Internet service screening method based on attention mechanism and time perception

Abstract

The invention discloses an Internet service screening method based on an attention mechanism and time perception. The method comprises the following steps: acquiring a corresponding service quality time sequence when a target user calls a plurality of alternative Internet services; inputting the service quality time sequence into a pre-trained Internet service quality prediction model to obtain a plurality of service quality prediction values corresponding to the alternative Internet services; and according to the service quality prediction value, selecting the Internet service with the optimal data throughput and response time when the target user calls the service. By extracting the user-service enhancement features of the change features of the user and the service in different time slices, the accuracy of predicting the Internet service quality value is improved, accurate selection of the current Internet service of the user is facilitated, the response time of the user in calling the data throughput of the Internet service is shortened, and the user experience is improved. The problem that the feature extraction capability is insufficient when the current model captures the QoS value is solved.

Description

Technical Field

[0001] The present invention relates to an Internet service screening method based on an attention mechanism and time perception, and belongs to the technical field of Internet service quality. Background Art

[0002] Quality of Service (QoS) is a commonly used metric for measuring the non-functional aspects of services (such as throughput and response time). Due to changes in network conditions, even when the same user invokes the same service, its QoS value may vary over time. Analyzing this phenomenon helps capture the time-varying characteristics of QoS values, thereby enhancing the ability to model the time characteristics of users and services, and further improving the accuracy of QoS prediction.

[0003] Traditional collaborative filtering methods have limitations in capturing the time-varying characteristics of QoS values. Deep learning can automatically extract high-dimensional features and mine hidden patterns, thus more effectively solving this problem. Long Short-Term Memory (LSTM) and Gated Recurrent Unit (GRU) are two common methods for processing sequential data. GRU is more suitable for building large-scale networks than LSTM because of its relatively simple structure and higher computational efficiency, but GRU still has certain deficiencies. For example, it cannot perform feature compensation, and its feature extraction ability is inferior to that of LSTM when dealing with long-time series data. Summary of the Invention

[0004] The purpose of the present invention is to overcome the deficiencies in the prior art and provide an Internet service screening method based on an attention mechanism and time perception. By extracting user-service enhanced features that change characteristics of users and services in different time slices, the accuracy of predicting Internet service quality values is improved, which is beneficial to the precise selection of the current Internet service for users, improves the response time of data throughput of users when invoking Internet services, and solves the problem of insufficient feature extraction ability of the current model when capturing QoS values.

[0005] To solve the above technical problems, the present invention is implemented by the following technical solutions:

[0006] An Internet service screening method based on an attention mechanism and time perception includes:

[0007] Obtain the service quality time series corresponding to a target user when invoking multiple alternative Internet services;

[0008] Input the service quality time series into a pre-trained Internet service quality prediction model to obtain multiple service quality prediction values corresponding to the alternative Internet services;

[0009] Select the Internet service with the best data throughput and response time when the target user calls the service according to the predicted service quality value;

[0010] Among them, the training of the pre-trained Internet service quality prediction model includes:

[0011] Obtain the user service quality sample time series from the user-service historical sample data set;

[0012] According to the user service quality sample time series, construct feature vector sets centered on users and services respectively;

[0013] Based on the feature vector sets, splice the feature vectors centered on users and the feature vectors centered on services to obtain user-service enhanced features;

[0014] Obtain and inherit the hidden state of the user-service enhanced features to obtain the time perception features of the user for the service;

[0015] Train the Internet service quality prediction model based on the time perception features and the target loss function to obtain the model parameters that minimize the loss value.

[0016] Furthermore, the Internet service quality prediction model includes:

[0017] A feature extraction module for constructing a feature vector set according to the input user service quality sample time series;

[0018] A feature enhancement neural network for obtaining enhanced features according to the input feature vector set;

[0019] A gated recurrent deep neural network for obtaining the time perception features of the user for the service according to the input enhanced features;

[0020] A prediction neural network for outputting the predicted service quality value of the corresponding Internet service according to the input time perception features.

[0021] Even further, the constructing of the feature vector sets centered on users and services respectively according to the user service quality sample time series includes:

[0022] Extract features from the user service quality sample time series to obtain the user feature vector of the user on the time slice and the service feature vector of the service on the time slice;

[0023] Based on the user feature vector and the service feature vector, determine the similarity between users and the similarity between services respectively;

[0024] Construct feature vector sets centered on users and services respectively according to the similarity between users and the similarity between services.

[0025] Further, determining the similarity between users and the similarity between services based on the user feature vector and the service feature vector respectively includes:

[0026] Calculating the similarity between users through the following formula:

[0027] ;

[0028] Where:

[0029] represents the feature vector of user at the th time slice, represents the feature vector of user at the th time slice;

[0030] represents the cosine similarity calculation function;

[0031] represents the norm of the vector;

[0032] Calculating the similarity between services through the following formula:

[0033] ;

[0034] Where:

[0035] represents the feature vector of service at the th time slice, represents the feature vector of service at the th time slice.

[0036] Further, constructing the feature vector sets centered on users and services according to the similarity between users and the similarity between services respectively includes:

[0037] Sorting all users in descending order of similarity to the target user, and selecting multiple users with top rankings to form a feature vector set centered on the target user;

[0038] Sorting all services in descending order of similarity to the target service, and selecting multiple services with top rankings to form a feature vector set centered on the target service.

[0039] Further, based on the feature vector set, performing splicing processing on the user-centered feature vector and the service-centered feature vector to obtain a user-service enhanced feature, including:

[0040] Obtain the query volume of the target user, the key vectors of other users, and the value vectors of other users in the user-centered feature vector set;

[0041] Based on the query volume of the target user, the key vectors of other users, and the value vectors of other users, use the multi-head attention mechanism to determine the first enhanced output set centered on the target user;

[0042] Perform the same operation as above on the service-centered feature vector set to determine the second enhanced output set centered on the target service;

[0043] Concatenate the first enhanced output set and the second enhanced output set to obtain the user-service enhanced feature.

[0044] Furthermore, the step of using the multi-head attention mechanism to determine the first enhanced output set centered on the target user based on the query volume of the target user, the key vectors of other users, and the value vectors of other users includes:

[0045] Calculate the output of the multi-head attention mechanism through the following formula:

[0046] ;

[0047] Where:

[0048] represents the output of the target user at the th time slice in the first convolutional layer; at the first convolutional layer;

[0049] represents the merge operation;

[0050] represents the th attention head, , represents the total number of attention heads;

[0051] represents the enhanced feature of the target user in the th attention head;

[0052] represents the output layer weight;

[0053] Perform a convolution operation on the output of the multi-head attention mechanism through the following formula:

[0054] ;

[0055] ;

[0056] Where:

[0057] Indicates the th time slice target user at the th convolutional layer output, Indicates the th time slice target user at the th convolutional layer output, Indicates the th time slice target user at the th convolutional layer output;

[0058] Indicates a convolution operation;

[0059] Indicates the th convolutional layer weights, Indicates the th convolutional layer bias term, Indicates the th convolutional layer stride, Indicates the th convolutional layer, Indicates the number of convolutional layers;

[0060] Indicates the enhanced feature, Indicates a batch normalization operation;

[0061] The concatenated feature is calculated using the following formula:

[0062] ;

[0063] Where:

[0064] Indicates the concatenated feature output after concatenation, Indicates a concatenation operation;

[0065] The output of the last convolutional layer is enhanced using the following formula to obtain the first enhanced output centered on the target user:

[0066] ;

[0067] Where:

[0068] Indicates a pooling operation, Indicates the output of the last layer L of the convolutional layer;

[0069] Indicates the The first enhanced output centered on the target user in a time slice is;

[0070] Then the first enhanced output set centered on the target user in a time slice is represented as:

[0071] ;

[0072] Wherein:

[0073] represents the first enhanced output set centered on the target user in a time slice is;

[0074] represents the first enhanced output centered on the target user in the th time slice, represents the number of the time slice and , ;

[0075] represents the first enhanced output centered on the target user in the th time slice;

[0076] represents the first enhanced output centered on the target user in the th time slice;

[0077] represents the first enhanced output centered on the target user in the th time slice.

[0078] Furthermore, obtaining and inheriting the hidden state of the user-service enhanced feature to obtain the time perception feature of the user for the service includes:

[0079] Using a gated recurrent deep neural network to perform time perception on the user-service enhanced feature to determine the time perception feature of the user for the service.

[0080] Furthermore, training the Internet service quality prediction model based on the time perception feature and the target loss function to obtain the model parameters that minimize the loss value includes:

[0081] Calculating the service quality prediction value of the target user calling a certain service through the following formula:

[0082] ;

[0083] Among them:

[0084] represents the output of the th fully connected layer;

[0085] represents the ReLU activation function; represents the bias term of the th fully connected layer, represents the weight of the th fully connected layer; represents the output of the th fully connected layer;

[0086] represents the number of fully connected layers of the prediction neural network, and the output of the last layer of the prediction neural network is the service quality prediction value;

[0087] When , , represents the time perception feature of the target user for the service ;

[0088] Update the model parameters based on the service quality prediction value and the target loss function.

[0089] Furthermore, the updating of the model parameters based on the service quality prediction value and the target loss function includes:

[0090] During the training of the Internet service quality prediction model, the following formula is used to calculate the loss value:

[0091] ;

[0092] Among them:

[0093] represents the function loss value;

[0094] represents the total number of users; represents the total number of Internet services;

[0095] represents at the time slice the actual service quality value of the target user calling the service ;

[0096] represents at the time slice the service quality prediction value of the target user calling the service ;

[0097] Update the model parameters through the gradient descent method, and update the model parameters using the following formula:

[0098] ;

[0099] ;

[0100] ;

[0101] where:

[0102] represents the learning rate that controls the speed of gradient descent during model training;

[0103] represents the partial derivative;

[0104] represents the first model parameter of the feature enhancement neural network;

[0105] represents the second model parameter in the gated recurrent depth neural network;

[0106] represents the third model parameter in the prediction neural network.

[0107] Compared with the prior art, the beneficial effects achieved by the present invention are as follows:

[0108] 1. Based on the quality-of-service time series corresponding to a target user when calling multiple alternative Internet services, the present invention extracts user-service enhancement features that characterize the changes of users and services in different time slices, obtains multiple quality-of-service prediction values corresponding to the alternative Internet services, and through the accurate quality-of-service prediction values, enables the target user to accurately select the current Internet service, improving the response time of the data throughput of the user when calling the Internet service.

[0109] 2. The present invention uses a multi-head attention mechanism to determine user-service enhancement features that characterize the changes of users and services in different time slices, and extracts feature vectors of users and services at different scales based on the query volume of the target user, the key vectors of other users, and the value vectors of other users, which helps to capture the interaction relationships of user and service features at different granularities, improve the feature expression ability, and thus improve the prediction accuracy.

[0110] 3. Based on the user-service enhancement features, the present invention uses a recurrent gated depth network to perform time perception on the user-service enhancement features, effectively captures and processes the dynamic changes and dependencies in the user-service enhancement features, obtains time-aware features, and finally obtains accurate quality-of-service prediction values based on the time-aware features. Brief Description of the Drawings

[0111] Figure 1 is a flowchart of an Internet service screening method based on attention mechanism and time perception provided in Embodiment 1 of the present invention;

[0112] Figure 2 is a training flowchart of an Internet service quality prediction model provided in Embodiment 1 of the present invention;

[0113] Figure 3 is a schematic structural diagram of a multi-head attention mechanism provided in Embodiment 1 of the present invention;

[0114] Figure 4 is a schematic structural diagram of a feature enhancement neural network provided in Embodiment 1 of the present invention;

[0115] Figure 5 is a schematic structural diagram of a gated recurrent deep neural network provided in Embodiment 1 of the present invention;

[0116] Figure 6 is a schematic diagram of the working principle of a recurrent gated deep unit provided in Embodiment 1 of the present invention. Detailed Description of the Embodiments

[0117] The present invention will be further described below with reference to the accompanying drawings. The following embodiments are only used to more clearly illustrate the technical solutions of the present invention and cannot be used to limit the protection scope of the present invention.

[0118] Embodiment 1:

[0119] As Figure 1 shown, the Internet service screening method based on attention mechanism and time perception includes:

[0120] Obtain the service quality time series corresponding to the target user when calling multiple alternative Internet services;

[0121] Input the service quality time series into a pre-trained Internet service quality prediction model to obtain multiple service quality prediction values corresponding to the alternative Internet services;

[0122] According to the service quality prediction values, select the Internet service with the best data throughput and response time when the target user calls the service; wherein, the Internet service quality prediction model includes:

[0123] A feature extraction module for constructing a feature vector set according to the input user service quality sample time series;

[0124] A feature enhancement neural network for obtaining enhanced features according to the input feature vector set;

[0125] A gated recurrent deep neural network is used to obtain the time perception features of users for services based on the input enhanced features;

[0126] A prediction neural network is used to output the service quality prediction value of the corresponding Internet service based on the input time perception features.

[0127] As Figure 2 shown, the training of the pre-trained Internet service quality prediction model includes:

[0128] Obtain the user service quality sample time series from the user-service historical sample data set;

[0129] According to the user service quality sample time series, construct a feature vector set centered on users and services respectively. Specifically:

[0130] Extract features from the user service quality sample time series to obtain the user feature vector of the user on the time slice and the service feature vector of the service on the time slice;

[0131] Based on the user feature vector and the service feature vector, determine the similarity between users and the similarity between services respectively, including:

[0132] For the user service quality sample time series , there is , represents the original matrix of service quality at the th time slice, represents the number of the time slice and , , , represents the matrix of , represents the total number of users, represents the total number of Internet services;

[0133] Perform non-negative matrix factorization on to obtain the user latent feature matrix at the th time slice, , represents the matrix of , represents the dimension of the user latent feature vector, is th column of , represents the feature vector of user at the th time slice. Similarly, represents the feature vector of user at the The eigenvector;

[0134] For Performing non - negative matrix factorization gives the service potential feature matrix on the th time slice , , Denote as the matrix of , and represents the dimension of the service potential eigenvector; ;

[0135] is the j - th column of Denote as the eigenvector of service on the th time slice. Similarly, denotes the eigenvector of service on the

[0136] Calculate the similarity between users through the following formula:

[0137] ;

[0138] Calculate the similarity between services through the following formula:

[0139] ;

[0140] denotes the cosine similarity calculation function; denotes the norm of the vector;

[0141] Construct a feature vector set centered on users and services respectively according to the similarity between users and the similarity between services, including:

[0142] Sort all users in descending order of similarity to the target user, and select multiple users with higher rankings to form a feature vector set centered on the target user;

[0143] Sort all services in descending order of similarity to the target service, and select multiple services with higher rankings to form a feature vector set centered on the target service.

[0144] Based on the feature vector set, splice the user - centered feature vector and the service - centered feature vector to obtain the user - service enhanced feature. Specifically:

[0145] As Figure 3 shown, obtain the query volume of the target user, the key vectors of other users, and the value vectors of other users in the user - centered feature vector set;

[0146] Based on the query volume of the target user, the key vectors of other users, and the value vectors of other users, a multi-head attention mechanism is used to determine a first enhanced output set centered on the target user, including:

[0147] Calculate the output of the multi-head attention mechanism through the following formula:

[0148] ;

[0149] Indicates a merging operation;

[0150] Indicates the output layer weight, Indicates the th time slice of the target user Output at the first convolutional layer; Indicates the th attention head, , Indicates the total number of attention heads;

[0151] Indicates the target user At the th attention head, the enhanced feature:

[0152] ;

[0153] Indicates the set of feature vectors centered on the target user ; Indicates other users At the th attention head, the value vector;

[0154] Indicates other users Value vector: , Indicates the weight of the value matrix;

[0155] Indicates other users At the th attention head, the importance to the target user ;

[0156] ;

[0157] Indicates the exponential function with the real number as the base;

[0158] Indicates the target user In the query vector of the th attention head, which represents the query vector of the target user:

[0159] ;

[0160] represents the weight of the query matrix;

[0161] T represents the transpose symbol, which represents the key vector of other users in the th attention head;

[0162] represents the key vector of other users :

[0163] ;

[0164] represents the weight of the key matrix; represents the key vector of other users in the th attention head;

[0165] represents the dimension of the key vector; represents the number of heads in the multi-head attention mechanism.

[0166] As Figure 4 shown, a convolution operation is performed on the output of the multi-head attention mechanism through the following formula:

[0167] ;

[0168] ;

[0169] represents the output of the target user in the th time slice in the th convolutional layer, represents the output of the target user in the th time slice in the th convolutional layer, represents the output of the target user in the th time slice in the th convolutional layer; in the th convolutional layer;

[0170] represents the convolution operation; represents the The weights of the convolutional layer, denote the bias term of the th convolutional layer, denote the stride of the th convolutional layer;

[0171] denote the enhanced features, denote the batch normalization operation;

[0172] The concatenated features are calculated using the following formula:

[0173] ;

[0174] where:

[0175] denote the concatenated features output after concatenation, denote the concatenation operation;

[0176] The output of the last convolutional layer is enhanced using the following formula to obtain the first enhanced output centered on the target user:

[0177] ;

[0178] denote the pooling operation, denote the output of the last layer L of the convolutional layer;

[0179] denote the th time slice of the first enhanced output centered on the target user ;

[0180] Then the first enhanced output sets centered on the target user on

[0181] time slices are represented as:

[0182] denote the first enhanced output sets centered on the target user on

[0183] denote the th time slice of the first enhanced output centered on the target user ;

[0184] denote the target user The first enhanced output as the center;

[0185] Indicates target user The first enhanced output as the center;

[0186] By Service The set of feature vectors centered on Through the same operation as above, we can get Targeted Service The second enhanced output is centered , thus obtaining Targeted Service The second enhanced output set centered on =[ ];

[0187] Indicates Targeted Service The second enhanced output as the center;

[0188] Indicates Targeted Service The second enhanced output as the center;

[0189] Indicates Targeted Service The second enhanced output as the center;

[0190] Based on the first enhanced output set and the second enhanced output set, the first enhanced output and the second enhanced output are concatenated using the following formula to obtain the user-service enhancement feature:

[0191] ;

[0192] Indicates Users on a time slice About Services User-service enhancement features;

[0193] but Users on a time slice About Services The user-service enhancement feature set is expressed as: ;

[0194] Indicates Users on a time slice User-service enhancement features for services ;

[0195] Indicates the user-service enhancement features of the user for the service on the th time slice User-service enhancement features for services ;

[0196] Indicates the user-service enhancement features of the user for the service on the th time slice User-service enhancement features for services ;

[0197] Obtain and inherit the hidden state of the user-service enhancement features to obtain the time perception features of the user for the service. Specifically:

[0198] Use a gated recurrent deep neural network to perform time perception on the user-service enhancement features to determine the time perception features of the user for the service, including:

[0199] As Figure 5 shown, the network structure in the figure includes F recursive gated deep units (RGDU);

[0200] The target user invokes the service at the time slice , , , which is the input for , indicating the user-service enhancement features of the user for the service in time slices, User-service enhancement features for services ; Indicates the number of time slices;

[0201] Indicates the th RGDU, Indicates the hidden state passed down from the previous network node. Combining and , the RGDU will obtain the hidden state passed to the next network node , Indicates the output of the (F-1)th network node. Taking the output of the Fth network node as the final result, the initial value of the hidden state ;

[0202] As Figure 6 shown, Indicates an addition operation; Denotes the Hadamard product of matrices, where the corresponding elements in two matrices of the same type are multiplied;

[0203] Fully connected layer denotes the fully connected layer;

[0204] Denotes the sigmoid function;

[0205] ReLU denotes the Rectified Linear Unit (ReLU);

[0206] GeLU denotes the activation function: ;

[0207] tanh denotes the hyperbolic tangent function;

[0208] x represents the input of the function.

[0209] The propagation process of RGDU is as follows:

[0210] Through the hidden state passed down from the previous network node (the -1-th network node) and the input of the current network node (the -th network node)

[0211] to obtain the reset gate and the update gate:

[0212] ;

[0213] ;

[0214] Denotes the intermediate variable of the reset gate;

[0215] Denotes the reset gate;

[0216] Denotes the hidden state weight matrix under the reset gate, Denotes the neighbor node weight matrix under the reset gate; Denotes the hidden state bias term under the reset gate, Denotes the neighbor node bias term under the reset gate;

[0217] Calculate the update gate through the following formula:

[0218] ;

[0219] ;

[0220] An intermediate variable representing the update gate

[0221] Represents the update gate;

[0222] Represents the hidden state weight matrix under the update gate, Represents the neighbor node weight matrix under the update gate, Represents the hidden state bias term under the update gate, Represents the neighbor node bias term under the update gate;

[0223] The hidden state of the current node information is calculated by the following formula:

[0224] ;

[0225] ;

[0226] ;

[0227] Represents the hidden state of the current node information;

[0228] Represents the intermediate variable of the current hidden state;

[0229] Represents the current hidden state;

[0230] Represents the neighbor node weight matrix under the gating component, Represents the hidden state weight matrix under the gating component, Represents the neighbor node bias term under the gating component, Represents the hidden state bias term under the gating component;

[0231] After time slices, the final result is output as the time-aware feature of the target user for the service . .

[0232] Train the Internet service quality prediction model based on the time-aware feature and the target loss function to obtain the model parameters that minimize the loss value. Specifically:

[0233] Calculate the service quality prediction value of the target user's call to a certain service by the following formula:

[0234] ;

[0235] Represents the output of the th fully connected layer;

[0236] represents the ReLU activation function; represents the bias term of the th fully connected layer, represents the weight of the th fully connected layer; represents the output of the th fully connected layer;

[0237] represents the number of fully connected layers of the prediction neural network, and the output of the last layer of the prediction neural network is the service quality prediction value;

[0238] When it is, , represents the time perception feature of the target user for the service ;

[0239] The output of the last layer of the prediction neural network is the service quality prediction value.

[0240] Update the model parameters based on the service quality prediction value and the target loss function, including:

[0241] During the process of training the Internet service quality prediction model, the loss value is calculated using the following formula:

[0242] ;

[0243] represents the function loss value;

[0244] represents the total number of users; represents the total number of Internet services;

[0245] represents the actual service quality value of the target user calling the service in the time slice ;

[0246] represents the predicted service quality value of the target user calling the service in the time slice ;

[0247] Update the model parameters by the gradient descent method, and update the model parameters using the following formula:

[0248] ;

[0249] ;

[0250] ;

[0251] represents the learning rate that controls the speed of gradient descent during model training;

[0252] represents the partial derivative;

[0253] represents the first model parameter of the feature enhancement neural network; represents the second model parameter in the gated recurrent depth neural network; represents the third model parameter in the prediction neural network; the initial values of all parameters in the network are obtained by generating random numbers.

[0254] To illustrate the superiority of the method proposed in the embodiments of the present invention, the following 1 to 5 traditional methods and 6 to 10 mainstream methods are selected for comparative experiments:

[0255] Method 1: UPCC (User-based CF using Pearson correlation coefficient, calculates the similarity between users using the Pearson correlation coefficient, and makes service recommendations based on similar users): This method is a user-based collaborative filtering algorithm that uses PCC to calculate user similarity and uses the relevant similarity data of other users to predict the missing QoS values;

[0256] Method 2: IPCC (Item-based CF using Pearson correlation coefficient, calculates the similarity between users using the Pearson correlation coefficient, and makes service recommendations based on similar services): This method is an item-based collaborative filtering algorithm that uses PCC to calculate service similarity and uses the relevant similarity data of other services to predict the missing QoS values;

[0257] Method 3: WSRec (Web service recommender system): This method is a hybrid collaborative algorithm that combines UPCC and IPCC to predict the missing QoS values;

[0258] Method 4: K-SLOPE (K-means clustering and Slope normal generation algorithm): This method uses K-means to exclude less similar users and simultaneously uses the Slope One algorithm to predict the missing QoS over time;

[0259] Method 5: NTF (Non - negative tensor factorization): This method is a generalized tensor decomposition model that takes into account the differences in QoS values at different times and replaces the user - service matrix in matrix factorization with a user - service - time interaction relationship;

[0260] Method 6: RTF (Recurrent tensor factorization): This method is a time - aware service prediction framework based on deep learning. It combines tensor decomposition and deep learning for time - aware service recommendation;

[0261] Method 7: TF - KMP (Temporal QoS forecasting with collaborative filtering QoS prediction based on K - means clustering);

[0262] Method 8: TUIPCC (Time - aware user - service item - based CF using Pearson correlation coefficient): This method is a collaborative - filtering - based method that divides QoS into multiple time slices and then uses a time - aware similarity calculation mechanism to select similar users (or services) to predict missing QoS;

[0263] Method 9: DeepTSQP (Deep learning - based approach for temporally aware service QoS prediction): It is a deep - learning - based method that uses feature representation and integration to perform time - aware QoS prediction for services;

[0264] Method 10: DeepTFO (Deep Learning Model to Predict Quality if Service Based on Time-aware Feature Optimization): This method is a feature compensation method based on a generative adversarial network. It extracts the time feature information of users and services through probabilistic matrix factorization and constructs a gated feature extraction network to compensate for feature losses to predict QoS values.

[0265] To compare the prediction accuracy of service quality prediction methods objectively, the commonly used evaluation metrics currently are Mean Absolute Error (MAE) and Root Mean Squared Error (RMSE):

[0266] ;

[0267] ;

[0268] represents the total number of users, represents the total number of Internet services;

[0269] represents user at the th time slice when calling service 's actual Quality of Service (QoS) value;

[0270] represents user at the th time slice when calling service 's predicted Quality of Service (QoS) value;

[0271] The mean absolute error is the average of the absolute values of the differences between the predicted values and the true values, representing the average deviation between the predicted values and the true values; the smaller the MAE value, the higher the accuracy of the prediction model;

[0272] The root mean square error is the square root of the average of the sum of the squares of the differences between the predicted values and the true values, representing the average error between the predicted values and the true values; the smaller the RMSE value, the higher the precision of the prediction model.

[0273] The dataset used in this comparative experiment is the real QoS dataset WS-DREAM2, which contains the records of 142 users' QoS calls to 4,500 Web services within 64 different time slices (each time slice is 15 minutes apart), including throughput and response time. Since this dataset is widely used in QoS time series prediction research, it has high representativeness and reference value;

[0274] In the comparative experiment, the dimension of the original user / service feature vector is set to 60, the number of similar users / services is set to 30, the number of attention module heads is set to 16, the number of convolutional layers is set to 3, the number of fully connected network layers is set to 3, and the matrix densities (Density) are 5%, 10%, 15%, and 20% respectively;

[0275] Table 1 Prediction accuracy of response time

[0276]

[0277] As shown in Table 1, the MAE and RMSE of the method proposed in this embodiment in terms of response time are better than those of the other 10 methods, and the gains are between 1.83% and 4.06%.

[0278] Table 2 Prediction accuracy of throughput

[0279]

[0280] As shown in Table 2, the MAE and RMSE of the method proposed in this embodiment in terms of throughput are better than those of the other methods, and the gains are from 2.48% to 10.01%.

[0281] In summary, the method proposed in this embodiment has the highest prediction accuracy for the two QoS attributes (response time and throughput), and can achieve the best results on matrices with different densities. As the matrix density increases, the prediction accuracy also improves accordingly because the high-density QoS matrix provides more data for the network, enabling the model to make more accurate predictions.

[0282] In the description of the present invention, it should be understood that terms such as "first", "second", etc. are only used for descriptive purposes and cannot be understood as indicating or implying relative importance or implicitly specifying the quantity of the indicated technical features. Thus, features defined with "first", "second", etc. may explicitly or implicitly include one or more of such features.

[0283] Those skilled in the art will appreciate that the embodiments of the present application may be provided as methods, systems, 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.

[0284] The present application is described with reference to the flowcharts of the methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each process in the flowchart 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 process in the flowchart. Figure 1 A process or multiple processes or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0285] 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 comprising an instruction device, which implements the process Figure 1 A function specified in a process or multiple processes.

[0286] 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 The steps of a specified function in a process or multiple processes.

[0287] The embodiments of the present invention are described above in conjunction with the accompanying drawings, but the present invention is not limited to the above-mentioned specific implementation methods. The above-mentioned specific implementation methods are merely illustrative and not restrictive. Under the enlightenment of the present invention, ordinary technicians in this field can also make many forms without departing from the scope of protection of the purpose of the present invention and the claims, which all fall within the protection of the present invention.

Claims

1. An Internet service screening method based on attention mechanism and time perception, characterized in that: include: Obtaining a time series of service quality corresponding to a target user when calling multiple alternative Internet services; Inputting the service quality time series into a pre-trained Internet service quality prediction model to obtain a plurality of service quality prediction values ​​corresponding to the candidate Internet services; According to the service quality prediction value, select the Internet service with the best data throughput and response time when the target user calls the service; The training of the pre-trained Internet service quality prediction model includes: Obtain user service quality sample time series from the user-service history sample dataset; According to the user service quality sample time series, construct the user-centered and service-centered feature vector sets respectively; Based on the feature vector set, the user-centric feature vector and the service-centric feature vector are concatenated to obtain the user-service enhanced feature; Obtain and inherit the hidden state of the user-service enhancement feature to obtain the user's time perception feature of the service; The Internet service quality prediction model is trained based on time-aware features and target loss function to obtain model parameters that minimize the loss value.

2. The Internet service screening method based on attention mechanism and time perception according to claim 1 is characterized in that: The Internet service quality prediction model includes: A feature extraction module is used to construct a feature vector set based on the input user service quality sample time series; Feature enhancement neural network, used to obtain enhanced features based on the input feature vector set; A gated recurrent deep neural network is used to obtain the user's time perception characteristics of the service based on the input enhanced features; The prediction neural network is used to output the service quality prediction value of the corresponding Internet service based on the input time perception characteristics.

3. The Internet service screening method based on attention mechanism and time perception according to claim 1 is characterized in that: The method constructs a user-centric and service-centric feature vector set based on the user service quality sample time series, including: Perform feature extraction on the user service quality sample time series to obtain the user feature vector of the user on the time slice and the service feature vector of the service on the time slice; Based on the user feature vector and the service feature vector, determining the similarity between users and the similarity between services respectively; According to the similarity between users and the similarity between services, feature vector sets centered on users and services are constructed respectively.

4. The method for screening Internet services based on attention mechanism and time perception according to claim 3 is characterized in that: The determining of the similarity between users and the similarity between services based on the user feature vector and the service feature vector respectively includes: The similarity between users is calculated by the following formula: ； in: Indicates Users on a time slice The characteristic vector of Indicates Users on a time slice The eigenvector of Represents the cosine similarity calculation function; represents the magnitude of a vector; The similarity between services is calculated by the following formula: ； in: Indicated in Time-slice service The characteristic vector of Indicated in Time-slice service The feature vector of .

5. The method for screening Internet services based on attention mechanism and time perception according to claim 3 is characterized in that: The method of constructing a user-centered and service-centered feature vector set based on the similarity between users and the similarity between services respectively includes: Sort all users by their similarity to the target user from high to low, and select multiple users with the highest ranking to form a feature vector set centered on the target user; All services are sorted from high to low in terms of similarity with the target service, and multiple services with the highest ranking are selected to form a feature vector set centered on the target service.

6. The method for screening Internet services based on attention mechanism and time perception according to claim 1, characterized in that: Based on the feature vector set, the user-centric feature vector and the service-centric feature vector are spliced ​​to obtain the user-service enhancement feature, including: Get the query volume of the target user in the user-centric feature vector set, the key vectors of other users, and the value vectors of other users; Based on the query volume of the target user, the key vectors of other users, and the value vectors of other users, a multi-head attention mechanism is used to determine the first enhanced output set centered on the target user; Applying the same operation as above to the service-centric feature vector set, determining a second enhanced output set centric to the target service; The first enhanced output set and the second enhanced output set are concatenated to obtain user-service enhancement features.

7. The method for screening Internet services based on attention mechanism and time perception according to claim 6, characterized in that: The method uses a multi-head attention mechanism to determine a first enhanced output set centered on the target user based on the query volume of the target user, the key vectors of other users, and the value vectors of other users, including: The output of the multi-head attention mechanism is calculated by the following formula: ； in: Indicates Time Slice Target User At the output of the first convolutional layer; Indicates a merge operation; Indicates Attention head, , represents the total number of attention heads; Indicates the target user In the Enhanced features in the attention heads; represents the output layer weight; The output of the multi-head attention mechanism is convolved as follows: ； ； in: Indicates Time Slice Target User In the The output of the convolutional layer, Indicates Time Slice Target User In the The output of the convolutional layer, Indicates Time Slice Target User In the Convolutional layer output; Represents the convolution operation; Indicates The weights of the convolutional layers, Indicates The bias term of the convolutional layer, Indicates The stride of the convolutional layer, Indicates convolutional layers, Indicates the number of convolutional layers; Indicates The enhanced features, Represents batch normalization operation; The following formula is used to calculate the splicing features: ； in: Express The splicing features output after splicing, Represents a splicing operation; The output of the last convolutional layer is enhanced by the following formula to obtain the first enhanced output centered on the target user: ； in: represents the pooling operation, Represents the output of the last L layers of the convolutional layer; Indicates target user The first enhanced output as the center; but target user The first enhanced output set centered on is expressed as: ； in: express target user The first enhanced output set centered on; Indicates target user The first enhanced output is centered, represents the number of the time slice and , ; Indicates target user The first enhanced output as the center; Indicates target user The first enhanced output as the center; Indicates target user The first enhanced output is centered.

8. The method for screening Internet services based on attention mechanism and time perception according to claim 2, characterized in that: The step of acquiring and inheriting the hidden state of the user-service enhancement feature to obtain the user's time perception feature of the service includes: A gated recursive deep neural network is used to perform time perception on the user-service enhancement features and determine the time perception characteristics of users for services.

9. The method for screening Internet services based on attention mechanism and time perception according to claim 8, characterized in that: The Internet service quality prediction model is trained based on the time perception feature and the target loss function to obtain the model parameters that minimize the loss value, including: The service quality prediction value of a target user calling a service is calculated by the following formula: ； in: Indicates The output of the fully connected layer; ReLU activation function. Indicates The bias term of the fully connected layer, Indicates The weights of the fully connected layers; Indicates The output of the fully connected layer; Indicates the number of fully connected layers of the prediction neural network, and the last layer of the prediction neural network outputs the service quality prediction value; when =1, , Indicates the target user About Services The time perception characteristics of Update the model parameters based on the service quality prediction value and the target loss function.

10. The method for screening Internet services based on attention mechanism and time perception according to claim 9, characterized in that: The updating of the model parameters based on the service quality prediction value and the target loss function includes: The following formula is used to calculate the loss value during the training of the Internet service quality prediction model: ； in: Represents the function loss value; Indicates the total number of users; represents the total number of Internet services; Indicates in time slice Target Users Calling a service The actual value of service quality; Indicates in time slice Target Users Calling a service The predicted value of service quality; The model parameters are updated by the gradient descent method, and the following formula is used to update the model parameters: ； ； ； in: Indicates the learning rate that controls the speed of gradient descent during model training; represents partial derivative; represents a first model parameter of a feature enhancement neural network; represents a second model parameter in a gated recurrent deep neural network; Represents the third model parameter in the prediction neural network.

Images