003551164200000022
Which indicates the ID of the user or the like,
Figure BDA0003551164200000023
which represents the longitude of the user's hand,
Figure BDA0003551164200000024
represents the longitude of the user;
service characteristic S is noted
Figure BDA0003551164200000025
Wherein
Figure BDA0003551164200000026
Which indicates the service ID of the service,
Figure BDA0003551164200000027
which represents the longitude of the service or services,
Figure BDA0003551164200000028
represents the longitude of the service;
interaction characteristic C is noted
Figure BDA0003551164200000029
S2: constructing a prediction neural network, wherein the prediction neural network comprises a prior network, a characteristic adopting layer, an object perception convolutional layer, a known information perception convolutional layer, a full information perception convolutional layer and two full connection layers which are sequentially arranged;
s21: the detailed network structure of the prior network is as follows:
Figure BDA00035511642000000210
wherein E isiRepresenting the ith prior network, f[2]Representing a one-dimensional convolution of the signals,
Figure BDA00035511642000000211
for calculating the mean value of the tensor along a given axis, [1,2 ]]A first and a second dimension representing a tensor,
Figure BDA00035511642000000212
for delete appointments [1]Is the dimension of the first dimension, XiRepresenting a prior probability of an input;
the prior network is used for estimating the prior probability of the given combined feature, and the prior probability distribution is modeled as axial Gaussian distribution, wherein the expected mu and the variance sigma are both in accordance with the Gaussian distribution;
calculating the parameter mu by a priori networkiAnd σiThe calculation formula is as follows:
μi＝fμ(Ei；Wμ) (2)
σi＝fσ(Ei；Wσ) (3)
wherein, muiRepresenting the expectation of the ith prior network, σiRepresenting the variance of the ith prior network, fμIs represented by EiCalculating the expected course, WμRepresents the weight of μ, fσIs represented by EiProcedure for calculating the variance, WσA weight representing σ;
inputting the user characteristics U of the training sample into the ith prior network, wherein i is 1, obtaining the spatial prior distribution P1 of the user potential network, and obtaining mu1And σ1Inputting U of a training sample and a QoS value corresponding to the training sample into the ith prior network, wherein i is 2, obtaining the posterior distribution Pr1 of the potential network space of the user, and obtaining mu2And σ2；
Inputting the service characteristics S of the training sample into the ith prior network, wherein i is 3, obtaining the spatial prior distribution P2 of the service potential network, and obtaining mu3And σ3To train the sampleS and the QoS value corresponding to the training sample are input to the ith prior network, i is 4, the posterior distribution Pr2 of the service potential network space is obtained, and μ is obtained4And σ4；
Inputting the interactive characteristic C of the training sample into the ith prior network, wherein i is 5, obtaining the spatial prior distribution P3 of the interactive potential network, and obtaining mu5And σ5Inputting C of a training sample and a QoS value corresponding to the training sample into the ith prior network, wherein i is 6, obtaining the posterior distribution Pr3 of the interactive potential network space, and obtaining mu6And σ6；
S22: the feature applying layer
The feature Z1 of the training sample is sampled from P1, as in equation (4):
Z1～P1(·|U)＝N(μ1,diag(σ1)) (4)
the feature Z2 of the training sample is sampled from P2, as in equation (5):
Z2～P2(·|U)＝N(μ2,diag(σ2)) (5)
the feature Z3 of the training sample is sampled from P3, as in equation (6):
Z3～P3(·|U)＝N(μ3,diag(σ3)) (6)
the three results are concatenated as the final feature representation of the training sample, as shown in equation (7):
Figure BDA0003551164200000031
s23: respectively inputting the final characteristic representation Z of the training sample into an object perception convolutional layer, a known information perception convolutional layer and a full information perception convolutional layer, and then sequentially inputting the two full connection layers to obtain a QoS (quality of service) predicted value of the training sample;
s24: defining a loss function of the prediction neural network, finishing training when the value of the loss function is not changed any more to obtain the trained prediction neural network, otherwise updating the parameters of the prediction neural network, and inputting the parameters into a training sample technology for training;
s3: for a sample to be predicted, the user characteristics, the service characteristics and the interactive characteristics of the sample to be predicted are obtained by adopting the method of S1, the real QoS value of the sample to be predicted is set as a random array, and then the sample to be predicted is input into the trained prediction neural network to obtain the QoS prediction value corresponding to the sample to be predicted.
Preferably, in S1, the obtained longitude of the user and the obtained longitude of the service are respectively increased by 180 degrees, and the longitude range is mapped from [ -180,180] to [0,360 ]; the acquired latitude of the user and the latitude of the service are respectively increased by 90, and the latitude range is mapped from-90, 90 to [0,180 ].
Preferably, the specific process of sensing the convolutional layer by the object in S23 is as follows: the convolution kernel is 6 x 1, the step size is 8, and the object-aware convolution performs two convolution operations;
X1＝f(Z；w1) (8)
wherein, w1Is a one-dimensional convolution kernel of size 6X 1, X1Representing the output of the object-aware convolutional layer.
Preferably, the specific process of sensing the convolutional layer by the known information in S23 is as follows: the convolution kernel size is 3 x 1, the step size is 11, and the known information sensing convolution executes two convolution operations;
X2＝f(Z；w2) (9)
wherein w2Is a one-dimensional convolution kernel of size 3X 1, X2Representing the output of the known information-aware convolutional layer.
Preferably, the specific process of the full information sensing convolutional layer in S23 is as follows: the convolution kernel size is 20 x 1, and only one convolution operation is executed;
X3＝f(Z；w3) (10)
wherein, w3Is a one-dimensional convolution kernel of size 20X 1, X3Represents the output of the full information-aware convolutional layer.
Preferably, the specific process of the two fully-connected layers in S23 is as follows:
Figure BDA0003551164200000041
Figure BDA0003551164200000042
wherein the content of the first and second substances,
Figure BDA0003551164200000043
denotes the merging operation, F denotes the flattening operation of the flattened layer of Keras, F[16,8,1]Denotes a fully connected layer, [16,8,1 ]]The number of the neurons in the layer is represented,
Figure BDA0003551164200000044
representing the QoS prediction value.
Preferably, the loss function defined in S24 is as in formula (13):
Loss＝Hloss+KLloss (13)
wherein the content of the first and second substances,
Figure BDA0003551164200000045
where y represents the true QoS value of the training sample,
Figure BDA0003551164200000051
expressing a QoS value, a delta constant term and a beta constant term of a predicted value obtained by a training sample through a prediction neural network;
KLloss＝Dkl(Pr1||P1)+Dkl(Pr2||P2)+Dkl(Pr3||P3) (15)
wherein, three KL divergence
Figure BDA0003551164200000052
Compared with the prior art, the invention has at least the following advantages:
1. the method can realize high-precision QoS prediction, and the network space of each specific position is known in the invention to make up for the defects of the existing method.
2. The method of the invention provides a depth model to carry out the network space mining of position perception, so that the network space is as close to a real network space as possible. The former is realized by a convolutional neural network and a probability model, and the latter is realized by introducing a joint loss function into a feature space.
3. Large scale experiments were conducted on two real-world datasets to assess the effectiveness of the method of the invention, which was compared to the 10 most advanced baselines. Experimental results show that the method is superior to all baselines in QoS prediction accuracy.
Drawings
FIG. 1 is a flow chart of the method of the present invention.
Detailed Description
The present invention will be described in further detail below.
The invention provides a network-space learning (LNSL) -based service quality prediction method, which consists of potential network space learning and QoS prediction. The first part aims at learning three potential network spaces (i.e., a user potential network space, a service potential network space, and an interaction potential network space). The second part samples each feature in three potential network spaces first, and then fuses different features through three different convolutional networks to realize QoS prediction. By the two innovative components, the LNSL generates potential network space for each user and service at different positions, and high-precision QoS prediction is realized.
It is a major technical challenge how to mine the cyber-spatial information of users and services on the basis of known information. In the QoS prediction problem, the only known information is the location information and QoS history of the user and service. Our main observation is that when a user accesses a service in another network state through his own network, the network information of each location is hidden in the true QoS record. In the whole process, the network condition of the user, the network condition of the service and the interaction state of the two networks all affect the quality of the service. Therefore, we fuse three location features to establish prior distributions of different network spaces. And then fitting three posterior distributions by fusing actual QoS information and position characteristics, and minimizing the Kullback-Leible (KL) distance between the corresponding prior distribution and the posterior distribution. Therefore, the method can mine potential cyber-spatial information from the real QoS call records.
Referring to fig. 1, a service quality prediction method based on network data space design includes the following steps:
s1: constructing a training set, wherein each training sample in the training set comprises a user characteristic, a service characteristic, an interactive characteristic and a real QoS value, wherein:
user characteristics U are noted
Figure BDA0003551164200000061
Wherein
Figure BDA0003551164200000062
Which indicates the ID of the user or the like,
Figure BDA0003551164200000063
which represents the longitude of the user's hand,
Figure BDA0003551164200000064
represents the longitude of the user;
service characteristic S is noted
Figure BDA0003551164200000065
Wherein
Figure BDA0003551164200000066
Which represents the service ID of the service-related service,
Figure BDA0003551164200000067
which represents the longitude of the service or services,
Figure BDA0003551164200000068
represents the longitude of the service;
interaction feature C is denoted as
Figure BDA0003551164200000069
S2: constructing a prediction neural network, wherein the prediction neural network comprises a prior network, a characteristic adopting layer, an object perception convolutional layer, a known information perception convolutional layer, a full information perception convolutional layer and two full connection layers which are sequentially arranged;
s21: the detailed network structure of the prior network is as follows:
Figure BDA00035511642000000610
wherein E isiRepresenting the ith prior network, f[2]Representing a one-dimensional convolution with a convolution kernel of 2 x 1,
Figure BDA00035511642000000611
for calculating the mean value of the tensor along a given axis, [1,2 ]]A first and a second dimension representing a tensor,
Figure BDA00035511642000000612
for delete appointments [1]Refers to the dimension of the first dimension, XiRepresenting a prior probability of an input;
the prior network is used for estimating the prior probability of the given combined feature, and the prior probability distribution is modeled as axial Gaussian distribution, wherein the expected mu and the variance sigma are both in accordance with the Gaussian distribution;
calculating the parameter mu by a priori networkiAnd σiThe calculation formula is as follows:
μi＝fμ(Ei；Wμ) (2)
σi＝fσ(Ei；Wσ) (3)
wherein, muiRepresenting the expectation of the ith prior network, σiRepresents the variance of the ith prior network, fμIs represented by EiCalculating the expected course, WμRepresents the weight of μ, fσIs represented by EiCalculating the variance overStroke, WσA weight representing σ;
inputting the user characteristics U of the training sample into the ith prior network, wherein i is 1, obtaining the spatial prior distribution P1 of the user potential network, and obtaining mu1And σ1Inputting U of a training sample and a QoS value corresponding to the training sample into the ith prior network, wherein i is 2, obtaining the posterior distribution Pr1 of the potential network space of the user, and obtaining mu2And σ2；
Inputting the service characteristics S of the training sample into the ith prior network, wherein i is 3, obtaining the spatial prior distribution P2 of the service potential network, and obtaining mu3And σ3Inputting S of a training sample and a QoS value corresponding to the training sample into the ith prior network, wherein i is 4, obtaining a service potential network space posterior distribution Pr2, and obtaining mu4And σ4；
Inputting the interactive characteristics C of the training sample into the ith prior network, and obtaining the spatial prior distribution P3 of the interactive potential network and mu, wherein i is 55And σ5Inputting C of a training sample and a QoS value corresponding to the training sample into the ith prior network, wherein i is 6, obtaining the posterior distribution Pr3 of the interactive potential network space, and obtaining mu6And σ6；
S22: the feature applying layer
The feature Z1 of the training sample is sampled from P1, as in formula (4):
Z1～P1(·|U)＝N(μ1,diag(σ1)) (4)
the feature Z2 of the training sample is sampled from P2, as in equation (5):
Z2～P2(·|U)＝N(μ2,diag(σ2)) (5)
the feature Z3 of the training sample is sampled from P3 as in equation (6):
Z3～P3(·|U)＝N(μ3,diag(σ3)) (6)
the three results are concatenated as the final characterization of the training sample, as shown in equation (7):
Figure BDA0003551164200000071
s23: respectively inputting the final characteristic representation Z of the training sample into an object perception convolutional layer, a known information perception convolutional layer and a full information perception convolutional layer, and then sequentially inputting the two full connection layers to obtain a QoS (quality of service) predicted value of the training sample;
s24: defining a loss function of the prediction neural network, finishing training when the value of the loss function is not changed any more to obtain the trained prediction neural network, otherwise updating the parameters of the prediction neural network, and inputting a training sample technology for training;
s3: for a sample to be predicted, the user characteristics, the service characteristics and the interactive characteristics of the sample to be predicted are obtained by adopting the method of S1, the real QoS value of the sample to be predicted is set as a random array, and then the sample to be predicted is input into the trained prediction neural network to obtain the QoS prediction value corresponding to the sample to be predicted.
And embedding the features. In order for the neural network to learn additional data features, the user identifier, user location information, service identifier, service location information and real-time response time are input into the Keras' embedded layer, which can be considered as a special unbiased term fully connected layer. Specifically, the inlining thermally encodes the input to generate a zero vector of a specified dimension, with the ith position of the vector set to 1. Since the embedding layer only accepts positive integer input data. We use some conversion operations to convert the location information and RT information into positive integers. I.e., the longitude in the location information is incremented by 180 and the longitude range is mapped from-180,180 to [0,360 ]. At the same time, we increase the latitude by 90 and map the latitude range from [ -90,90] to [0,180 ]. Specifically, the acquired longitude of the user and the longitude of the service are respectively increased by 180 degrees, and the longitude range is mapped from [ -180,180] to [0,360 ]; the acquired latitude of the user and the latitude of the service are increased by 90, respectively, and the latitude range is mapped from-90, 90 to [0,180 ].
The embedding method uses dense vectors to represent words or documents, similar to natural language processing. By this operation, the classification features are mapped to a high-dimensional dense embedding vector.
And combining the feature vectors into three different combined features according to the physical significance of the features, namely user features, service features and interactive features. The user characteristics are composed of a user ID, a user longitude and a user latitude. The service characteristics are composed of a service ID, a service longitude and a service latitude. The interactive characteristics include user longitude, latitude and service longitude, latitude.
A priori distribution of potential network space. Central component (P) of the architecture of the present invention1,P2,P3) Is a prior distribution of three dimensions. Each position in this space encodes a variant. Estimating the probability of these variants of a given combined feature by a "prior network" parameterized by a weight w
Figure BDA0003551164200000081
These prior probability distributions are modeled as axial gaussian distributions.
Posterior distribution of network space. In the potential network space learning module, three different network space prior distributions are established through known position characteristics. During the training process, we fit the posterior distribution Pr with the true QoS value and the known location featuresi. By minimizing PriAnd PiKL divergence between, a priori distribution P of known featuresiCloser to the posterior distribution Pri. The specific network structure is as follows:
Figure BDA0003551164200000082
and is
Figure BDA0003551164200000083
Finally, three posterior distributions Pr are obtained through a prior networki。
Figure BDA0003551164200000084
XriRepresents the ith posterior scoreCloth, UrRepresenting user information with a genuine label, SrRepresenting service information with real labels, CrRepresenting the interaction information with the real tag,
Figure BDA0003551164200000085
representing a true tag vector.
Specifically, three different convolution kernels are designed to realize the classification perception of the features.
1) The object-aware convolution. Object-Con considers the status of a user or service separately by generating Object features. It can focus on the status of the user object and the service object itself to judge the quality of the service. The convolution kernel is 6 x 1 and the step size is 8. The object-aware convolution performs two convolution operations. The first convolution operation fuses all the user features to obtain the user object features. The second convolution operation fuses all the service features to obtain service object features.
X1＝f(Z；w1) (8)
Wherein w1Is a one-dimensional convolution kernel of size 6X 1, X1Representing the output of the object-aware convolutional layer.
2) Known information aware convolution. The convolution kernel size is 3 x 1 with step size 11. The known information-aware convolution performs two convolution operations. The first convolution operation fuses all the features known to the user to obtain the fused features known to the user. And fusing all the known service characteristics by the second convolution operation to obtain the known service fusion characteristics.
X2＝f(Z；w2) (9)
Wherein w2Is a one-dimensional convolution kernel of size 3X 1, X2Representing the output of the known information-aware convolutional layer.
3) Full information aware convolution. All-Con performs only one convolution operation. And (4) fusing all the features through convolution operation to obtain fused features.
X3＝f(Z；w3) (10)
Wherein w3Is a one-dimensional convolution kernel of size 20X 1, X3Represents the output of the full information-aware convolutional layer.
And predicting the network. We combine the fused and sampled features and pass through a convolutional layer. Finally, QoS prediction is achieved through a fully connected network. The method comprises the following specific steps:
Figure BDA0003551164200000091
Figure BDA0003551164200000092
wherein
Figure BDA0003551164200000093
Indicating a merge operation. F represents the flattening operation of the flattened layer of Keras. F[16,8,1]Denotes a fully connected layer, [16,8,1 ]]
The number of the neurons in the layer is represented,
Figure BDA0003551164200000094
indicating the predicted value.
The Huber loss, which is a loss function used in robust regression, is less sensitive to outliers in the data analysis,
therefore, the loss function defined in S24 is as in formula (13):
Loss＝Hloss+KLloss (13)
wherein
Figure BDA0003551164200000095
Where y represents the true QoS value of the training sample,
Figure BDA0003551164200000096
and the QoS value represents a predicted value obtained by the training sample through a prediction neural network, delta is a constant item, the default value is 0.5, and beta is a constant item, and the default value is 10.
KLloss＝Dkl(Pr1||P1)+Dkl(Pr2||P2)+Dkl(Pr3||P3) (15)
Wherein, three KL divergence
Figure BDA0003551164200000097
In the training process, KLlossThree prior distributions P are calculatediAnd corresponding posterior distribution PriThe distance between them. Instead of using a random array (e.g., an array of all 1 s) during testing, we use a random array
Figure BDA0003551164200000098
Since the network parameters are not updated during the training process, PriThe test result is not affected.
And (3) experimental verification:
data set
We have experimented on two reference data sets, which are Web services QoS data collected by WS-Dream system. This is a large real-world Web service data set containing 1,974675 QoS values for Web services collected from 339 users on 5825 services, including user and service location information. In this context, the QoS data set exists in the form of a user-service matrix, where the row index represents a user identifier and the column index represents a service identifier, each value in the matrix representing a corresponding Response Time (RT).
In calculating MAE and root mean square error, from
Figure BDA0003551164200000104
Outliers are eliminated. We have no true label for outliers. Therefore, we need to detect outliers from the dataset. In the present invention, we adopt the same detection method and parameter setting as the use of the ifforest (short for forest isolation) method for outlier detection. forest detects outliers based on the concept of isolation without using any distance or density metric, which makes the algorithm very goodIs efficient and robust. ifforest will calculate an outlier for each datum. The value range of the fraction is [0,1 ]]A larger value indicates a greater likelihood of an outlier. From the outliers, we can flexibly set the number of outliers. Here set to 0.1.
1.2 evaluation index
In the QoS prediction problem, an important criterion for evaluation is prediction accuracy. In most studies, two metrics are used to measure accuracy. The first metric is the mean absolute error function (MAE):
Figure BDA0003551164200000101
wherein r isi,jIn the form of an actual value of the value,
Figure BDA0003551164200000102
the predicted value of the QoS attribute is N, and the number of the QoS records is N. The second metric is Root Mean Square Error (RMSE):
Figure BDA0003551164200000103
the lower the values of these two indicators, the more accurate our predictions are represented.
For convenience, we compare the inventive method LNSL with the following six methods:
IPCC calculates the similarity between two services by using the idea of project-based collaborative filtering algorithm.
UPCC uses similar user behavior information to predict QoS.
UIPCC, which is a hybrid collaborative filtering method, combines the results of UPCC and IPCC.
HMF is also a collaborative filtering method based on a model, and uses position information to cluster users and services, and then combines local matrix decomposition prediction and global matrix decomposition prediction.
LDCF is an advanced deep learning method with location awareness for QoS prediction.
CMF is an anomaly resilient QoS prediction method that uses cauchy loss to measure the difference between observed and predicted QoS values.
For all methods, outliers will be removed when calculating MAE and RMSE. Furthermore, in the experiments we run 10 times per method and reported the average results for a fair comparison.
To evaluate the performance of the method of the present invention, we selected six baselines for comparison, including the traditional algorithm and the most advanced method. We performed experiments with different data densities on two data sets with and without outliers.
As shown in table 1, LNSL outperformed the comparative method in any case. Among all comparison methods, CMF is the latest method based on outlier optimization and LDCF is a deep learning method based on location awareness. CMF performed better than other comparison methods on datasets without outliers. On a data set with outliers, the LDCF is more accurate than other comparison methods by adding distance information in a deep learning model. On the RT data set with outliers, LNSL increased on average 22.82% on MAE index and 5.34% on RMSE compared to the two baseline methods. On the RT data set without outliers, LNSL of MAE index increased 50.62% on average and LNSL of RMSE index increased 22.43% on average.
Table 1 shows the comparison of the predicted results of the method of the present invention and the prior art
Figure BDA0003551164200000111
Experimental results show that the method has the smallest mean square error and mean square error in all cases. This means that the proposed method performs better than other baselines. The accuracy of this method has significant advantages, particularly when the known information in the dataset does not have outliers.
Finally, the above embodiments are only for illustrating the technical solutions of the present invention and not for limiting, although the present invention has been described in detail with reference to the preferred embodiments, it should be understood by those skilled in the art that modifications or equivalent substitutions may be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all of them should be covered in the claims of the present invention.

Claims (7)

1. A service quality prediction method based on network data space design is characterized by comprising the following steps:
s1: constructing a training set, wherein each training sample in the training set comprises a user characteristic, a service characteristic, an interactive characteristic and a real QoS value, wherein:
user characteristics U are noted
Figure FDA0003551164190000011
Wherein
Figure FDA0003551164190000012
Which indicates the ID of the user or the like,
Figure FDA0003551164190000013
which represents the longitude of the user and is,
Figure FDA0003551164190000014
represents the longitude of the user;
service characteristic S is noted
Figure FDA0003551164190000015
Wherein
Figure FDA0003551164190000016
Which represents the service ID of the service-related service,
Figure FDA0003551164190000017
which represents the longitude of the service or services,
Figure FDA0003551164190000018
represents the longitude of the service;
interaction characteristic C is noted
Figure FDA0003551164190000019
S2: constructing a prediction neural network, wherein the prediction neural network comprises a prior network, a characteristic adoption layer, an object perception convolutional layer, a known information perception convolutional layer, a full information perception convolutional layer and two full connection layers which are arranged in sequence;
s21: the detailed network structure of the prior network is as follows:
Figure FDA00035511641900000110
wherein E isiRepresenting the ith prior network, f[2]Representing a one-dimensional convolution of the signals,
Figure FDA00035511641900000111
for calculating the mean value of the tensor along a given axis, [1,2 ]]A first and a second dimension representing a tensor,
Figure FDA00035511641900000112
for delete appointments [1]Is the dimension of the first dimension, XiRepresenting a prior probability of an input;
the prior network is used for estimating the prior probability of the given combined feature, and the prior probability distribution is modeled as axial Gaussian distribution, wherein the expected mu and the variance sigma are both in accordance with the Gaussian distribution;
calculating the parameter mu by a priori networkiAnd σiThe calculation formula is as follows:
μi＝fμ(Ei；Wμ) (2)
σi＝fσ(Ei；Wσ) (3)
wherein, muiRepresenting the expectation of the ith prior network, σiRepresents the variance of the ith prior network, fμIs represented by EiCalculating what is expectedProcess, WμRepresents the weight of μ, fσIs represented by EiProcedure for calculating the variance, WσA weight representing σ;
inputting the user characteristics U of the training sample into the ith prior network, wherein i is 1, obtaining the spatial prior distribution P1 of the user potential network, and obtaining mu1And σ1Inputting U of a training sample and a QoS value corresponding to the training sample into the ith prior network, wherein i is 2, obtaining the posterior distribution Pr1 of the potential network space of the user, and obtaining mu2And σ2；
Inputting the service characteristics S of the training sample into the ith prior network, wherein i is 3, obtaining the spatial prior distribution P2 of the service potential network, and obtaining mu3And σ3Inputting S of a training sample and a QoS value corresponding to the training sample into the ith prior network, wherein i is 4, obtaining a service potential network space posterior distribution Pr2, and obtaining mu4And σ4；
Inputting the interactive characteristic C of the training sample into the ith prior network, wherein i is 5, obtaining the spatial prior distribution P3 of the interactive potential network, and obtaining mu5And σ5Inputting C of a training sample and a QoS value corresponding to the training sample into the ith prior network, wherein i is 6, obtaining the posterior distribution Pr3 of the interactive potential network space, and obtaining mu6And σ6；
S22: the feature applying layer
The feature Z1 of the training sample is sampled from P1, as in equation (4):
Z1～P1(·|U)＝N(μ1,diag(σ1)) (4)
the feature Z2 of the training sample is sampled from P2, as in equation (5):
Z2～P2(·|U)＝N(μ2,diag(σ2)) (5)
the feature Z3 of the training sample is sampled from P3, as in equation (6):
Z3～P3(·|U)＝N(μ3,diag(σ3)) (6)
and connecting the three adoption results in series to serve as the final characteristic expression of the training sample, as shown in formula (7):
Figure FDA0003551164190000021
s23: respectively inputting the final characteristic representation Z of the training sample into an object perception convolutional layer, a known information perception convolutional layer and a full information perception convolutional layer, and then sequentially inputting the two full connection layers to obtain a QoS (quality of service) predicted value of the training sample;
s24: defining a loss function of the prediction neural network, finishing training when the value of the loss function is not changed any more to obtain the trained prediction neural network, otherwise updating the parameters of the prediction neural network, and inputting the parameters into a training sample technology for training;
s3: for a sample to be predicted, the user characteristic, the service characteristic and the interaction characteristic of the sample to be predicted are obtained by adopting the method of S1, the real QoS value of the sample to be predicted is set as a random array, and then the sample to be predicted is input into the trained prediction neural network to obtain the QoS prediction value corresponding to the sample to be predicted.
2. The method of claim 1, wherein the method comprises: in the S1, the obtained longitude of the user and the obtained longitude of the service are respectively increased by 180 degrees, and the longitude range is mapped from [ -180,180] to [0,360 ]; the acquired latitude of the user and the latitude of the service are respectively increased by 90, and the latitude range is mapped from-90, 90 to [0,180 ].
3. The method of claim 1 or 2, wherein the method for predicting quality of service based on network data space design comprises: the specific process of the object-aware convolutional layer in S23 is as follows: the convolution kernel is 6 x 1, the step size is 8, and the object-aware convolution performs two convolution operations;
X1＝f(Z；w1) (8)
wherein, w1Is a one-dimensional convolution kernel of size 6X 1, X1Representing the output of the object-aware convolutional layer.
4. The method of claim 3, wherein the method comprises: the specific process of the known information sensing convolutional layer in S23 is as follows: the convolution kernel size is 3 x 1, the step size is 11, and the known information sensing convolution executes two convolution operations;
X2＝f(Z；w2) (9)
wherein, w2Is a one-dimensional convolution kernel of size 3X 1, and X2 represents the output of the known information-aware convolutional layer.
5. The method of claim 4, wherein the method comprises: the specific process of the full information sensing convolutional layer in S23 is as follows: the convolution kernel size is 20 x 1, and only one convolution operation is executed;
X3＝f(Z；w3) (10)
wherein, w3Is a one-dimensional convolution kernel of size 20X 1, X3Representing the output of the full information aware convolutional layer.
6. The method of claim 5 for predicting quality of service based on network data space design, wherein: the specific process of the two full connection layers in S23 is as follows:
Figure FDA0003551164190000031
Figure FDA0003551164190000032
wherein the content of the first and second substances,
Figure FDA0003551164190000033
denotes the merging operation, F denotes the flattening operation of the flattened layer of Keras, F[16，8，1]Showing a fully connected layer, [16,8,1 ]]The number of the layer of neurons is represented,
Figure FDA0003551164190000034
indicating the QoS prediction value.
7. The method of claim 6, wherein the method comprises: the loss function defined in S24 is as in equation (13):
Loss＝Hloss+KLloss (13)
wherein the content of the first and second substances,
Figure FDA0003551164190000035
where y represents the true QoS value of the training sample,
Figure FDA0003551164190000036
expressing a QoS value, a delta constant term and a beta constant term of a predicted value obtained by a training sample through a prediction neural network;
KLloss＝Dkl(Pr1||P1)+Dkl(Pr2||P2)+Dkl(Pr3||P3) (15)
wherein, three KL divergence
Figure FDA0003551164190000037
CN202210264626.8A 2022-03-17 2022-03-17 Network data space design and application method oriented to service quality prediction Active CN114757391B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202210264626.8A CN114757391B (en) 2022-03-17 2022-03-17 Network data space design and application method oriented to service quality prediction

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202210264626.8A CN114757391B (en) 2022-03-17 2022-03-17 Network data space design and application method oriented to service quality prediction

Publications (2)

Publication Number Publication Date
CN114757391A true CN114757391A (en) 2022-07-15
CN114757391B CN114757391B (en) 2024-05-03

Family

ID=82326748

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202210264626.8A Active CN114757391B (en) 2022-03-17 2022-03-17 Network data space design and application method oriented to service quality prediction

Country Status (1)

Country Link
CN (1) CN114757391B (en)

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN117274616A (en) * 2023-09-26 2023-12-22 南京信息工程大学 Multi-feature fusion deep learning service QoS prediction system and prediction method

Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111368888A (en) * 2020-02-25 2020-07-03 重庆邮电大学 Service function chain fault diagnosis method based on deep dynamic Bayesian network
CN111541570A (en) * 2020-04-22 2020-08-14 北京交通大学 Cloud service QoS prediction method based on multi-source feature learning
WO2021186158A1 (en) * 2020-03-17 2021-09-23 The University Court Of The University Of Edinburgh A distributed network traffic data decomposition method
CN113743675A (en) * 2021-09-13 2021-12-03 南京信息工程大学 Cloud service QoS deep learning prediction model
CN113810954A (en) * 2021-09-08 2021-12-17 国网宁夏电力有限公司信息通信公司 Virtual resource dynamic expansion and contraction method based on flow prediction and deep reinforcement learning
CN114117945A (en) * 2022-01-26 2022-03-01 南京信息工程大学 Deep learning cloud service QoS prediction method based on user-service interaction graph

Patent Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111368888A (en) * 2020-02-25 2020-07-03 重庆邮电大学 Service function chain fault diagnosis method based on deep dynamic Bayesian network
WO2021186158A1 (en) * 2020-03-17 2021-09-23 The University Court Of The University Of Edinburgh A distributed network traffic data decomposition method
CN111541570A (en) * 2020-04-22 2020-08-14 北京交通大学 Cloud service QoS prediction method based on multi-source feature learning
CN113810954A (en) * 2021-09-08 2021-12-17 国网宁夏电力有限公司信息通信公司 Virtual resource dynamic expansion and contraction method based on flow prediction and deep reinforcement learning
CN113743675A (en) * 2021-09-13 2021-12-03 南京信息工程大学 Cloud service QoS deep learning prediction model
CN114117945A (en) * 2022-01-26 2022-03-01 南京信息工程大学 Deep learning cloud service QoS prediction method based on user-service interaction graph

Non-Patent Citations (3)

* Cited by examiner, † Cited by third party
Title
ZILIANG WANG 等: "HSA-Net: Hidden-State-Aware Networks for High-Precision QoS Prediction", 《IEEE TRANSACTIONS ON PARALLEL AND DISTRIBUTED SYSTEMS》, vol. 33, no. 6, 10 September 2021 (2021-09-10), pages 1421 - 1435, XP011886008, DOI: 10.1109/TPDS.2021.3111810 *
冯欣 等: "基于视觉注意力变化的网络丢包视频质量评估", 《自动化学报》, vol. 37, no. 11, 15 November 2011 (2011-11-15), pages 1322 - 1331 *
高继勋;马小雨;: "基于增强学习与QoE需求函数的自适应无线网络择取优化机制", 计算机应用与软件, no. 07, 15 July 2015 (2015-07-15), pages 143 - 147 *

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN117274616A (en) * 2023-09-26 2023-12-22 南京信息工程大学 Multi-feature fusion deep learning service QoS prediction system and prediction method
CN117274616B (en) * 2023-09-26 2024-03-29 南京信息工程大学 Multi-feature fusion deep learning service QoS prediction system and prediction method

Also Published As

Publication number Publication date
CN114757391B (en) 2024-05-03

Similar Documents

Publication Publication Date Title
Zhang et al. Location-aware deep collaborative filtering for service recommendation
Yang et al. A location-based factorization machine model for web service QoS prediction
CN109062962B (en) Weather information fused gated cyclic neural network interest point recommendation method
CN112084427A (en) Interest point recommendation method based on graph neural network
Xu et al. Predicting destinations by a deep learning based approach
Gong et al. Missing value imputation for multi-view urban statistical data via spatial correlation learning
Du et al. GAN-based anomaly detection for multivariate time series using polluted training set
CN112800237B (en) Prediction method and device based on knowledge graph embedded representation and computer equipment
CN112115967B (en) Image increment learning method based on data protection
Xia et al. Joint deep networks based multi-source feature learning for QoS prediction
Huynh et al. Context-similarity collaborative filtering recommendation
CN117061322A (en) Internet of things flow pool management method and system
Chen et al. A cluster feature based approach for QoS prediction in web service recommendation
CN114757391B (en) Network data space design and application method oriented to service quality prediction
Zhu et al. Difftraj: Generating gps trajectory with diffusion probabilistic model
CN114445121A (en) Advertisement click rate prediction model construction and advertisement click rate prediction method
Zhan et al. Neural networks for geospatial data
Yang et al. Gated graph convolutional network based on spatio-temporal semi-variogram for link prediction in dynamic complex network
Rohilla et al. An Empirical Framework for Recommendation-based Location Services Using Deep Learning
Liao et al. Traj2Traj: A road network constrained spatiotemporal interpolation model for traffic trajectory restoration
CN116467466A (en) Knowledge graph-based code recommendation method, device, equipment and medium
CN116310728A (en) Browser identification method based on CNN-Linformer model
CN114462707B (en) Web service multidimensional QoS joint prediction method based on feature depth fusion
CN115664970A (en) Network abnormal point detection method based on hyperbolic space
CN116258923A (en) Image recognition model training method, device, computer equipment and storage medium

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination
GR01 Patent grant
GR01 Patent grant