Multi-feature fusion deep learning QoS prediction method and device and medium

Abstract

The invention discloses a multi-feature fusion deep learning QoS prediction method and device and a medium, and the method comprises the steps: carrying out the initialization of an original QoS matrix Q, obtaining a user potential feature matrix and a service potential feature matrix, inputting the user potential feature matrix and the service potential feature matrix into a similar feature set extraction network, obtaining a similar user potential feature vector set of the user and a similar service potential feature vector set of the service; inputting the similar user potential feature vector set and the similar service potential feature vector set into a feature conversion inference network to obtain feature vectors of the users and feature vectors of the services; inputting the position information into a position feature extraction network, and extracting position feature vectors of the user and the service; and inputting the feature vector of the user, the feature vector of the service and the location feature vector of the user and the service into a feature fusion and prediction network, and outputting a predicted QoS value. The problem that proper services are difficult to recommend to users can be solved.

Description

Technical Field

[0001] The present invention relates to the field of communication technology, and in particular to a multi-feature fusion deep learning QoS prediction method, device and medium. Background Art

[0002] With the popularity of service computing, more and more service providers choose to publish services on the cloud for users to call. However, with the rapid growth of the number of services, many services have the same or similar functions. How to choose the right service to meet user needs has become a key issue. In this case, Quality of Service (QoS) has become an important basis for service selection.

[0003] QoS is used to describe the non-functional attributes of services, such as response time and throughput. However, in many cases it is difficult to obtain the QoS value of the service called by the user. There are two main reasons for this: on the one hand, the QoS called by the user depends on the contextual information of the user and the service. For example, different users are generally in different network environments, so when different users call the same service, their QoS values ​​may be different; on the other hand, users will not call all services, so the QoS of many services is unknown to users who have not called the service, and the lack of service QoS values ​​makes it difficult to recommend services to users. To solve this problem, researchers predict the QoS value of unknown services by using the QoS value of known called services. However, how to accurately predict the QoS value of services remains one of the main challenges facing service computing.

[0004] There are many prediction methods currently, among which collaborative filtering (CF) is the most widely used method. CF-based QoS prediction methods are based on memory-based and model-based collaborative filtering. The memory-based collaborative filtering method predicts QoS values ​​through similar users or services, where similar users or services can be calculated through historical call records. However, this type of method is easily affected by data sparsity, resulting in reduced QoS prediction accuracy. To alleviate the impact of data sparsity, model-based collaborative filtering methods usually use matrix factorization (MF) methods to learn user and service potential features from historical QoS values. This method usually achieves better QoS prediction performance than the memory-based collaborative filtering method, but the traditional matrix factorization method is to linearly combine the potential feature vectors of users and services through inner products, which is difficult to fully capture the deeper relationship between users and services. It is easy to lose information in low-dimensional space transformation, which affects the prediction effect of the QoS value of unknown services, making it difficult to recommend suitable services to users. Summary of the invention

[0005] The purpose of the present invention is to overcome the deficiencies in the prior art and provide a multi-feature fusion deep learning QoS prediction method, device and medium, which can solve the problem that the traditional matrix decomposition method linearly combines the potential feature vectors of users and services through inner products, which is difficult to fully capture the deeper relationship between users and services, and is prone to information loss in low-dimensional space transformation, thereby affecting the prediction effect of the QoS value of unknown services, making it difficult to recommend suitable services to users.

[0006] To achieve the above object, the present invention is implemented by adopting the following technical solutions:

[0007] In a first aspect, the present invention provides a multi-feature fusion deep learning QoS prediction method, comprising:

[0008] Construct the original QoS matrix Q by collecting the real QoS values ​​returned by users calling the service;

[0009] Initialize the original QoS matrix Q to obtain a user potential feature matrix and a service potential feature matrix, input the user potential feature matrix and the service potential feature matrix into a similar feature set extraction network of a pre-built prediction model, and obtain a user similar user potential feature vector set and a service similar service potential feature vector set;

[0010] Inputting the similar user potential feature vector set and the similar service potential feature vector set into a feature conversion inference network of a pre-built prediction model to obtain a feature vector of the user and a feature vector of the service;

[0011] Obtaining the location information of the user and the service through the real QoS value returned by the user calling the service, inputting the location information into the location feature extraction network of the pre-built prediction model, and extracting the location feature vector of the user and the service;

[0012] The user's feature vector, the service's feature vector, and the user and service's location feature vectors are input into a feature fusion and prediction network of a pre-built prediction model, and a predicted QoS value is output.

[0013] Furthermore, the original QoS matrix Q is initialized to obtain a user potential feature matrix and a service potential feature matrix including:

[0014] The original QoS matrix Q∈ Decomposed into user potential feature matrix and the service latent feature matrix ;

[0015] Minimize the original QoS matrix Q and the user latent feature matrix by the multiplication update rule and the service latent feature matrix The difference between:

[0016] ,

[0017] ,

[0018] ,

[0019] ,

[0020] Among them, t and w are the number of users and the number of services, respectively, l is the dimension of the user latent feature vector and the service latent feature vector, ||·|| F is the Frobenius norm, represents the kth feature of service j in the service potential feature matrix S, represents the kth feature of user i in the user potential feature matrix U, and T represents the transposed matrix.

[0021] Furthermore, inputting the user potential feature matrix and the service potential feature matrix into a similar feature set extraction network of a pre-built prediction model to obtain a user similar user potential feature vector set and a service similar service potential feature vector set includes:

[0022] The similarity between two users or services is calculated by the Pearson coefficient. The calculation formula is as follows:

[0023] ,

[0024] represents the Pearson correlation coefficient between user i and user e, and are the QoS values ​​returned by user i and user e when they call service h, and Respectively represent the average QoS values ​​of the services called by user i and user e, represents the set of common services called by user i and user e,

[0025] By calculating the Pearson correlation coefficient similarity between user i and other users, the top K users with the highest similarity to user i are selected, and the potential feature vectors of these users are combined into a set of similar user potential feature vectors for user i. ;

[0026] By calculating the Pearson correlation coefficient similarity between service j and other services, the top K services with the highest similarity to service j are selected, and the potential feature vectors of these services are combined into a similar service potential feature vector set of service j. .

[0027] Further, the similar user potential feature vector set Input the feature conversion inference network of the pre-built prediction model to obtain the feature vector V of user i i include:

[0028] The feature conversion inference network includes a feature conversion module, a multi-layer feature extraction module and a feature inference module;

[0029] The similar user potential feature vector set is transformed into Perform shape transformation operations into matrices , and the matrix Input the feature conversion module to obtain the feature matrix of user i , expressions include:

[0030] ,

[0031] ,

[0032] ,

[0033] in, is the output of the mth convolutional layer, Bm, Cm and Ym are the weight, bias and stride of the mth convolutional layer respectively, Conv() operation represents the convolution operation, Bn() operation represents the batch normalization operation, () represents the shape transformation operation, is the output of the Nth convolutional layer;

[0034] The multi-layer feature extraction module further extracts features of different scales output by each convolution layer in the feature conversion module to obtain a merge matrix , expressions include:

[0035] ,

[0036]

[0037] in Represents the output of the enhanced mth convolutional layer , , and Represent the weight matrix, bias term and stride of the mth convolutional layer respectively, Represents the merge matrix of the output features of the multi-layer feature extraction module; © represents the merge operation;

[0038] The outputs of the feature conversion module and the multi-level feature extraction module are subjected to feature fusion by the feature inference module, and the vector after feature fusion is converted into a one-dimensional feature vector by a shape transformation operation. , and finally the one-dimensional feature vector Input feature inference module to extract features of user i and obtain feature vector V of user i i ;

[0039] ,

[0040] ,

[0041] ,

[0042] in, represents the output of the mth fully connected layer, and They respectively show the weight and bias of the mth fully connected layer, () represents a rectified linear unit.

[0043] Furthermore, inputting the location information into a location feature extraction network of a pre-built prediction model to extract location feature vectors of users and services includes:

[0044] The location information is input into the location input layer of the location feature extraction network to generate initial features of user and service location information. The initial features of the user and service location information are expressed as:

[0045] ,

[0046] ,

[0047] in, is the initial feature of the user's location information, , and denote the ID, RE and AS of user i respectively, is the initial feature of the service location information, , and denote the ID, RE and AS of service j respectively;

[0048] One-hot encoding is used to convert the initial features of the user and service location information into a high-dimensional sparse vector matrix. The high-dimensional sparse vector matrix of the user and service location information is expressed as:

[0049] ,

[0050] ,

[0051] in, is the high-dimensional sparse vector matrix of user location information, and They represent the corresponding high-dimensional sparse vectors obtained after one-hot encoding of user i’s ID, RE, and AS. is a high-dimensional sparse vector matrix of service location information, and They represent the corresponding high-dimensional sparse vectors obtained after one-hot encoding of the ID, RE and AS of service j;

[0052] The high-dimensional sparse vector of the user and service location information is converted into a low-dimensional dense embedding feature vector through the location embedding layer of the location feature extraction network. The conversion process is expressed as:

[0053] ,

[0054] ,

[0055] ,

[0056] ,

[0057] ,

[0058] ,

[0059] In the formula, , and represents the embedding weight matrix of user i, , and represents the embedding weight matrix of service j, D i , R i and A i are the embedded features of user i’s ID, RE and AS output in the embedding layer, D j , R j and A j are the embedded features output by the embedding layer of service j’s ID, RE, and AS respectively;

[0060] The embedded features of user i and service j are merged to generate the corresponding embedded feature vector F i and F j , the corresponding embedding feature vector F i and F j The embedding feature matrix L is obtained by combining the merging and shape transformation operations. ij :

[0061] ,

[0062] ,

[0063] ,

[0064] The embedding feature matrix L is extracted through the convolution layer of the potential feature extraction and output layer in the position feature extraction network. ij The features of are transformed and a shape transformation operation is performed to obtain a conversion embedding feature matrix, the conversion embedding feature matrix is ​​input into the fully connected layer of the potential feature extraction and output layer to obtain a feature output vector, the feature output vector is input into the Softmax() function of the potential feature extraction and output layer to calculate the probability distribution of the feature output vector, and the probability distribution of the feature output vector is subjected to a Hadamard product operation with the feature output vector to obtain the position feature vector of user i and service j. :

[0065] ,

[0066] ,

[0067] ,

[0068] in, represents the transformed embedding feature matrix, represents the feature output vector, Z, M and E represent the weight matrix, bias term and stride of the convolutional layer, and represents the weight matrix of the fully connected layer, and represents the bias term of the fully connected layer, and the Softmax() function represents the probability distribution operation function.

[0069] Furthermore, the user's feature vector, the service's feature vector, and the user and service's location feature vector are input into a feature fusion and prediction network of a pre-built prediction model, and the output prediction QoS value includes:

[0070] The feature vector of the user and the feature vector of the service are merged to obtain a merged feature vector, the merged feature vector is merged with the position feature vector of the user and the service to obtain a final merged feature vector, the final merged feature vector is input into the Γ fully connected layers of the feature fusion and prediction network, the interaction relationship between the features is learned, and the predicted QoS value is output in the last Γth fully connected layer , the specific expression is as follows:

[0071] ,

[0072] ,

[0073] ,

[0074] in, represents the merged eigenvector, represents the final merged eigenvector, and represents the weight and bias of the mth fully connected layer in the network, It is the output of the mth layer of the feature fusion prediction network.

[0075] Furthermore, the training method of the prediction model includes:

[0076] The mean square error is used as the loss function, and the expression of the loss function is:

[0077] ,

[0078] Among them, q ij The QoS value returned by user i when calling service j, is the predicted QoS value of user i for service j output by the prediction model;

[0079] The parameters of the prediction model are updated by the stochastic gradient descent method so that the loss function of the prediction model is minimized. The specific expression is:

[0080] ,

[0081] ,

[0082] ,

[0083] in, are the parameters in the feature transformation inference network, Extract the parameters in the network for position features, Parameters in the feature fusion prediction network, The learning rate is used to control the speed of gradient descent during the pre-set iteration.

[0084] In a second aspect, the present invention provides a multi-feature fusion deep learning QoS prediction device, comprising:

[0085] The matrix construction module is used to construct the original QoS matrix Q by collecting the real QoS values ​​returned by the user calling the service;

[0086] A matrix processing module, used for initializing the original QoS matrix Q, obtaining a user potential feature matrix and a service potential feature matrix, inputting the user potential feature matrix and the service potential feature matrix into a similar feature set extraction network of a pre-built prediction model, and obtaining a user similar user potential feature vector set and a service similar service potential feature vector set;

[0087] A feature vector acquisition module, used for inputting the similar user potential feature vector set and the similar service potential feature vector set into a feature conversion inference network of a pre-built prediction model to obtain the user's feature vector and the service's feature vector;

[0088] A location feature vector acquisition module, used to obtain the location information of the user and the service through the real QoS value returned by the user calling the service, input the location information into the location feature extraction network of the pre-built prediction model, and extract the location feature vector of the user and the service;

[0089] The output module is used to input the user's feature vector, the service's feature vector and the user and service's location feature vector into a feature fusion and prediction network of a pre-built prediction model, and output a predicted QoS value.

[0090] In a third aspect, the present invention provides an electronic terminal, comprising a processor and a memory connected to the processor, wherein a computer program is stored in the memory, and when the computer program is executed by the processor, the steps of the method described in the first aspect are performed.

[0091] In a fourth aspect, the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the method described in the first aspect.

[0092] Compared with the prior art, the present invention has the following beneficial effects:

[0093] For the QoS matrix, features are constructed through matrix decomposition. However, information is easily lost in low-dimensional space transformation. Therefore, the present invention first promotes the potential feature vector of the user / service to a high-dimensional user / service feature matrix to enrich the feature information, and then further extracts the feature information output in the feature conversion module through a multi-layer feature extraction module to further extract richer and more complex features; finally, the feature inference module is used to fuse the features in the feature conversion module and the multi-layer feature extraction module to further extract the potential features of the user / service, and output the feature vector of the user / service, thereby optimizing the features of the user / service, improving the prediction effect of the QoS value of the unknown service, and being able to recommend suitable services to the user. BRIEF DESCRIPTION OF THE DRAWINGS

[0094] Figure 1A method flow chart of a multi-feature fusion deep learning QoS prediction method provided in Embodiment 1 of the present invention;

[0095] Figure 2 A prediction model framework diagram of a multi-feature fusion deep learning QoS prediction method provided in the first embodiment of the present invention;

[0096] Figure 3 A schematic diagram of a feature conversion inference network of a prediction model in a multi-feature fusion deep learning QoS prediction method provided in the first embodiment of the present invention;

[0097] Figure 4 A schematic diagram of a network for extracting location features of a prediction model in a multi-feature fusion deep learning QoS prediction method provided in the first embodiment of the present invention;

[0098] Figure 5 A schematic diagram of feature fusion and prediction network of a prediction model in a multi-feature fusion deep learning QoS prediction method provided in the first embodiment of the present invention;

[0099] Figure 6 A schematic diagram of response time MAE with parameter c as an independent variable in a multi-feature fusion deep learning QoS prediction method provided in the first embodiment of the present invention;

[0100] Figure 7 A schematic diagram of response time RMSE with parameter c as an independent variable in a multi-feature fusion deep learning QoS prediction method provided in the first embodiment of the present invention;

[0101] Figure 8 A schematic diagram of response time MAE with parameter l as an independent variable in a multi-feature fusion deep learning QoS prediction method provided in the first embodiment of the present invention;

[0102] Fig. 9 A schematic diagram of response time RMSE with parameter l as an independent variable in a multi-feature fusion deep learning QoS prediction method provided in the first embodiment of the present invention;

[0103] Fig.10 A schematic diagram of response time MAE with parameter K as the independent variable in a multi-feature fusion deep learning QoS prediction method provided in the first embodiment of the present invention;

[0104] Fig.11 A schematic diagram of response time RMSE with parameter K as the independent variable in a multi-feature fusion deep learning QoS prediction method provided in the first embodiment of the present invention;

[0105] Fig.12A schematic diagram of response time MAE with parameter N as the independent variable in a multi-feature fusion deep learning QoS prediction method provided in the first embodiment of the present invention;

[0106] Fig.13 A schematic diagram of response time RMSE with parameter N as the independent variable in a multi-feature fusion deep learning QoS prediction method provided in the first embodiment of the present invention;

[0107] Fig.14 A schematic diagram of response time MAE with parameter Γ as an independent variable in a multi-feature fusion deep learning QoS prediction method provided in the first embodiment of the present invention;

[0108] Fig.15 A schematic diagram of response time RMSE with parameter Γ as the independent variable in a multi-feature fusion deep learning QoS prediction method provided in Example 1 of the present invention. DETAILED DESCRIPTION

[0109] The technical solution of the present invention is described in detail below through the accompanying drawings and specific embodiments. It should be understood that the embodiments of the present application and the specific features in the embodiments are detailed descriptions of the technical solution of the present application, rather than limitations on the technical solution of the present application. In the absence of conflict, the embodiments of the present application and the technical features in the embodiments can be combined with each other.

[0110] The term "and / or" in this article is only a description of the association relationship of the associated objects, indicating that there can be three relationships. For example, A and / or B can mean: A exists alone, A and B exist at the same time, and B exists alone. In addition, the character " / " in this article generally indicates that the associated objects before and after are in an "or" relationship.

[0111] Embodiment 1:

[0112] Figure 1 This is a flowchart of the multi-feature fusion deep learning QoS prediction method based on location perception in the first embodiment of the present invention. This flowchart only shows the logical sequence of the method described in this embodiment. Under the premise of no conflict, in other possible embodiments of the present invention, different Figure 1 The steps shown or described are accomplished in the order shown.

[0113] like Figure 1 As shown, the method disclosed in this embodiment specifically includes the following steps:

[0114] Step 1: Construct the original QoS matrix Q by collecting the real QoS values ​​returned by the user calling the service;

[0115] Step 2: Initialize the original QoS matrix Q to obtain the user potential feature matrix and the service potential feature matrix:

[0116] The original QoS matrix Q∈ Decomposed into user potential feature matrix and the service latent feature matrix , where NMF stands for Nonnegative Matrix Factorization (NMF), which can intuitively represent the relationship between users and services and takes up less storage space;

[0117] Minimize the original QoS matrix Q and the user latent feature matrix by the multiplication update rule and the service latent feature matrix The difference between:

[0118] ,

[0119] ,

[0120] ,

[0121] ,

[0122] where t and w are the number of users and services, respectively, l is the dimension of the user latent feature vector and the service latent feature vector, ||·|| F is the Frobenius norm, represents the kth feature of service j in the service potential feature matrix S, represents the kth feature of user i in the user potential feature matrix U, and T represents the transposed matrix;

[0123] The user potential feature matrix and the service potential feature matrix are input into the similar feature set extraction network of the pre-built prediction model. Based on the user potential feature matrix, a similar user potential feature vector set of the user can be obtained through similarity calculation. Based on the service potential feature matrix, a similar service potential feature vector set of the service can be obtained through similarity calculation:

[0124] Because PCC is easy to implement, the direction and relative size of the vectors are taken into account when calculating the similarity between two vectors, and it can more comprehensively measure the linear correlation between vectors and is widely used in recommendation systems. Therefore, this embodiment calculates the similarity between two users or services through the Pearson Correlation Coefficient (PCC), and the calculation formula is as follows:

[0125] ,

[0126] represents the Pearson correlation coefficient between user i and user e, and are the QoS values ​​returned by user i and user e when they call service h, and Respectively represent the average QoS values ​​of the services called by user i and user e, represents the set of common services called by user i and user e,

[0127] By calculating the Pearson correlation coefficient similarity between user i and other users, the top K users with the highest similarity to user i are selected, and the potential feature vectors of these users are combined into a set of similar user potential feature vectors for user i. ;

[0128] By calculating the Pearson correlation coefficient similarity between service j and other services, the top K services with the highest similarity to service j are selected, and the potential feature vectors of these services are combined into a similar service potential feature vector set of service j. ;

[0129] Step 3: Inputting the similar user potential feature vector set and the similar service potential feature vector set into the feature conversion inference network of the pre-built prediction model to obtain the user's feature vector and the service's feature vector includes:

[0130] The prediction model is as follows: Figure 2 As shown, the feature conversion inference network is as follows Figure 3 As shown, the feature conversion inference network includes a feature conversion module, a multi-layer feature extraction module and a feature inference module;

[0131] The similar user potential feature vector set is transformed into Perform shape transformation operations into matrices , and the matrix Input the feature conversion module to obtain the feature matrix of user i , expressions include:

[0132] ,

[0133] ,

[0134] ,

[0135] in, is the output of the mth convolutional layer, Bm, Cm and Ym are the weight, bias and stride of the mth convolutional layer respectively, Conv() operation represents the convolution operation, Bn() operation represents the batch normalization operation, () represents the shape transformation operation, is the output of the Nth convolutional layer;

[0136] The multi-layer feature extraction module further extracts features of different scales output by each convolution layer in the feature conversion module to obtain a merge matrix , expressions include:

[0137] ,

[0138]

[0139] in Represents the output of the enhanced mth convolutional layer , , and Represent the weight matrix, bias term and stride of the mth convolutional layer respectively, Represents the merge matrix of the output features of the multi-layer feature extraction module;

[0140] The outputs of the feature conversion module and the multi-level feature extraction module are subjected to feature fusion by the feature inference module, and the vector after feature fusion is converted into a one-dimensional feature vector by a shape transformation operation. , and finally the one-dimensional feature vector Input feature inference module to extract features of user i and obtain feature vector V of user i i ;

[0141] ,

[0142] ,

[0143] ,

[0144] in, represents the output of the mth fully connected layer, and They respectively show the weight and bias of the mth fully connected layer, () represents the rectified linear unit;

[0145] The similar service potential feature vector set Input the feature conversion inference network of the pre-built prediction model to obtain the feature vector H of service j j The process of obtaining The acquisition process is similar:

[0146] The similar service potential feature vector set is transformed by the feature conversion module Perform shape transformation operations into matrices , and the matrix Input the feature conversion module to obtain the feature matrix of service j , expressions include:

[0147] ,

[0148] ,

[0149] ,

[0150] in, is the output of the mth convolutional layer, , and are the weight, bias, and stride of the mth convolutional layer, respectively. The Conv() operation represents the convolution operation, and the Bn() operation represents the batch normalization operation. () represents the shape transformation operation, is the output of the Nth convolutional layer;

[0151] The multi-layer feature extraction module further extracts features of different scales output by each convolution layer in the feature conversion module to obtain a merge matrix , expressions include:

[0152] ,

[0153]

[0154] in Represents the output of the enhanced mth convolutional layer , , and Represent the weight matrix, bias term and stride of the mth convolutional layer respectively, Represents the merge matrix of the output features of the multi-layer feature extraction module;

[0155] The outputs of the feature conversion module and the multi-level feature extraction module are subjected to feature fusion by the feature inference module, and the vector after feature fusion is converted into a one-dimensional feature vector by a shape transformation operation. , and finally the one-dimensional feature vector Input feature inference module to extract the features of service j and obtain the feature vector H of service j j ;

[0156] ,

[0157] ,

[0158] ,

[0159] in, represents the output of the mth fully connected layer, and They respectively show the weight and bias of the mth fully connected layer, () represents the rectified linear unit;

[0160] Step 4: Obtain the location information of the user and the service through the real QoS value returned by the user calling the service. The location information of the user and the service includes regions and autonomous systems. Since the number of users, services, regions and autonomous systems is limited, and each user and service corresponds to only one region (RE) and one autonomous system (AS), this embodiment uses non-negative integers as unique identifiers to represent them respectively, inputs the location information into the location feature extraction network of the pre-built prediction model, and extracts the location feature vectors of the user and the service:

[0161] Wherein, the position feature extraction network is as follows: Figure 4 As shown, the location information is input into the location input layer of the location feature extraction network to generate initial features of user and service location information. The initial features of the user and service location information are expressed as:

[0162] ,

[0163] ,

[0164] in, is the initial feature of the user's location information, , and denote the ID, RE and AS of user i respectively, is the initial feature of the service location information, , and denote the ID, RE and AS of service j respectively;

[0165] One-hot encoding is used to convert the initial features of the user and service location information into a high-dimensional sparse vector matrix. The high-dimensional sparse vector matrix of the user and service location information is expressed as:

[0166] ,

[0167] ,

[0168] in, is the high-dimensional sparse vector matrix of user location information, and They represent the corresponding high-dimensional sparse vectors obtained after one-hot encoding of user i’s ID, RE, and AS. is a high-dimensional sparse vector matrix of service location information, and They represent the corresponding high-dimensional sparse vectors obtained after one-hot encoding of the ID, RE and AS of service j;

[0169] The high-dimensional sparse vector of the user and service location information is converted into a low-dimensional dense embedding feature vector through the location embedding layer of the location feature extraction network. The conversion process is expressed as:

[0170] ,

[0171] ,

[0172] ,

[0173] ,

[0174] ,

[0175] ,

[0176] In the formula, , and represents the embedding weight matrix of user i, , and represents the embedding weight matrix of service j, D i , R i and A i are the embedded features of user i’s ID, RE and AS output in the embedding layer, D j , R j and A j are the embedded features output by the embedding layer of service j’s ID, RE, and AS respectively;

[0177] The embedded features of user i and service j are merged to generate the corresponding embedded feature vector F i and F j , the corresponding embedding feature vector F i and F j The embedding feature matrix L is obtained by combining the merging and shape transformation operations. ij :

[0178] ,

[0179] ,

[0180] ,

[0181] The embedding feature matrix L is extracted through the convolution layer of the potential feature extraction and output layer in the position feature extraction network. ij The features of are transformed and a shape transformation operation is performed to obtain a conversion embedding feature matrix, the conversion embedding feature matrix is ​​input into the fully connected layer of the potential feature extraction and output layer to obtain a feature output vector, the feature output vector is input into the Softmax() function of the potential feature extraction and output layer to calculate the probability distribution of the feature output vector, and the probability distribution of the feature output vector is subjected to a Hadamard product operation with the feature output vector to obtain the position feature vector of user i and service j. :

[0182] ,

[0183] ,

[0184] ,

[0185] in, represents the transformed embedding feature matrix, represents the feature output vector, Z, M and E represent the weight matrix, bias term and stride of the convolutional layer, and represents the weight matrix of the fully connected layer, and represents the bias term of the fully connected layer, and the Softmax() function represents the probability distribution operation function;

[0186] Step 5: Input the user's feature vector, the service's feature vector, and the user and service's location feature vector into the feature fusion and prediction network of the pre-built prediction model, and output the predicted QoS value:

[0187] Among them, the feature fusion and prediction network is as follows Figure 5 As shown;

[0188] The feature vector of the user and the feature vector of the service are merged to obtain a merged feature vector, the merged feature vector is merged with the position feature vector of the user and the service to obtain a final merged feature vector, the final merged feature vector is input into the Γ fully connected layers of the feature fusion and prediction network, the interaction relationship between the features is learned, and the predicted QoS value is output in the last Γth fully connected layer , the specific expression is as follows:

[0189] ,

[0190] ,

[0191] ,

[0192] in, represents the merged eigenvector, represents the final merged eigenvector, and represents the weight and bias of the mth fully connected layer in the network, For the output of the mth layer of the feature fusion prediction network, please note that .

[0193] The training method of the pre-built prediction model includes:

[0194] The mean square error is used as the loss function, and the expression of the loss function is:

[0195] ,

[0196] Among them, q ij The QoS value returned by user i when calling service j, is the predicted QoS value of user i for service j output by the prediction model;

[0197] The parameters of the prediction model are updated by the stochastic gradient descent method so that the loss function of the prediction model is minimized. The specific expression is:

[0198] ,

[0199] ,

[0200] ,

[0201] in, are the parameters in the feature conversion inference network, including the weights and biases of the feature conversion inference network, Extract the parameters of the position feature network, including the weights and biases of the position feature extraction network. are the parameters in the feature fusion prediction network, including the weights and biases of the feature fusion prediction network. The learning rate is used to control the speed of gradient descent during the pre-set iteration.

[0202] To verify the performance of the proposed method, this paper conducts experiments based on two datasets, namely, response time and throughput datasets. Both datasets are derived from the WS-DREAM dataset, which is a large-scale QoS dataset covering real calls and is widely used in QoS prediction. WS-DREAM contains 1,974,675 historical QoS call records of 339 users and 5,825 services. In addition, WS-DREAM also provides identifiers of users and services as well as contextual information such as regions and autonomous systems. In actual service-oriented application scenarios, a single user only calls a limited number of services, which leads to the user-service QoS matrix having great sparsity. In order to simulate actual applications with very high sparsity, the experiment creates datasets with different matrix densities by randomly removing a specified number of QoS values ​​from WS-DREAM. Let μ be the matrix density:

[0203] ,

[0204] in is the number of selected items in the original user service QoS matrix used for training, =|Q|. Experiments are conducted at densities of 5%, 10%, 15%, and 20% (regardless of response time or throughput). At each density, the remaining data is randomly selected as the test set. For example, when the QoS matrix density is 5%, 5% of the data is selected as the training set, and the remaining 95% of the data is used as the test set. This embodiment does not use a validation set because it has little effect on the experimental results. This embodiment achieves accurate prediction results by taking the average value of multiple experiments.

[0205] About evaluation indicators: The experimental evaluation indicators selected mean absolute error (MAE) and root mean square error (RMSE), and the calculation method is as follows:

[0206] ,

[0207] ,

[0208] in and They represent the original QoS value and predicted value of user i calling service j, t and w represent the number of users and the number of services, respectively. MAE mainly evaluates the overall accuracy of QoS prediction by calculating the mean absolute error, while RMSE is a more sensitive evaluation of the performance of the QoS prediction model on outliers.

[0209] Specifically, in the attached Figure 6 -Attached Fig.15The MAE diagram of response time and the RMSE diagram of response time with different parameters as independent variables are shown in FIG.

[0210] Comparison method: The location-aware multi-feature fusion deep learning QoS prediction method applied in the embodiment is compared with ten typical comparison methods, including two memory-based collaborative filtering methods, one hybrid collaborative filtering method, three model-based collaborative filtering methods and four deep neural network methods, which are specifically described as follows:

[0211] 1) UPCC: This is a user-centric QoS prediction method that uses the Pearson Correlation Coefficient (PCC) to find a group of users similar to the target user as its domain, and then predicts the QoS value by combining the average QoS of the target user and the deviation adjustment of these similar users.

[0212] 2) IPCC: This is a service-centric QoS prediction method that uses the Pearson Correlation Coefficient (PCC) to find a set of services similar to the target service as its domain, and then predicts the QoS value by combining the average QoS of the target service and the deviation adjustment of these similar services.

[0213] 3) UIPCC: This is a hybrid collaborative filtering method that combines UPCC and IPCC. It uses weighted coefficients to adjust the importance of UPCC and IPCC and integrates similar users and services to predict QoS values.

[0214] 4) NMF: This method decomposes the QoS matrix into two non-negative matrices to learn the latent features in the data and then predict the QoS value.

[0215] 5) PMF: This method optimizes matrix decomposition by using a probabilistic model. It is a representative model-based method that predicts QoS values ​​by improving traditional matrix decomposition techniques.

[0216] 6) SPP+LLMF: This is a location-aware low-quality matrix factorization method that combines matrix factorization and location information to predict QoS values.

[0217] 7) NDMF: This is a domain-integrated deep matrix factorization method that captures the nonlinear interaction relationship between users and services through a neural network and selects user domains through user geographic location information and historical QoS records, and integrates user domains into the model to predict QoS values.

[0218] 8) LDCF: This method is a prediction method based on deep learning. This method predicts the QoS value by embedding the location information of users and services into feature representation and using multi-layer fully connected layers to learn the high-dimensional linear relationship between users and services.

[0219] 9) MM-DNN: This method is a deep learning-based prediction method that predicts QoS values ​​by fusing multi-level and multi-scale features and introducing individual evaluation into the deep learning model.

[0220] 10) DFMI: This method is a prediction method based on deep learning. It obtains the deep interaction features of users and services through feature mapping and reasoning, and predicts the QoS value by enhancing the data flow information through feature transformation.

[0221] In order to analyze the impact of different model parameter settings on the model prediction accuracy, this embodiment conducted the following experiments to determine the optimal values ​​of the model parameters when the matrix density is 5%, 10%, 15% and 20%.

[0222] 1) The dimension c of the embedding layer

[0223] The parameter c refers to the dimension of the embedding vector in the embedding layer. To verify the effect of c on the prediction accuracy of the model, the experiment sets N to 3, Γ to 3, K to 30, l to 50, and c to 2, 4, 8, 16, 32, and 64. The experimental results are shown in Figure 2. Figure 6 and Figure 7 As shown in the figure, as the dimension increases, both MAE and RMSE show a downward trend. When c<16, both MAE and RMSE show a downward trend. When the dimension c>16, MAE and RMSE have an upward trend. This is because when the dimension is too small, it may not be able to fully express the characteristic information of users and services, thereby affecting the feature expression ability and reducing the prediction accuracy. When the dimension is too large, the prediction accuracy may be reduced due to the fitting problem caused by data sparsity. Considering the above, this embodiment sets c to 16.

[0224] 2) User / service original feature vector dimension l

[0225] The parameter l refers to the number of individual features of users and services obtained after the original matrix Q is decomposed by non-negative matrix. To study the impact of l on the accuracy of QoS prediction, the experiment sets N to 3, Γ to 3, K to 30, c to 16, and l to 40, 50, 60, 70, 80, and 90. The experimental results are shown in Figure 8 and Fig. 9 As shown, when l<70, both MAE and RMSE show a downward trend. This is because the higher the dimension, the more features can be learned. However, when l>70, MAE and RMSE show an upward trend. This is mainly because as the dimension continues to increase, the model may overfit, resulting in a decrease in prediction accuracy. Therefore, in this embodiment, the value of l is 70.

[0226] 3) Number of similar users / services K

[0227] The parameter K refers to the number of similar users / services obtained after similarity calculation of the target user / service. To verify the influence of K on the prediction accuracy of the model, the experiment sets N to 3, Γ to 3, c to 16, l to 70, and K to 10, 15, 20, 25, 30, 35. The experimental results are shown in Figure 2. Fig.10 and Fig.11 As shown in the figure, when K<20, both MAE and RMSE show a downward trend, and when K>20, both MAE and RMSE show an upward trend. This is because as the number of similar users / services increases, the feature information provided by users or services with low similarity to the target user or service makes it difficult for the model to extract useful feature information, thus affecting the prediction accuracy. Therefore, in this embodiment, K is set to 20.

[0228] 4) Number of network layers N in feature conversion / feature inference blocks

[0229] To verify the effect of N on the prediction accuracy of the model, the experiment sets Γ to 3, K to 20, c to 16, l to 70, and N to 3, 4, 5, and 6. The experimental results are shown in Fig.12 and Fig.13 As shown in the figure, when N<4, as N increases, both MAE and RMSE show a downward trend, because a deeper network structure can better learn the intrinsic relationship between users and services. When N>4, both MAE and RMSE show an upward trend, because as the number of network layers continues to increase, the network may have an overfitting problem, resulting in a decrease in prediction accuracy. Therefore, in this embodiment, N is set to 4.

[0230] 5) QoS prediction network fully connected layers Γ

[0231] To verify the effect of Γ on the prediction accuracy of the model, the experiment sets K to 20, c to 16, l to 70, N to 4, and Γ to 2, 3, 4, and 5. The experimental results are shown in Fig.14 and Fig.15 As shown in FIG. 1 , when Γ<3, as Γ increases, the prediction accuracy improves. When Γ>3, the prediction accuracy decreases. This is because as the number of network layers increases, the network falls into overfitting, resulting in a decrease in prediction accuracy. Therefore, in this embodiment, Γ is set to 3.

[0232] In order to study the impact of the features extracted by the multi-layer feature extraction module and the features extracted by the position feature extraction network on the model, this section conducts an ablation experiment to compare three different experimental settings of the prediction model, where Response time represents response time and Throughput represents throughput.

[0233] Case 1: Only the user / service feature matrix extracted by the feature conversion module is used for QoS prediction;

[0234] Case 2: QoS prediction using features extracted by feature conversion module and multi-layer feature extraction module;

[0235] Case 3: Use the features extracted by the feature conversion module, multi-layer feature extraction module and location feature extraction network for QoS prediction.

[0236] Table 1 shows the MAE and RMSE of the three cases set in the experiment when the matrix density is 10%. It can be seen from Table 1 that the MAE and RMSE of response time and throughput gradually decrease from Case 1 to Case 3. Therefore, the features extracted by the fusion feature conversion module, the multi-layer feature extraction module and the position feature extraction network can achieve better experimental results.

[0237] Table 1 Ablation experiment

[0238]

[0239] Under four different matrix densities (5%, 10%, 15% and 20%), the location-aware multi-feature fusion deep learning QoS prediction method (LDMF) applied in this embodiment is compared with other methods. The experimental results are shown in Tables 2 and 3. Table 2 shows the MAE and RMSE of the response time. Under 5%, 15% and 20% matrix density, the MAE of the method LDMF disclosed in this embodiment is slightly higher than DFMI, but performs better in RMSE. Table 3 shows the MAE and RMSE of throughput. Under different matrix densities, the method LDMF disclosed in this embodiment outperforms other methods in terms of throughput prediction accuracy by 2.10% to 5.31%. The method LDMF disclosed in this embodiment is superior to other methods because the present application first pre-processes the data, performs matrix decomposition on the original QoS matrix to obtain the user and service potential feature matrix, and performs similarity calculation on the target user and service to obtain a similar user and service vector set; secondly, the vector set is input into the feature conversion inference network to obtain the high-dimensional features of the user and service, and obtain the feature vectors of the user and service; then, the user and service location context information is extracted through the location feature extraction network to obtain the location feature vectors of the user and service; finally, a QoS prediction network is constructed to fuse the above-obtained features to predict the QoS value.

[0240] Table 2 MAE and RMSE of response time

[0241]

[0242] Table 3. MAE and RMSE of throughput

[0243]

[0244] The experimental environment of this embodiment is Intel Core i7-11700KF CPU @3.60GHz, NVIDIA GeForceRTX3090GPU, Windows 11 64bit, and programming adopts Python 3.8 and Pytorch 1.8.0.

[0245] The experimental parameter settings of this embodiment are shown in Table 4.

[0246] Table 4 Experimental parameter settings

[0247]

[0248] Within the parameter range of Table 4, the optimal parameter settings of the proposed model are determined through real additional parameter experiments. Since many methods have differences in evaluation indicators and experimental environments, the experiments use the same evaluation indicators and data density.

[0249] Embodiment 2:

[0250] This embodiment provides a multi-feature fusion deep learning QoS prediction method and device, including:

[0251] The matrix construction module is used to construct the original QoS matrix Q by collecting the real QoS values ​​returned by the user calling the service;

[0252] A matrix processing module, used for initializing the original QoS matrix Q, obtaining a user potential feature matrix and a service potential feature matrix, inputting the user potential feature matrix and the service potential feature matrix into a similar feature set extraction network of a pre-built prediction model, and obtaining a user similar user potential feature vector set and a service similar service potential feature vector set;

[0253] A feature vector acquisition module, used for inputting the similar user potential feature vector set and the similar service potential feature vector set into a feature conversion inference network of a pre-built prediction model to obtain the user's feature vector and the service's feature vector;

[0254] A location feature vector acquisition module, used to obtain the location information of the user and the service through the real QoS value returned by the user calling the service, input the location information into the location feature extraction network of the pre-built prediction model, and extract the location feature vector of the user and the service;

[0255] The output module is used to input the user's feature vector, the service's feature vector and the user and service's location feature vector into a feature fusion and prediction network of a pre-built prediction model, and output a predicted QoS value.

[0256] A multi-feature fusion deep learning QoS prediction method device provided in an embodiment of the present invention can execute a multi-feature fusion deep learning QoS prediction method provided in an embodiment of the present invention, and has functional modules and beneficial effects corresponding to the execution method.

[0257] Embodiment three:

[0258] The present invention provides an electronic terminal, comprising a processor and a memory connected to the processor, wherein a computer program is stored in the memory, and when the computer program is executed by the processor, the steps of the method described in Example 1 are executed, and the terminal has functional modules and beneficial effects corresponding to the execution method.

[0259] Embodiment 4:

[0260] The embodiment of the present invention further provides a computer-readable storage medium on which a computer program is stored. When the program is executed by a processor, the steps of the method described in the first embodiment are implemented, and the program has functional modules and beneficial effects corresponding to the execution method.

[0261] Those skilled in the art will appreciate that the embodiments of the present application may be provided as methods, devices, or computer program products. Therefore, the present application may take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Moreover, the present application may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program codes.

[0262] The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (apparatus), and computer program products according to the embodiments of the present application. It should be understood that each process and / or block in the flowchart and / or block diagram, as well as the combination of processes and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0263] These computer program instructions may also be stored in a computer readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer readable memory produce an article of manufacture including an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1A function specified in one or more boxes.

[0264] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0265] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the technical principles of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims

1. A multi-feature fusion deep learning QoS prediction method, characterized in that: include: Construct the original QoS matrix Q by collecting the real QoS values ​​returned by users calling the service; Initialize the original QoS matrix Q to obtain a user potential feature matrix and a service potential feature matrix, input the user potential feature matrix and the service potential feature matrix into a similar feature set extraction network of a pre-built prediction model, and obtain a user similar user potential feature vector set and a service similar service potential feature vector set; Inputting the similar user potential feature vector set and the similar service potential feature vector set into a feature conversion inference network of a pre-built prediction model to obtain a feature vector of the user and a feature vector of the service; Obtaining the location information of the user and the service through the real QoS value returned by the user calling the service, inputting the location information into the location feature extraction network of the pre-built prediction model, and extracting the location feature vector of the user and the service; The user's feature vector, the service's feature vector, and the user and service's location feature vectors are input into a feature fusion and prediction network of a pre-built prediction model, and a predicted QoS value is output.

2. According to claim 1, a multi-feature fusion deep learning QoS prediction method is characterized in that: Initializing the original QoS matrix Q, obtaining the user potential feature matrix and the service potential feature matrix includes: The original QoS matrix Q∈ Decompose into user potential feature matrix and the service latent feature matrix ; Minimize the original QoS matrix Q and the user latent feature matrix by the multiplication update rule and the service latent feature matrix The difference between: ， ， ， ， Among them, t and w are the number of users and the number of services respectively. l are the dimensions of user latent feature vector and service latent feature vector, ||·|| F is the Frobenius norm, represents the kth feature of service j in the service potential feature matrix S, represents the kth feature of user i in the user potential feature matrix U, and T represents the transposed matrix.

3. According to claim 1, a multi-feature fusion deep learning QoS prediction method is characterized in that: Inputting the user potential feature matrix and the service potential feature matrix into the similar feature set extraction network of the pre-built prediction model to obtain the user's similar user potential feature vector set and the service's similar service potential feature vector set comprises: The similarity between two users or services is calculated by the Pearson coefficient. The calculation formula is as follows: ， represents the Pearson correlation coefficient between user i and user e, and are the QoS values ​​returned by user i and user e when they call service h, and Respectively represent the average QoS values ​​of the services called by user i and user e, represents the set of common services called by user i and user e, By calculating the Pearson correlation coefficient similarity between user i and other users, the top K users with the highest similarity to user i are selected, and the potential feature vectors of these users are combined into a set of similar user potential feature vectors for user i. ; By calculating the Pearson correlation coefficient similarity between service j and other services, the top K services with the highest similarity to service j are selected, and the potential feature vectors of these services are combined into a similar service potential feature vector set of service j. .

4. According to claim 3, a multi-feature fusion deep learning QoS prediction method is characterized in that: The similar user potential feature vector set Input the feature conversion inference network of the pre-built prediction model to obtain the feature vector V of user i i include: The feature conversion inference network includes a feature conversion module, a multi-layer feature extraction module and a feature inference module; The similar user potential feature vector set is transformed into Perform shape transformation operations into matrices , and the matrix Input the feature conversion module to obtain the feature matrix of user i , expressions include: ， ， ， in, is the output of the mth convolutional layer, Bm, Cm and Ym are the weight, bias and stride of the mth convolutional layer respectively, Conv() operation represents the convolution operation, Bn() operation represents the batch normalization operation, () represents the shape transformation operation, is the output of the Nth convolutional layer; The multi-layer feature extraction module further extracts features of different scales output by each convolution layer in the feature conversion module to obtain a merge matrix , expressions include: ， ， in Represents the output of the enhanced mth convolutional layer , , and Represent the weight matrix, bias term and stride of the mth convolutional layer respectively, Represents the merge matrix of the output features of the multi-layer feature extraction module; © represents the merge operation; The outputs of the feature conversion module and the multi-level feature extraction module are subjected to feature fusion by the feature inference module, and the vector after feature fusion is converted into a one-dimensional feature vector by a shape transformation operation. , and finally the one-dimensional feature vector Input feature inference module to extract features of user i and obtain feature vector V of user i i ; ， ， ， in, represents the output of the mth fully connected layer, and They respectively show the weight and bias of the mth fully connected layer, () represents a rectified linear unit.

5. The multi-feature fusion deep learning QoS prediction method according to claim 4 is characterized in that: Inputting the location information into a location feature extraction network of a pre-built prediction model, extracting location feature vectors of users and services includes: The location information is input into the location input layer of the location feature extraction network to generate initial features of user and service location information. The initial features of the user and service location information are expressed as: ， ， in, is the initial feature of the user's location information, , and denote the ID, RE and AS of user i respectively, is the initial feature of the service location information, , and denote the ID, RE and AS of service j respectively; One-hot encoding is used to convert the initial features of the user and service location information into a high-dimensional sparse vector matrix. The high-dimensional sparse vector matrix of the user and service location information is expressed as: ， ， in, is the high-dimensional sparse vector matrix of user location information, and They represent the corresponding high-dimensional sparse vectors obtained after one-hot encoding of user i’s ID, RE, and AS. is a high-dimensional sparse vector matrix of service location information, and They represent the corresponding high-dimensional sparse vectors obtained after one-hot encoding of the ID, RE and AS of service j; The high-dimensional sparse vector of the user and service location information is converted into a low-dimensional dense embedding feature vector through the location embedding layer of the location feature extraction network. The conversion process is expressed as: ， ， ， ， ， ， In the formula, , and represents the embedding weight matrix of user i, , and represents the embedding weight matrix of service j, D i , R i and A i are the embedded features of user i’s ID, RE and AS output in the embedding layer, D j , R j and A j are the embedded features output by the embedding layer of service j’s ID, RE, and AS respectively; Merge the embedded features of user i and service j to generate the corresponding embedded feature vector F i and F j , the corresponding embedding feature vector F i and F j The embedding feature matrix L is obtained by combining the merging and shape transformation operations. ij : ， ， ， The embedding feature matrix L is extracted through the convolution layer of the potential feature extraction and output layer in the position feature extraction network. ij The features of are transformed and a shape transformation operation is performed to obtain a conversion embedding feature matrix, the conversion embedding feature matrix is ​​input into the fully connected layer of the potential feature extraction and output layer to obtain a feature output vector, the feature output vector is input into the Softmax() function of the potential feature extraction and output layer to calculate the probability distribution of the feature output vector, and the probability distribution of the feature output vector is subjected to a Hadamard product operation with the feature output vector to obtain the position feature vector of user i and service j. : ， ， ， in, represents the transformed embedding feature matrix, represents the feature output vector, Z, M, and E represent the weight matrix, bias term, and stride of the convolutional layer, and represents the weight matrix of the fully connected layer, and It represents the bias term of the fully connected layer, and the Softmax() function represents the probability distribution operation function.

6. The multi-feature fusion deep learning QoS prediction method according to claim 5 is characterized in that: The user's feature vector, the service's feature vector, and the user and service's location feature vector are input into a feature fusion and prediction network of a pre-built prediction model, and the output prediction QoS value includes: The feature vector of the user and the feature vector of the service are merged to obtain a merged feature vector, the merged feature vector is merged with the position feature vector of the user and the service to obtain a final merged feature vector, the final merged feature vector is input into the Γ fully connected layers of the feature fusion and prediction network, the interaction relationship between the features is learned, and the predicted QoS value is output in the last Γth fully connected layer , the specific expression is as follows: ， ， ， in, represents the merged eigenvector, represents the final merged eigenvector, and represents the weight and bias of the mth fully connected layer in the network, It is the output of the mth layer of the feature fusion prediction network.

7. The multi-feature fusion deep learning QoS prediction method according to claim 2 is characterized in that: The training method of the prediction model includes: The mean square error is used as the loss function, and the expression of the loss function is: ， Among them, q ij The QoS value returned by user i when calling service j, is the predicted QoS value of user i for service j output by the prediction model; The parameters of the prediction model are updated by the stochastic gradient descent method so that the loss function of the prediction model is minimized. The specific expression is: ， ， ， in, are the parameters in the feature transformation inference network, Extract the parameters in the network for position features, Parameters in the feature fusion prediction network, The learning rate is used to control the speed of gradient descent during the pre-set iteration.

8. A multi-feature fusion deep learning QoS prediction device, characterized in that: include: The matrix construction module is used to construct the original QoS matrix Q by collecting the real QoS values ​​returned by the user calling the service; A matrix processing module, used for initializing the original QoS matrix Q, obtaining a user potential feature matrix and a service potential feature matrix, inputting the user potential feature matrix and the service potential feature matrix into a similar feature set extraction network of a pre-built prediction model, and obtaining a user similar user potential feature vector set and a service similar service potential feature vector set; A feature vector acquisition module, used for inputting the similar user potential feature vector set and the similar service potential feature vector set into a feature conversion inference network of a pre-built prediction model to obtain the user's feature vector and the service's feature vector; A location feature vector acquisition module, used to obtain the location information of the user and the service through the real QoS value returned by the user calling the service, input the location information into the location feature extraction network of the pre-built prediction model, and extract the location feature vector of the user and the service; The output module is used to input the user's feature vector, the service's feature vector and the user and service's location feature vector into a feature fusion and prediction network of a pre-built prediction model, and output a predicted QoS value.

9. An electronic terminal, characterized in that: comprising a processor and a memory connected to said processor, A computer program is stored in the memory, and when the computer program is executed by the processor, the steps of the method according to any one of claims 1 to 7 are performed.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 7 are implemented.

Images