CN113743675A - Cloud service QoS deep learning prediction model - Google Patents

Cloud service QoS deep learning prediction model Download PDF

Info

Publication number
CN113743675A
CN113743675A CN202111066541.0A CN202111066541A CN113743675A CN 113743675 A CN113743675 A CN 113743675A CN 202111066541 A CN202111066541 A CN 202111066541A CN 113743675 A CN113743675 A CN 113743675A
Authority
CN
China
Prior art keywords
qos
matrix
individual
user
features
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
CN202111066541.0A
Other languages
Chinese (zh)
Other versions
CN113743675B (en
Inventor
张佩云
黄文君
陈禹同
王轩
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.)
Nanjing University of Information Science and Technology
Original Assignee
Nanjing University of Information Science and Technology
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 Nanjing University of Information Science and Technology filed Critical Nanjing University of Information Science and Technology
Priority to CN202111066541.0A priority Critical patent/CN113743675B/en
Publication of CN113743675A publication Critical patent/CN113743675A/en
Application granted granted Critical
Publication of CN113743675B publication Critical patent/CN113743675B/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
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/22—Matching criteria, e.g. proximity measures
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/25—Fusion techniques
    • G06F18/253—Fusion techniques of extracted features
    • 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
    • G06Q50/00—Information and communication technology [ICT] specially adapted for implementation of business processes of specific business sectors, e.g. utilities or tourism
    • G06Q50/10—Services
    • H—ELECTRICITY
    • H04—ELECTRIC COMMUNICATION TECHNIQUE
    • H04L—TRANSMISSION OF DIGITAL INFORMATION, e.g. TELEGRAPHIC COMMUNICATION
    • H04L67/00—Network arrangements or protocols for supporting network services or applications
    • H04L67/01—Protocols
    • H04L67/10—Protocols in which an application is distributed across nodes in the network
    • H—ELECTRICITY
    • H04—ELECTRIC COMMUNICATION TECHNIQUE
    • H04L—TRANSMISSION OF DIGITAL INFORMATION, e.g. TELEGRAPHIC COMMUNICATION
    • H04L67/00—Network arrangements or protocols for supporting network services or applications
    • H04L67/50—Network services
    • H04L67/60—Scheduling or organising the servicing of application requests, e.g. requests for application data transmissions using the analysis and optimisation of the required network resources
    • H04L67/61—Scheduling or organising the servicing of application requests, e.g. requests for application data transmissions using the analysis and optimisation of the required network resources taking into account QoS or priority requirements

Landscapes

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

Abstract

The invention discloses a cloud service QoS deep learning prediction model, which comprises the following steps: s1: determining the basic concept of the model; s2: according to the determined basic concept, a QoS deep learning prediction model is provided; s3: an algorithm process is provided, and comprises multi-scale feature extraction and MM-DNN learning; s4: and evaluating the performance of the QoS deep learning prediction model. The model fuses features of three scales: global, local and individual features, and global and individual feature matrixes are obtained through non-negative matrix decomposition, and global and individual features can be extracted from the global and individual features. Because the QoS value of the cloud service can be influenced by the geographic position, similar users and similar services are obtained based on distance similarity calculation, and local characteristics of the users and the services are formed. The three characteristics are gradually fused in stages through the deep neural network, characteristic processing and learning are carried out, and the characteristics of each stage are corrected by using individual evaluation in each stage, so that the QoS value of the cloud service is predicted more accurately.

Description

Cloud service QoS deep learning prediction model
Technical Field
The invention relates to the technical field of QoS (quality of service), in particular to a cloud service QoS deep learning prediction model.
Background
Cloud computing provides users with fast and safe services, and with the rapid development of cloud computing, the number of cloud-based services will be increased continuously, and among a plurality of services with similar functions, users are difficult to select candidate services meeting the requirements of the users. In this case, the QoS of the cloud service needs to be further compared to obtain a better choice.
QoS is a set of parameters used to describe non-functional attributes of a service (e.g., throughput, response time, etc.) and is a key indicator often used in cloud computing to evaluate the performance of a service. Due to the uncertainty of the user information (such as network status, personal preference, etc.), when different users invoke the same service, the QoS evaluation values (QoS values for short) for the service may also have a large difference. For a used service, the user may evaluate through his historical QoS records. For an un-invoked service, the user will not be able to obtain the QoS value of the service to judge its performance. How to accurately predict the QoS value of a service, and help a user select a better quality and more appropriate service becomes one of the main challenges.
To solve this problem, many methods have been developed to predict the QoS value of a service. Most of these methods are inspired by collaborative filtering in service recommendation, and predict missing QoS by collecting historical information of users or services. However, most of these methods only use the information of the user-service QoS primitive matrix, and ignore other factors affecting the QoS value. Interaction of users and services based on different geographic positions is shown in fig. 1, and many application programs are deployed in a cloud by taking the internet as a center to provide safe and rapid services for customers. Not only the differences in user information, service characteristics, and network status may cause differences in QoS values of services, but also the geographic location may affect the QoS values of services. Generally, users or services in close geographic locations typically have similar QoS values. Therefore, it is important to consider the influence of geographical location on QoS prediction and service recommendation based on collecting user and service history information. With the rapid development of cloud computing, the number of cloud services providing similar functions is increasing, and Quality of Service (QoS) prediction becomes one of the main challenges for user Service recommendation. There are some work currently applying deep learning to QoS prediction, but further improvements in prediction accuracy are needed. For the situation, a cloud service QoS deep learning prediction model is proposed.
Disclosure of Invention
The invention aims to provide a cloud service QoS deep learning prediction model which can effectively solve the defects in the background technology. The model fuses features of three scales: global, local, and individual features. Global and individual feature matrices are obtained by Non-Negative Matrix Factorization (NMF), from which global and individual features can be extracted. Because the QoS value of the cloud service can be influenced by the geographic position, similar users and similar services are obtained based on distance similarity calculation, and local characteristics of the users and the services are formed. The three features are gradually fused in stages through MM-DNN, feature processing and learning are carried out, and the features of the stages are corrected by using individual evaluation in each stage, so that the QoS value of the cloud service is predicted more accurately.
The purpose of the invention can be realized by the following technical scheme:
a cloud service QoS deep learning prediction model, wherein the model building process comprises the following steps:
s1: determining the basic concept of the model, wherein the basic concept is as follows: multi-scale feature, global feature matrix
Figure BDA0003258606270000021
Local feature matrix (G), individual feature matrix and individual evaluation
Figure BDA0003258606270000022
S2: according to the determined basic concept, a QoS deep learning prediction model is provided, and the model is divided into three working stages: preprocessing, feature extraction and deep neural network prediction, wherein the preprocessing comprises the steps of carrying out nonnegative matrix decomposition on a user-service QoS (quality of service) original matrix, the feature extraction is to use the nonnegative matrix decomposition on the user-service QoS original matrix to obtain a global feature matrix and an individual feature matrix, respectively extracting global features and individual features from the global feature matrix and extracting local features based on distance attributes, and the deep neural network prediction processes the features and finally predicts a missing QoS value by designing MM-DNN;
s3: an algorithm process is provided, and comprises multi-scale feature extraction and MM-DNN learning;
s4: evaluating the performance of a QoS deep learning prediction model, performing an experiment on a QoS data set WS-DREAM of a real world, adopting an average absolute error (MAE) and a Root Mean Square Error (RMSE) as evaluation indexes for evaluating a prediction result, setting test parameters, determining parameters for optimizing the performance of the model in a set parameter range, evaluating the performance of the model by using the optimal parameters in VI-F, analyzing the influence of different parameter settings on the prediction accuracy of the model, determining the optimal values of the parameters of the model, performing an ablation experiment, comparing three different settings of the model, analyzing the influence of a network structure and matrix density of MM-DNN on time consumption, and finally comparing the MAE and the RMSE of the four methods under four different matrix densities (5%, 10%, 15% and 20%).
Further, the multi-scale features in S1 include global features, local features, and individual features;
the global feature matrix is obtained by fitting a user-service QoS (quality of service) original matrix Q after non-negative matrix decomposition, and global features can be extracted from the matrix;
the local feature matrix is composed of a global feature matrix
Figure BDA0003258606270000031
The matrix which is formed by the fitting QoS values of the similar users to the similar cloud service is extracted, and local features can be obtained from the matrix;
the individual characteristic matrix comprises a user individual characteristic U and a cloud service individual characteristic matrix S;
the individual evaluation
Figure BDA0003258606270000032
Used for describing evaluation of cloud service by user, from global feature matrix
Figure BDA0003258606270000033
And the user i corrects the characteristics of each stage of the model according to the fitted QoS value of the cloud service j.
Further, the preprocessing in S2 is to perform non-negative matrix decomposition on the user-service QoS original matrix Q to obtain a non-negative individual feature matrix of the user
Figure BDA0003258606270000034
Non-negative individual feature matrix of sum service
Figure BDA0003258606270000035
Each column U of UiAn l-dimensional individual feature vector representing user i, each column S in SjAn l-dimensional individual feature vector representing service j;
performing non-negative matrix factorization on Q to enable U, S inner products to be close to Q, wherein the formula (1) shows that:
Figure BDA0003258606270000041
Figure BDA0003258606270000042
obtaining a QoS value from a user-service QoS original matrix Q, learning to obtain U and S, and using the square of an error between the user-service QoS original matrix and a global feature matrix as a loss function, as shown in a formula (2):
Figure BDA0003258606270000043
the loss function in equation (2) is minimized using a multiplicative update rule, as shown in equation (3):
Figure BDA0003258606270000044
Figure BDA0003258606270000045
further, the global feature extraction in S2 is to extract from the global feature matrix
Figure BDA0003258606270000046
Extracting evaluation of user i to all cloud services
Figure BDA0003258606270000047
And all users' evaluations (U) of service jTSj) As global features, global features
Figure BDA0003258606270000048
Is recorded as:
Figure BDA0003258606270000049
further, in the S2, local feature extraction is firstly performed to calculate the user similarity and the cloud service similarity based on the distance, and the distance d between the users b and vbvCalculating as shown in equation (5):
c＝sin(γb)·sin(γv)+cos(γb)·cos(γv)·cos(θb-θv)
Figure BDA00032586062700000410
then sorting the distances, selecting similar users and similar services according to the distances, fitting QoS values to form a local feature matrix G, and converting the local feature matrix G into a one-dimensional vector serving as a vectorLocal features
Figure BDA00032586062700000411
As shown in formula (6);
Figure BDA00032586062700000412
further, the individual feature extraction in S2 is to extract a feature vector U of the user i from the user individual feature matrix UiExtracting a feature vector S of the cloud service j from the service individual feature matrix S as the user individual featurejAs serving individual characteristics, individual characteristics
Figure BDA00032586062700000413
Expressed as:
Figure BDA0003258606270000051
further, the deep neural network prediction structure in S2 is divided into four stages:
the stage 1 has a merging layer and L full-connection layers, global features are input, after the global features pass through one full-connection layer, the features are merged in the merging layer for individual evaluation and are corrected, information with the same size as the local features in the features is further extracted through the full-connection layer, and a feed-forward process of the stage can be expressed as follows:
Figure BDA0003258606270000052
Figure BDA0003258606270000053
Figure BDA0003258606270000054
Figure BDA0003258606270000055
Figure BDA0003258606270000056
to activate the function ReLU (Rectified Linear Unit, ReLU), as shown in equation (9):
Figure BDA0003258606270000057
the stage 2 comprises two merging layers and M full-connection layers, local features are merged in one merging layer, after passing through one full-connection layer, individual evaluation is merged in the merging layer, then the full-connection layer is fed in, and the process is as shown in the formula (10):
Figure BDA0003258606270000058
Figure BDA0003258606270000059
Figure BDA00032586062700000510
stage 3 has two merging layers and Z full-connection layers, connects individual characteristics in one merging layer, merges individual evaluation in the merging layer after one full-connection layer, learns through the full-connection layer, and the process is as shown in formula (11):
Figure BDA00032586062700000511
Figure BDA00032586062700000512
Figure BDA00032586062700000513
the first three stages all use BN and Dropout;
stage 4 has a merging layer in which individual evaluations are concatenated and a full-link layer
Figure BDA0003258606270000061
And (3) further learning through a full connection layer, thereby correcting the result and outputting a QoS predicted value, as shown in formula (12):
Figure BDA0003258606270000062
in the parameter training process, the Mean Absolute Error (MAE) is used as a loss function, as shown in formula (13):
Figure BDA0003258606270000063
parameters in the network are optimized, and a gradient descent method is adopted for model training, so that loss is minimized, as shown in a formula (14).
Figure BDA0003258606270000064
Figure BDA0003258606270000065
Further, in the S3, the multi-scale features are extracted as an algorithm 1, the MM-DNN is learned as an algorithm 2, the algorithm 1 obtains global features, local features, and individual features, the three features are used as input of the algorithm 2, the MM-DNN is subjected to staged feature fusion, and finally the algorithm 2 returns a QoS predicted value of the user to the cloud service.
Further, in S4, the Mean Absolute Error (MAE) and the Root Mean Square Error (RMSE) are used as evaluation indicators for evaluating the prediction result, as shown in formulas (15) and (16):
Figure BDA0003258606270000066
Figure BDA0003258606270000067
further, the cloud service QoS deep learning prediction model is stored in a memory configured to run the cloud service QoS deep learning prediction model, executing the cloud service QoS deep learning prediction model of any of claims 1 to 9.
The invention has the beneficial effects that:
the model of the invention integrates the characteristics of three scales: global, local and individual characteristics, global and individual characteristic matrixes are obtained through non-Negative Matrix Factorization (NMF), global and individual characteristics can be extracted from the global and individual characteristic matrixes, as the QoS value of the cloud service can be influenced by the geographic position, similar users and similar services are obtained through distance similarity calculation, local characteristics of the users and the services are formed, the three characteristics are gradually fused in stages through MM-DNN, characteristic processing and learning are carried out, and the characteristics of the stage are corrected by using individual evaluation in each stage, so that the QoS value of the cloud service is predicted more accurately.
Drawings
The invention will be further described with reference to the accompanying drawings.
FIG. 1 is a service invocation graph based on different geographical locations in the present invention;
FIG. 2 is an exemplary diagram of a user invoking a service in accordance with the present invention;
FIG. 3 is a diagram of a user-service QoS raw matrix of the present invention;
FIG. 4 is a diagram of a user-service QoS prediction matrix of the present invention;
FIG. 5 is a diagram of the multi-scale feature extraction process of the present invention;
FIG. 6 is a diagram of a cloud service QoS deep learning prediction model of the present invention;
FIG. 7 is an exploded view of a non-negative matrix of the present invention;
FIG. 8 is a diagram illustrating a merge operation of the present invention;
FIG. 9 is a partial feature matrix diagram of the present invention;
FIG. 10 is a partial feature extraction diagram of the present invention;
FIG. 11 is a block diagram of a deep neural network according to the present invention;
FIG. 12 is a graph of experimental parameter settings for the present invention;
FIG. 13 is a graph of experimental results of the impact of the number of similar users on the prediction accuracy of the present invention;
FIG. 14 is a graph of experimental results of the impact of similar services of the present invention on prediction accuracy;
FIG. 15 is a graph of experimental results of the impact of individual feature dimensions on prediction accuracy in accordance with the present invention;
FIG. 16 is a graph of the impact of phase 1-3 full link layer number on prediction accuracy for the present invention;
FIG. 17 is an experimental ablation map of the present invention;
FIG. 18 is a time consumption graph for various network configurations of the present invention;
FIG. 19 is a graph of time consumption for different matrix densities according to the present invention;
FIG. 20 is a graph of prediction accuracy versus response time attribute of the present invention;
FIG. 21 is a comparison graph of prediction accuracy for throughput attributes of the present invention.
Detailed Description
The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the drawings in the embodiments of the present invention, and it is obvious that the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. All other embodiments, which can be derived by a person skilled in the art from the embodiments given herein without making any creative effort, shall fall within the protection scope of the present invention.
Generally, one user may invoke multiple cloud services, while one service may be invoked by a different user. As the number of cloud services continues to increase, numerous functions have emergedSimilar services, users want to select candidate services meeting the requirements of the users, and services with better performance can be obtained by comparing the QoS of cloud services with similar functions. And the QoS value of the cloud service is recorded through the cloud service system or user feedback to represent the service performance. Three users u are given as in fig. 21～u3And 5 services s1～s5In a calling relationship, e.g. user u1 has called service s1、s2And s4The call relationship is shown by the solid arrow.
User-service QoS raw matrix Q: after the user calls the cloud service, the QoS value of the cloud service is collected through the cloud service system or the user feedback, and a user-service QoS original matrix Q is constructed. In the matrix Q, a row represents a user, a column represents a service, and each entry represents a QoS value fed back by the user in the row after invoking the service in the column. Assuming that m users in the cloud have called n cloud services, note:
Figure BDA0003258606270000081
wherein each term qijE Q represents a QoS evaluation value (QoS value for short) of the user i to the cloud service j. FIG. 3 shows the QoS rating obtained after 3 users invoke 5 services, for example, 1.984 shows user u1To service s1Because of user u1Without calling service s3And thus the corresponding position of fig. 3 is blank (missing item).
Since the users may not invoke all cloud services, only some users have QoS values for some services in the matrix Q, resulting in Q being very sparse. The missing entries in the original user-service QoS matrix Q may be predicted by the existing QoS values in the matrix, the QoS prediction values being shown in bold numbers in fig. 4.
The application aims to improve the accuracy of a QoS predicted value of a cloud service for a user, and provides a cloud service QoS deep learning prediction model based on staged multi-scale feature fusion and individual evaluation. Mainly comprises the following steps:
1. improving QoS prediction accuracy by improving feature extraction
At present, many QoS prediction works are carried out, and some work only utilizes a user-service QoS original matrix to carry out similarity calculation, and a QoS value is predicted according to similar users or services; some work only uses matrix factorization to make predictions from historical QoS records; other efforts have introduced context information to further improve prediction accuracy. However, they predict QoS values using only a single scale feature, which may leave the information incomplete, resulting in low prediction accuracy. The problem can therefore be solved from various perspectives using multi-scale features to achieve higher QoS prediction accuracy. Therefore, global, local and individual characteristics can be extracted and fused for QoS prediction, and therefore the prediction accuracy of the cloud service QoS value is improved.
2. Improving QoS prediction accuracy by improving deep neural network structure
In recent years, deep learning has achieved enormous results in many fields, and some work has also emerged to predict QoS using deep neural networks. Most of the works use a convolutional neural network or a multilayer perceptron for training, but the network structure is not perfect enough, the extraction and processing of features are required to be further improved, and the prediction accuracy is also in a space of great improvement. Therefore, based on the consideration, a new staged multi-scale feature fusion deep neural network is constructed, and extracted multi-scale features are fused in stages, so that the feature learning of the model is more effective, and a more accurate QoS predicted value is obtained.
A cloud service QoS deep learning prediction model needs to determine the basic concept of the model, as shown in fig. 5, the basic concept of the model is as follows:
1. multi-scale features: the multi-scale features refer to different features which are divided by using multiple scales for measurement, and comprise global features, local features and individual features;
2. global feature matrix
Figure BDA0003258606270000101
The global feature matrix is obtained by fitting a user-service QoS (quality of service) original matrix Q after non-negative matrix decomposition (matrix elements are called pseudo-ones)A resultant QoS value), from which a user-service QoS global feature (global feature for short) can be extracted;
3. local feature matrix (G): the local feature matrix is composed of global feature matrix
Figure BDA0003258606270000102
The extracted matrix formed by the fitting QoS values of the similar users to the similar cloud services is obtained by calculation based on distance attributes of the similar users and the similar services, and user-service QoS local features (local features for short) can be obtained from the matrix;
4. an individual feature matrix: the cloud service QoS system comprises a user individual characteristic matrix U and a cloud service individual characteristic matrix S, which are obtained by carrying out nonnegative matrix decomposition on a user-service QoS original matrix Q;
5. individual evaluation: individual evaluation
Figure BDA0003258606270000103
Used for describing evaluation of cloud service by user, from global feature matrix
Figure BDA0003258606270000104
And the QoS value of the user i fitting the cloud service j is used for correcting the characteristics of each stage of the model.
A cloud service QoS deep learning prediction model is proposed, as shown in figure 6,
the model mainly comprises four stages of work: preprocessing, feature extraction, deep neural network prediction and determination algorithm;
1. pretreatment: carrying out nonnegative matrix decomposition on a user-service QoS (quality of service) original matrix;
the preprocessing module mainly carries out nonnegative matrix decomposition on a user-service QoS (quality of service) original matrix Q to respectively obtain nonnegative individual feature matrices of users
Figure BDA0003258606270000105
Non-negative individual feature matrix of sum service
Figure BDA0003258606270000106
Wherein each column U of UiAn l-dimensional individual feature vector representing user i, each column S in SjAn l-dimensional individual feature vector representing service j. The non-negative matrix factorization of Q aims to find the most suitable U, S to make the inner product of the two approach Q as close as possible, as shown in equation (1):
Figure BDA0003258606270000107
Figure BDA0003258606270000111
wherein
Figure BDA0003258606270000112
Is the global feature matrix fitted by U, S.
Referring to fig. 7, fig. 7 shows a specific example of non-negative matrix decomposition, in which a user-service QoS raw matrix Q (fig. 7(a)) composed of 3 users and 5 cloud services is decomposed into a user individual feature matrix U (fig. 7(b)) and a service individual feature matrix S (fig. 7(c)), where fk(where k ∈ {1,2,3}) is the kth feature of the user or service, column 1 of the matrix [1.16759,0,0.18754 ] in FIG. 7(b) ]]Representing user u1Fig. 7(d) is a global feature matrix fitted to U, S.
Since U and S are unknown in advance, it is necessary to obtain QoS values from the user-service QoS primitive matrix Q and learn them. Using the square of the error between the user-service QoS raw matrix and the global feature matrix as a loss function, as shown in equation (2):
Figure BDA0003258606270000113
the loss function in equation (2) is minimized using a multiplicative update rule, as shown in equation (3):
Figure BDA0003258606270000114
Figure BDA0003258606270000115
wherein u iskiFor the k-th feature, s, of the user i in the user individual feature matrixkjThe kth feature of the service j in the service individual feature matrix is used.
2. Feature extraction: performing multi-scale feature extraction, obtaining a global feature matrix and an individual feature matrix by using nonnegative matrix decomposition on a user-service QoS (quality of service) original matrix, and extracting global features and individual features from the global feature matrix and the individual feature matrix respectively; extracting local features based on the distance attribute;
2.1, global feature extraction: from a global feature matrix
Figure BDA0003258606270000116
The evaluation of all cloud services by the user i (namely:
Figure BDA0003258606270000117
) And all users' evaluations of service j (i.e.: u shapeTSj) As global features, then global features
Figure BDA0003258606270000118
Can be written as:
Figure BDA0003258606270000119
wherein the content of the first and second substances,
Figure BDA0003258606270000126
for the merge operation, the two vectors are connected together. An example of the merge operation is given based on fig. 7(d), as shown in fig. 8. Wherein [1.98201,0.29433,0.06292,0.26363, 0.0025%]For user u1Evaluation of all services, [0.0025,0.36584,0.25229]Serving s for all pairs of users5Evaluation of (3). Through merging operationTwo vectors are sequentially connected together to become [1.98201,0.29433,0.06292,0.26363,0.0025,0.0025,0.36584,0.25229]。
2.2 local feature extraction
2.2.1, calculating the user similarity and the cloud service similarity based on the distance:
users or services with close geographical positions usually have similar QoS values, and users with close distances can be selected as similar users, and cloud services with close distances can be selected as similar services. Distance d between users b, vbvCalculating as shown in equation (5):
c＝sin(γb)·sin(γv)+cos(γb)·cos(γv)·cos(θb-θv)
Figure BDA0003258606270000121
wherein, γb，θb∈(-180,180]Respectively, the latitude and longitude of the user b, and r is the earth radius. dbvSmaller means that two users are geographically closer together, and have more similar QoS values. Let cxyThe distance similarity between x and y is served for the cloud, which is calculated in a similar manner to equation (5).
2.2.2, local feature extraction:
firstly, after calculating the user similarity and the service similarity based on the distance, respectively sequencing the distance, and selecting J users closest to the user i as similar users of the user i and K cloud services closest to the service J as similar services of the service J according to the distance. Then, from
Figure BDA0003258606270000122
Extracting the fitting QoS values of the J similar users to the K similar services to form a local feature matrix G, and converting the local feature matrix G into a one-dimensional vector serving as a local feature
Figure BDA0003258606270000123
As shown in equation (6).
Figure BDA0003258606270000124
Wherein the content of the first and second substances,
Figure BDA0003258606270000127
for the shape transformation function, the conversion of the matrix into a vector is implemented.
From the global feature matrix of FIG. 7(d)
Figure BDA0003258606270000131
Extracts user u from1、u3For cloud services s4、s5To generate a local feature matrix G, as shown in fig. 9, for user u1Selecting two users u nearest to the user1、u3(including itself); for service s5Selecting two services s closest to him4、s5(including itself), the generated local feature matrix G is converted into a one-dimensional vector as a local feature, as shown in fig. 10.
2.2.3, individual feature extraction:
extracting feature vector U of user i from user individual feature matrix UiExtracting a feature vector S of the cloud service j from the service individual feature matrix S as the user individual featurejAs a service individual feature. Individual characteristics
Figure BDA0003258606270000132
Can be expressed as shown in formula (7):
Figure BDA0003258606270000133
3. and (3) deep neural network prediction:
by designing MM-DNN, processing features and finally predicting a missing QoS value, namely, inputting global, local and individual features into a deep neural network in stages, processing the features by utilizing a full-connection layer, performing feature fusion in a merging layer, learning the fused features by utilizing a plurality of full-connection layers, correcting the features of each stage based on individual evaluation, and finally outputting a QoS predicted value.
Deep Neural Networks (DNNs) are the basis of Deep learning, and are Neural Networks including multiple hidden layers, and can be classified into Multi-Layer perceptrons (MLPs), Convolutional Neural Networks (CNNs), Recurrent Neural Networks (RNNs), and the like according to the characteristics of neurons. Neurons with different characteristics have different roles, such as: MLP can extract more abundant information, CNN is more suitable for two-dimensional features, and RNN is more sensitive to time information. However, most of the current cloud service QoS prediction work uses a convolutional neural network or a multilayer perceptron for training, but the network structure is not perfect enough, the extraction and processing of features need to be further improved, and the prediction accuracy has a space for greatly improving. Therefore, a new deep neural network model is constructed by improving the MLP structure to perform QoS prediction, as shown in fig. 11.
The deep neural network structure shown in fig. 11 is composed of a plurality of fully-connected layers and merging layers alternately, and a final prediction result is output through one fully-connected layer. After the multi-scale features are extracted, the model processes and learns the features through a deep neural network to obtain a final QoS predicted value. The model is divided into four stages, and the global features, the local features and the individual features are sequentially fused in the first three stages. In each stage, an individual evaluation is introduced to correct the characteristics. If the characteristics of three scales are simultaneously input into the model for prediction in the stage 1, the problem of overlarge parameter quantity is caused, so that the calculation expense is increased, and the multi-scale characteristics are fused in stages, so that the problem can be well relieved. In stage 4, individual evaluation is introduced, and the QoS predicted value is further corrected. FIG. 11 is specifically analyzed as follows:
stage 1:
in this stage, there is one merging layer and L fully connected layers. The model inputs global features, after passing through a full connection layer, the features are corrected by merging individual evaluation in a merging layer, and then information with the same size as the local features in the features is further extracted through the full connection layer. The feed forward process at this stage can be expressed as:
Figure BDA0003258606270000141
Figure BDA0003258606270000142
Figure BDA0003258606270000143
Figure BDA0003258606270000144
in the formula, y0As input of MM-DNN, ykIs the output of the kth fully-connected layer, αkAnd betakRespectively representing the weight and the offset of the kth fully-connected layer.
Figure BDA0003258606270000145
The function ReLU (Rectified Linear Unit, ReLU) is activated as shown in equation (9).
Figure BDA0003258606270000146
And (2) stage:
at this stage, there are two merging layers and M fully connected layers. Firstly merging local characteristics in a merging layer, merging individual evaluation in the merging layer after passing through a full connecting layer, and then feeding into the full connecting layer, wherein the process is shown as a formula (10):
Figure BDA0003258606270000151
Figure BDA0003258606270000152
Figure BDA0003258606270000153
and (3) stage:
in this stage, there are two merging layers and Z fully connected layers. Firstly, connecting individual characteristics in a merging layer, merging individual evaluation in the merging layer after passing through a full connecting layer, and learning through the full connecting layer, wherein the process is shown as a formula (11).
Figure BDA0003258606270000154
Figure BDA0003258606270000155
Figure BDA0003258606270000156
And (4) stage:
in this phase, there is one merging layer and one full connection layer. Since the model aims to obtain the QoS predicted value of user i to cloud service j, individual evaluation is connected in the merging layer
Figure BDA0003258606270000157
(i.e. the
Figure BDA0003258606270000158
The fitted QoS value of the user i to the service j), the result is further learned through a layer of full connection layer, and the QoS prediction value is finally output, as shown in formula (12).
Figure BDA0003258606270000159
In order to solve the problems of gradient explosion, gradient loss, overfitting and the like in the training process, BN and Dropout are used in the first three stages, and the training efficiency and the learning capacity of the model are improved. The BN is a means of resisting overfitting and accelerating model convergence, which is widely applied at present, and is used for regularizing dispersed data, namely preprocessing the data. Dropout is to ignore the influence of parameters in a part of neural network layers on the result according to a certain probability in the training process of the deep learning network to prevent overfitting, thereby improving the performance of the neural network.
In the parameter training process of the model, Mean Absolute Error (MAE) is used as a loss function, and the formula (13) is as follows:
Figure BDA0003258606270000161
wherein p isijAnd obtaining a QoS predicted value of the user i to the cloud service j by the QoS deep learning prediction model, wherein N is the number of the QoS values needing to be predicted.
In order to optimize parameters in the network, a gradient descent method is adopted for model training, and loss is minimized.
As shown in equation (14).
Figure BDA0003258606270000162
Figure BDA0003258606270000163
Wherein, λ is the learning rate for controlling the gradient descending speed in the iterative process, k belongs to {1,2, …, L + M + Z }, and the initial value α is1And beta1Obtained by generating random numbers.
4. The algorithm is as follows:
the algorithm process of the cloud service QoS deep learning prediction model based on staged multi-scale feature fusion and individual evaluation comprises the following steps: multi-scale feature extraction (algorithm 1) and MM-DNN learning (algorithm 2).
Global, local and individual characteristics are obtained through the algorithm 1, the three characteristics are used as input of the algorithm 2, staged characteristic fusion is carried out on MM-DNN, and finally the algorithm 2 returns a QoS predicted value of a user to the cloud service. Algorithm 1 is a multi-scale feature extraction algorithm.
Figure BDA0003258606270000164
Figure BDA0003258606270000171
Some of the explanations regarding algorithm 1 are as follows:
1.1, obtaining a global characteristic matrix (a 2 nd row) by carrying out NMF on Q to obtain an individual characteristic matrix (a 1 st row) and carrying out matrix multiplication.
1.2, respectively extracting individual characteristics of the user i and the service j (3 rd to 4 th lines).
1.3, merging
Figure BDA0003258606270000172
And UTSjTo obtain the global feature (line 5). Local features are extracted in line 6.
1.4 merging UiAnd SjTo obtain individual features (line 7) and return all extracted features (line 8).
After the multi-scale features are extracted, staged feature fusion and feature learning is performed by MM-DNN, as shown in algorithm 2.
Figure BDA0003258606270000173
Some of the explanations regarding algorithm 2 are as follows:
2.1 learning global features in phase 1 (lines 1-3), where F1For storing the global features learned by one full link layer in this stage
Figure BDA0003258606270000181
F2Used for storing the merging result C after the learning of the full connection layer in the stage1。
2.2, in stage 2, local feature fusion is performed through one merging layer (line 4), followed by feature learning (lines 5-7), where F3Used for storing the merging result C after one layer of full connection layer learning in the stage2，F4Used for storing the merging result C after the learning of the full connection layer in the stage3。
2.3 in stage 3, individual feature fusion through one merging layer (line 8) followed by feature learning (lines 9-11), where F5Used for storing the merging result C after one layer of full connection layer learning in the stage4，F6Used for storing the merging result C after the learning of the full connection layer in the stage5。
2.4 in stage 4, Individual assessment
Figure BDA0003258606270000182
Is input into the merging layer (line 12). Feature learning through fully connected layers (line 13), where pijIs the QoS prediction value learned by the full connection layer from the merging result C6. Finally, a QoS prediction value of the cloud service is returned (line 14).
And (3) model verification:
1. a data set;
to evaluate the performance of the QoS deep learning prediction model, experiments were conducted on the real world QoS dataset WS-DREAM. The data set includes QoS (throughput and response time) records for 5825 Web services, 339 users, and 1,947,675 users invoking the services, and each user and service record contains geographic information. This data set is widely used in QoS prediction work. Experiments were performed with matrix densities of 5%, 10%, 15%, 20%, respectively, taking existing items at different matrix densities as training sets, and selecting 200000 data at each density randomly as test sets.
2. Evaluation indexes are as follows:
the Mean Absolute Error (MAE) and the Root Mean Square Error (RMSE) are used as evaluation indexes for evaluating the prediction results, and the indexes are widely used in QoS prediction research work, as shown in formulas (15) and (16):
Figure BDA0003258606270000191
Figure BDA0003258606270000192
the MAE is an average value of absolute values of errors and can represent an average distance between a predicted value and a true value, and the smaller the value is, the more accurate the QoS prediction is. RMSE is the square root of the mean and the square of the deviation between the predicted and true values, with smaller values indicating smaller QoS prediction error margins.
3. Parameter setting and comparing method
The experimental parameter settings are shown in fig. 12.
In VI-D, we will experimentally determine the parameters that optimize the model performance in the parameter ranges set forth above, and evaluate the model performance using the optimal parameters in VI-F.
To evaluate the model herein, a comparison was made with the following exemplary methods:
3.1, PMF: a classical model-based collaborative filtering method uses probability matrix factorization to predict QoS values. It is used as a benchmark in NDMF and CNMF.
3.2, NDMF: a QoS prediction collaborative filtering model based on a deep neural network only uses global and local features, reveals implicit features of users and services through a complex nonlinear interaction function, and predicts a QoS value by finding out a user field through fusing user geographic information and user service call records.
3.3, CNMF: a QoS collaborative filtering model based on neighborhood perception matrix decomposition only uses local features to integrate neighborhood information of users and services into a matrix decomposition model for QoS prediction.
4. Parameter analysis
To analyze the effect of different parameter settings on the prediction accuracy of the model, the following experiment was performed to determine the optimal values of the parameters of the model.
4.1, number of similar users and similar services:
parameters J and K represent the number of similar users and similar services used in extracting local features, respectively. To study the effect of J and K on the prediction accuracy of the model herein, the individual feature dimension was set to 50 and the number of first three stages of fully connected layers was set to 3 with matrix densities (i.e., the ratio of data existing in the user-service QoS raw matrix to total data) of 5%, 10%, 15%, and 20%, respectively. The results of the experiment are shown in FIGS. 12 and 13.
As shown in fig. 12 and 13, when the number of similar users or services is set to 5 to 30, respectively, the prediction accuracy at different matrix densities fluctuates only slightly. The proposed model has better results in response time and throughput if Q has a higher density.
4.2, individual feature dimension:
the individual feature dimension represents the number of individual features of the user and the cloud service decomposed by the user-service original matrix Q, and determines how many individual features are used for QoS prediction. To investigate the effect of the individual feature dimensions on the prediction accuracy, we set J to 20, K to 10, and the number of first three-stage fully-connected layers to 3 at four different matrix densities (5%, 10%, 15%, 20%). The results of the experiment are shown in FIG. 14.
As shown in fig. 13 and 14, in the case of four matrix densities, the accuracy of the cloud service QoS prediction is improved as the feature dimension of each individual increases. Because more individual features can be mined out with higher dimensionality, the MM-DNN network is enabled to learn the features more effectively, and higher prediction accuracy is achieved. Meanwhile, as the matrix density is increased from 5% to 20% and the dimension is increased from 10 to 50, the higher the matrix density is, the more obvious the prediction accuracy is, and the prediction accuracy is obviously improved. The reason is that when the user-service original QoS matrix is too sparse, individual characteristics of the user and the service cannot be accurately obtained through non-negative matrix decomposition, thereby affecting the model performance. And as the density of the user-service original QoS matrix is increased, the individual characteristics can be more accurately extracted by the non-negative matrix decomposition, so that the prediction accuracy of the model is improved. Therefore, we can use a higher individual feature dimension to obtain better QoS prediction accuracy.
4.3, number of layers:
to study the effect of the number of fully-connected layers per stage on the prediction accuracy in this MM-DNN, the matrix density was set to 15%, the individual feature dimension was set to 10, J was set to 10, and K was set to 20. The results of the experiment are shown in FIG. 16.
Experiments show that the prediction accuracy is increased with the increase of the number of fully connected layers. This is because deeper networks can better learn the intrinsic relationships in the sample. However, with the increase of the number of layers, the improvement range of the prediction accuracy is smaller and smaller, the network parameter number and the calculation complexity are greatly improved, and meanwhile, the risk of overfitting is also improved. The number of full-connection layers was set to 4 in the following experiment.
5. Ablation experiment
This section has carried out the ablation experiment, has compared three kinds of different settings of model: global feature + local feature; global features + individual features; local features + individual features.
5.1, global feature + local feature: the global features and the local features are fed into the MM-DNN together for QoS prediction.
5.2, global feature + individual feature: the global features and the individual features are fed into the MM-DNN together for QoS prediction.
5.3, local features + individual features: the local features and the individual features are fed into the MM-DNN together for QoS prediction.
5.4, global feature + local feature + individual feature: all features are sent to the MM-DNN for QoS prediction.
As can be seen from fig. 17, from the first row to the fourth row, both MAE and RMSE of response time and throughput gradually decrease. Thus, the best results are obtained using three dimensions of features. The reason is that the global features are a summary of the overall information and the individual features are a detailed description of the detailed information. The local features are in between, and are a detailed description of the global features and a generalization of the individual features. As can be seen from the above ablation experiments, different characteristics contribute to the improvement of the prediction result. Therefore, the model using all the features is optimized in terms of QoS prediction accuracy.
6. Time consumption
In order to analyze the influence of the network structure of MM-DNN on time consumption, the matrix density was set to 15%, and the number of fully-connected layers for stages 1 to 3 was set to 2,3, and 4, respectively, as shown in fig. 15. The results show that the time consumption increases with the increase of the fully connected layer. Because the more the number of network layers, the more parameters need to be trained, resulting in high computational complexity.
In order to analyze the influence of the matrix density on the time consumption, experiments were performed with the number of fully-connected layers of stages 1 to 3 set to 4, and the matrix densities set to 5%, 10%, 15%, and 20%, respectively, as shown in fig. 16. As can be seen from fig. 16, as the matrix density increases, the training time consumption increases, and the prediction time consumption remains almost constant. The reason is that the higher the matrix density, the more data in the training set, while the test set remains unchanged. When the matrix density is set to 5% and 10%, the training time consumption is greater than the prediction time consumption because in these cases the amount of data in the training set is less than the amount of data in the test set.
7. Comparative experiment
The four methods were compared for MAE and RMSE at four different matrix densities (5%, 10%, 15%, 20%). Fig. 17 shows the MAE and RMSE of the response times of the different prediction methods, and fig. 18 shows the MAE and RMSE of the throughputs of the different prediction methods.
The results show that the prediction accuracy of MM-DNN is comprehensively superior to the other three methods for different QoS attributes. It can be seen that MM-DNN achieves the best results at different matrix densities. This demonstrates the effectiveness of staged multi-scale feature fusion. As the density of the matrix increases, the prediction accuracy of all methods increases, because more data can effectively improve the prediction accuracy.
The cloud service QoS deep learning prediction model is stored within a computer readable memory.
The computer-readable storage medium may be an internal storage unit of the terminal device in any of the foregoing embodiments, for example, a hard disk or a memory of the terminal device; the computer-readable storage medium may also be an external storage device of the terminal device, such as a plug-in hard disk, a Smart Media Card (SMC), a Secure Digital (SD) Card, a Flash memory Card (Flash Card), and the like provided in the terminal device. Further, the computer-readable storage medium may include both an internal storage unit and an external storage device of the terminal device. The computer-readable storage medium stores the computer program and other programs and data required by the terminal device. The above-described computer-readable storage medium may also be used to temporarily store data that has been output or is to be output.
The units described as separate parts may or may not be physically separate, and parts displayed as units may or may not be physical units, may be located in one place, or may be distributed on a plurality of network units. Some or all of the units can be selected according to actual needs to achieve the purpose of the solution of the embodiment of the present invention.
In addition, functional units in the embodiments of the present invention may be integrated into one processing unit, or each unit may exist alone physically, or two or more units are integrated into one unit. The integrated unit can be realized in a form of hardware, and can also be realized in a form of a software functional unit.
The above description is only an alternative embodiment of the application and is illustrative of the technical principles applied. It will be appreciated by those skilled in the art that the scope of the invention herein disclosed is not limited to the particular combination of features described above, but also encompasses other arrangements formed by any combination of the above features or their equivalents without departing from the spirit of the invention. For example, the above features may be replaced with (but not limited to) features having similar functions disclosed in the present application.
The foregoing is illustrative of only alternative embodiments of the present application and is not intended to limit the present application, which may be modified or varied by those skilled in the art. Any modification, equivalent replacement, improvement and the like made within the spirit and principle of the present application shall be included in the protection scope of the present application.

Claims (10)

1. A cloud service QoS deep learning prediction model is characterized in that the model building process comprises the following steps:
s1: determining the basic concept of the model, wherein the basic concept is as follows: multi-scale feature, global feature matrix
Figure FDA0003258606260000011
Local feature matrix (G), individual feature matrix and individual evaluation
Figure FDA0003258606260000012
S2: according to the determined basic concept, a QoS deep learning prediction model is provided, and the model is divided into three working stages: preprocessing, feature extraction and deep neural network prediction, wherein the preprocessing comprises the steps of carrying out nonnegative matrix decomposition on a user-service QoS (quality of service) original matrix, the feature extraction is to use the nonnegative matrix decomposition on the user-service QoS original matrix to obtain a global feature matrix and an individual feature matrix, respectively extracting global features and individual features from the global feature matrix and extracting local features based on distance attributes, and the deep neural network prediction processes the features and finally predicts a missing QoS value by designing MM-DNN;
s3: an algorithm process is provided, and comprises multi-scale feature extraction and MM-DNN learning;
s4: evaluating the performance of a QoS deep learning prediction model, performing an experiment on a QoS data set WS-DREAM of a real world, adopting an average absolute error (MAE) and a Root Mean Square Error (RMSE) as evaluation indexes for evaluating a prediction result, setting test parameters, determining parameters for optimizing the performance of the model in a set parameter range, evaluating the performance of the model by using the optimal parameters in VI-F, analyzing the influence of different parameter settings on the prediction accuracy of the model, determining the optimal values of the parameters of the model, performing an ablation experiment, comparing three different settings of the model, analyzing the influence of a network structure and matrix density of MM-DNN on time consumption, and finally comparing the MAE and the RMSE of the four methods under four different matrix densities (5%, 10%, 15% and 20%).
2. The cloud service QoS deep learning prediction model of claim 1, wherein the multi-scale features in S1 comprise global features, local features and individual features;
the global feature matrix is obtained by fitting a user-service QoS (quality of service) original matrix Q after non-negative matrix decomposition, and global features can be extracted from the matrix;
the local feature matrix is composed of a global feature matrix
Figure FDA0003258606260000021
The matrix which is formed by the fitting QoS values of the similar users to the similar cloud service is extracted, and local features can be obtained from the matrix;
the individual characteristic matrix comprises a user individual characteristic U and a cloud service individual characteristic matrix S;
the individual evaluation
Figure FDA0003258606260000022
Used for describing evaluation of cloud service by user, from global feature matrix
Figure FDA0003258606260000023
And the user i corrects the characteristics of each stage of the model according to the fitted QoS value of the cloud service j.
3. According to claim 1The cloud service QoS deep learning prediction model is characterized in that the preprocessing in S2 is to perform nonnegative matrix decomposition on a user-service QoS original matrix Q to obtain a nonnegative individual feature matrix of a user
Figure FDA0003258606260000024
Non-negative individual feature matrix of sum service
Figure FDA0003258606260000025
Each column U of UiAn l-dimensional individual feature vector representing user i, each column S in SjAn l-dimensional individual feature vector representing service j;
performing non-negative matrix factorization on Q to enable U, S inner products to be close to Q, wherein the formula (1) shows that:
Figure FDA0003258606260000026
Figure FDA0003258606260000027
in the formula (1), the reaction mixture is,
Figure FDA0003258606260000028
is a global feature matrix fitted by U, S;
obtaining a QoS value from a user-service QoS original matrix Q, learning to obtain U and S, and using the square of an error between the user-service QoS original matrix and a global feature matrix as a loss function, as shown in a formula (2):
Figure FDA0003258606260000029
the loss function in equation (2) is minimized using a multiplicative update rule, as shown in equation (3):
Figure FDA00032586062600000210
Figure FDA00032586062600000211
in the formula (3), ukiFor the k-th feature, s, of the user i in the user individual feature matrixkjThe kth feature of the service j in the service individual feature matrix is used.
4. The cloud service QoS deep learning prediction model of claim 1, wherein in the S2, the global feature is extracted from a global feature matrix
Figure FDA0003258606260000031
Extracting evaluation of user i to all cloud services
Figure FDA0003258606260000032
And all users' evaluations (U) of service jTSj) As global features, global features
Figure FDA0003258606260000033
Is recorded as:
Figure FDA0003258606260000034
in the formula (4), the reaction mixture is,
Figure FDA0003258606260000035
() For the merge operation, the two vectors are concatenated together.
5. The cloud service QoS deep learning prediction model of claim 1, wherein in the S2, local feature extraction is performed first based on user similarity based on distanceDegree and cloud service similarity calculation, distance d between users b, vbvCalculating as shown in equation (5):
c＝sin(γb)·sin(γv)+cos(γb)·cos(γv)·cos(θb-θv)
Figure FDA0003258606260000036
in the formula (5), γb，θb∈(-180,180]Respectively representing the latitude and longitude of the user b, r is the earth radius, dbvSmall means that two users are geographically close and have more similar QoS values, cxyServing the distance similarity between x and y for the cloud;
then sorting the distances, selecting similar users and similar services according to the distances, fitting QoS values to form a local feature matrix G, converting the local feature matrix G into a one-dimensional vector serving as a local feature
Figure FDA0003258606260000037
As shown in formula (6);
Figure FDA0003258606260000038
in the formula (6), the reaction mixture is,
Figure FDA0003258606260000039
() For the shape transformation function, the matrix is converted into the vector, and the user u is extracted1、u3For cloud services s4、s5And (4) fitting the QoS value to generate a local feature matrix G, and converting the generated local feature matrix G into a one-dimensional vector as a local feature.
6. The cloud service QoS deep learning prediction model according to claim 1, wherein the individual feature extraction in S2 is to extract a feature vector of a user i from a user individual feature matrix UQuantity UiExtracting a feature vector S of the cloud service j from the service individual feature matrix S as the user individual featurejAs serving individual characteristics, individual characteristics
Figure FDA00032586062600000310
Expressed as:
Figure FDA0003258606260000041
7. the cloud service QoS deep learning prediction model of claim 1, wherein the deep neural network prediction structure in S2 is divided into four stages:
the stage 1 has a merging layer and L full-connection layers, global features are input, after the global features pass through one full-connection layer, the features are merged in the merging layer for individual evaluation and are corrected, information with the same size as the local features in the features is further extracted through the full-connection layer, and a feed-forward process of the stage can be expressed as follows:
Figure FDA0003258606260000042
Figure FDA0003258606260000043
Figure FDA0003258606260000044
Figure FDA0003258606260000045
in the formula (8), y0As input of MM-DNN, ykIs the output of the kth fully-connected layer, αkAnd betakRespectively representing the weight and the offset of the k-th fully-connected layer,
Figure FDA0003258606260000046
to activate the function ReLU (Rectified Linear Unit, ReLU), as shown in equation (9):
Figure FDA0003258606260000047
the stage 2 comprises two merging layers and M full-connection layers, local features are merged in one merging layer, after passing through one full-connection layer, individual evaluation is merged in the merging layer, then the full-connection layer is fed in, and the process is as shown in the formula (10):
Figure FDA0003258606260000048
Figure FDA0003258606260000049
Figure FDA00032586062600000410
stage 3 has two merging layers and Z full-connection layers, connects individual characteristics in one merging layer, merges individual evaluation in the merging layer after one full-connection layer, learns through the full-connection layer, and the process is as shown in formula (11):
Figure FDA0003258606260000051
Figure FDA0003258606260000052
Figure FDA0003258606260000053
the first three stages all use BN and Dropout;
stage 4 has a merging layer in which individual evaluations are concatenated and a full-link layer
Figure FDA0003258606260000054
And (3) further learning through a full connection layer, thereby correcting the result and outputting a QoS predicted value, as shown in formula (12):
Figure FDA0003258606260000055
in the parameter training process, the Mean Absolute Error (MAE) is used as a loss function, as shown in formula (13):
Figure FDA0003258606260000056
in the formula (13), pijThe QoS predicted value of the user i to the cloud service j is obtained for the QoS deep learning prediction model, and N is the number of the QoS values needing to be predicted;
parameters in the network are optimized, a gradient descent method is adopted for model training, and loss is minimized, as shown in formula (14):
Figure FDA0003258606260000057
Figure FDA0003258606260000058
in the formula (14), λ is a learning rate for controlling gradient descent speed in an iterative process, k is equal to {1,2, …, L + M + Z }, and an initial value α is1And beta1Obtained by generating random numbers.
8. The cloud service QoS deep learning prediction model of claim 1, wherein in S3, multi-scale features are extracted as algorithm 1, MM-DNN is learned as algorithm 2, algorithm 1 obtains global features, local features and individual features, the three features are used as input of algorithm 2 and are subjected to staged feature fusion in MM-DNN, and finally algorithm 2 returns a QoS prediction value of a user for a cloud service.
9. The cloud service QoS deep learning prediction model of claim 1, wherein in S4, a Mean Absolute Error (MAE) and a Root Mean Square Error (RMSE) are used as evaluation indicators for evaluating a prediction result, as shown in formulas (15) (16):
Figure FDA0003258606260000061
Figure FDA0003258606260000062
in the formula (15), MAE is an average value of absolute values of errors and represents an average distance between a predicted value and a true value;
in equation (16), RMSE is the square root of the mean and the square of the deviation between the predicted value and the true value.
10. The cloud service QoS deep learning prediction model according to claim 1, wherein the cloud service QoS deep learning prediction model is stored in a computer-readable memory, and the memory is configured to run the cloud service QoS deep learning prediction model, and execute the cloud service QoS deep learning prediction model according to any one of claims 1 to 9.
CN202111066541.0A 2021-09-13 2021-09-13 Construction method and system of cloud service QoS deep learning prediction model Active CN113743675B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202111066541.0A CN113743675B (en) 2021-09-13 2021-09-13 Construction method and system of cloud service QoS deep learning prediction model

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202111066541.0A CN113743675B (en) 2021-09-13 2021-09-13 Construction method and system of cloud service QoS deep learning prediction model

Publications (2)

Publication Number Publication Date
CN113743675A true CN113743675A (en) 2021-12-03
CN113743675B CN113743675B (en) 2024-01-30

Family

ID=78738192

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202111066541.0A Active CN113743675B (en) 2021-09-13 2021-09-13 Construction method and system of cloud service QoS deep learning prediction model

Country Status (1)

Country Link
CN (1) CN113743675B (en)

Cited By (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN114757391A (en) * 2022-03-17 2022-07-15 重庆大学 Service quality prediction method based on network data space design
CN115277521A (en) * 2022-06-06 2022-11-01 浙大城市学院 Multi-view-angle-based dynamic prediction method for QoS (quality of service) of Internet of things
CN116011662A (en) * 2023-02-02 2023-04-25 南京信息工程大学 Service QoS prediction method based on pyramid structure multi-feature extraction
CN117649153A (en) * 2024-01-29 2024-03-05 南京典格通信科技有限公司 Mobile communication network user experience quality prediction method based on information integration

Citations (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN106447103A (en) * 2016-09-26 2017-02-22 河海大学 Deep learning based QoS prediction method of Web service
US20210105338A1 (en) * 2020-01-06 2021-04-08 Intel Corporation Quality of service (qos) management with network-based media processing (nbmp)

Patent Citations (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN106447103A (en) * 2016-09-26 2017-02-22 河海大学 Deep learning based QoS prediction method of Web service
US20210105338A1 (en) * 2020-01-06 2021-04-08 Intel Corporation Quality of service (qos) management with network-based media processing (nbmp)

Non-Patent Citations (1)

* Cited by examiner, † Cited by third party
Title
WENJUN HUANG 等: "QoS Prediction Model of Cloud Services Based on Deep Learning", 《IEEE/CAA JOURNAL OF AUTOMATICA SINICA》, vol. 9, no. 3, pages 564, XP011896360, DOI: 10.1109/JAS.2021.1004392 *

Cited By (7)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN114757391A (en) * 2022-03-17 2022-07-15 重庆大学 Service quality prediction method based on network data space design
CN114757391B (en) * 2022-03-17 2024-05-03 重庆大学 Network data space design and application method oriented to service quality prediction
CN115277521A (en) * 2022-06-06 2022-11-01 浙大城市学院 Multi-view-angle-based dynamic prediction method for QoS (quality of service) of Internet of things
CN115277521B (en) * 2022-06-06 2023-10-03 浙大城市学院 Multi-view-based dynamic prediction method for QoS (quality of service) of Internet of things
CN116011662A (en) * 2023-02-02 2023-04-25 南京信息工程大学 Service QoS prediction method based on pyramid structure multi-feature extraction
CN117649153A (en) * 2024-01-29 2024-03-05 南京典格通信科技有限公司 Mobile communication network user experience quality prediction method based on information integration
CN117649153B (en) * 2024-01-29 2024-04-16 南京典格通信科技有限公司 Mobile communication network user experience quality prediction method based on information integration

Also Published As

Publication number Publication date
CN113743675B (en) 2024-01-30

Similar Documents

Publication Publication Date Title
Zhang et al. A cross-domain recommender system with kernel-induced knowledge transfer for overlapping entities
CN113743675A (en) Cloud service QoS deep learning prediction model
Dash et al. An outliers detection and elimination framework in classification task of data mining
Tölö Predicting systemic financial crises with recurrent neural networks
US11210368B2 (en) Computational model optimizations
CN112508265A (en) Time and activity multi-task prediction method and system for business process management
CN112949821B (en) Network security situation awareness method based on dual-attention mechanism
Jiang et al. A fast deep autoencoder for high-dimensional and sparse matrices in recommender systems
CN115983984A (en) Multi-model fusion client risk rating method
CN115080868A (en) Product pushing method, product pushing device, computer equipment, storage medium and program product
CN115310589A (en) Group identification method and system based on depth map self-supervision learning
Sadouk et al. A novel cost‐sensitive algorithm and new evaluation strategies for regression in imbalanced domains
Enikeeva et al. Change-point detection in dynamic networks with missing links
Lee et al. An entropy decision model for selection of enterprise resource planning system
Renström et al. Fraud Detection on Unlabeled Data with Unsupervised Machine Learning
Lo Predicting software reliability with support vector machines
CN114462707B (en) Web service multidimensional QoS joint prediction method based on feature depth fusion
CN115659277A (en) E-commerce session recommendation method, system, device and medium based on multi-behavior feature fusion
CN115759036A (en) Method for constructing recommendation-based event detection model and method for detecting event by using model
CN112463964B (en) Text classification and model training method, device, equipment and storage medium
Xiong et al. L-RBF: A customer churn prediction model based on lasso+ RBF
Badyal et al. Insightful Business Analytics Using Artificial Intelligence-A Decision Support System for E-Businesses
Lu et al. Tensor mutual information and its applications
CN117274616B (en) Multi-feature fusion deep learning service QoS prediction system and prediction method
Davis et al. NIF: A framework for quantifying neural information flow in deep networks

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