A quality of service prediction method and device for adaptive feature learning

Abstract

The application discloses a quality of service prediction method and device based on self-adaptive feature learning, and the method comprises the following steps: acquiring the context information of a user and a service, and an interaction matrix of the user and the service, so as to obtain input features of a model; extracting multi-source features of the user and the service based on the input features; encoding the multi-source features to obtain interaction features of the user and the service; inputting the interaction features into a deep neural network to obtain a predicted quality of service value; calculating a contrast loss; modifying each parameter of the deep neural network, repeating the above steps until the contrast loss no longer decreases, saving the parameters of the deep neural network to obtain a trained model; and inputting real-time acquired input features into the trained model to obtain a prediction result of the quality of service; the application has the advantage that the quality of service prediction is relatively accurate.

Description

Technical Field

[0001] This invention relates to the field of service computing, and more specifically to a service quality prediction method and apparatus based on adaptive feature learning. Background Technology

[0002] With the rapid development of technologies such as cloud computing and the Internet of Things, the demand for software systems in social production and daily life is increasing daily. Service-oriented architecture needs to strike a balance between stable service integration and flexible adaptation to needs. Based on this, microservice technology, with its independent processes and deployment capabilities, has emerged. It boasts advantages such as distributed storage, high availability, scalability, and intelligent operation and maintenance, and continues to evolve to meet the diverse and personalized needs of users. This has led to the development and release of more and more services with the same or similar functions. In this context, recommending ideal services to users has become a significant challenge. Service quality refers to some non-functional attributes of web services, which can intuitively reflect the degree of matching between users and services. Therefore, service quality has become an important criterion for users when choosing from many similar services. However, in real-world service call scenarios, it is impractical for users to conduct large-scale evaluation and testing of service quality. On the one hand, the quality of services called by users is easily affected by multifaceted environments. For example, when users call the same service, changes in user geographical location and network environment may lead to different QoS values ​​for the called services. On the other hand, evaluating the service quality of all services is a time-consuming and resource-intensive task. Against this backdrop, accurate and personalized service quality prediction has become a significant challenge in the field of service computing.

[0003] Chinese Patent Publication No. CN109873726A discloses a robust quality of service (QoS) prediction and assurance method based on deep learning in SDN. This method preprocesses historical traffic data from switch ports in the network topology, establishes a deep learning model, and trains the model using the preprocessed traffic data. It then collects real-time traffic data from switch ports in the network topology via the SDN controller, preprocesses it, and feeds it into the trained deep learning model to output QoS parameters for the current video stream service. Based on the QoS parameters of the previous video stream service, it quantifies and calculates the QoE index. Using all the QoS parameters of the current video stream service, it employs machine learning methods to predict, record, and monitor the network traffic of the current user's video stream service. The QoE index prompts the client or the SDN central controller to take corresponding actions, thus achieving the traffic prediction function. However, this patent application does not consider the relationship between users and services when performing QoS prediction. Changes in the user's geographical location may lead to different QoS values ​​for the called services, resulting in inaccurate QoS predictions. Summary of the Invention

[0004] The technical problem to be solved by this invention is that the existing service quality prediction methods do not consider the impact of different contextual features on service quality, resulting in inaccurate service quality prediction.

[0005] This invention solves the above-mentioned technical problems through the following technical means: an adaptive feature learning service quality prediction method, the method comprising:

[0006] Step 1: Obtain the contextual information of users and services, and the interaction matrix of users and services to obtain the input features of the model;

[0007] Step 2: Extract multi-source features for users and services based on input features;

[0008] Step 3: Encode the multi-source features to obtain the interaction features of users and services;

[0009] Step 4: Input the interaction features into the deep neural network to obtain the predicted quality of service value;

[0010] Step 5: Calculate the contrast loss;

[0011] Step 6: Modify the parameters of the deep neural network, repeat steps 2 to 5 until the contrastive loss no longer decreases, and save the parameters of the deep neural network to obtain the trained model;

[0012] Step 7: Input the real-time acquired input features into the trained model to obtain the service quality prediction results.

[0013] This invention obtains the input features of the model based on the contextual information of users and services, and the interaction matrix between users and services, trains the model, and uses the trained model to predict service quality. It fully considers the impact of different contextual information on service quality, assigns different weights to different contextual information, and the service quality prediction is relatively accurate.

[0014] Further, step one includes:

[0015] The input features of the model are represented as

[0016] Xe=(U-ID,U-AS,U-Region,S-ID,S-AS,S-Region,S-WSDL,S-ServiceProvider)

[0017] Wherein, U-ID and S-ID represent the IDs of the user and the service, respectively; U-AS and S-AS represent the Autonomous System Numbers of the user and the service, respectively; U-Region and S-Region represent the regions where the user and the service are located, respectively; S-WSDL represents the WSDL address of the service; and S-ServiceProvider represents the service provider.

[0018] Furthermore, step two includes:

[0019] Step 201: Perform one-hot encoding on the input features Xe of the model to obtain...

[0020] Xe=(0100,010,100,1000,100,001,0010,0100)

[0021] Step 202: Map each one-hot encoding to a dense vector of fixed length d. For each domain's text information, obtain an embedding matrix. To represent all textual features in this domain;

[0022] Step 203: Extract the embedding representation of the vectors with a value of 1 in each domain, and remove the embedding representation of the vectors with a value of 0 in each domain to obtain the multi-source feature matrix of the service call. Where e i Let f represent the i-th row vector of E, and let f represent the number of files containing all text information.

[0023] Furthermore, step three includes:

[0024] Step 301: Encode the multi-source feature matrix E using an encoder:

[0025]

[0026] a ij The weight representing the relationship between context feature i and context feature j is calculated as follows:

[0027] a′ ij =relu(W(e i ⊙e j )+b)

[0028]

[0029] Where ReLU represents the activation function, W represents the weight matrix, b represents the bias, and the ⊙ symbol indicates that corresponding positions of two vectors are multiplied, and the result is still a vector rather than a number.

[0030] Step 302: For each row vector p of matrix Pi Normalization operation yields:

[0031]

[0032] Where, p i Let represent the i-th row vector in matrix P, where Represents each row vector p i The mean of ε is a minimum value that prevents the denominator of the equation from being zero, while The variance of the row vectors is represented by γ, and β are the parameters that the model can be trained on.

[0033] Furthermore, step four includes:

[0034] The normalized matrix P is input into the deep neural network to calculate the predicted service quality value.

[0035] Q = flatten(P)

[0036] y1 = relu(W1Q + b1),

[0037] y2=relu(W2y1+b2),…,

[0038] y L =relu(W L y L-1 +b L )

[0039] Here, `flatten()` represents the operation of converting a two-dimensional matrix into a one-dimensional vector, y L The output of the Lth layer of the neural network, where ReLU represents the activation function, W L Let b represent the weight matrix of the Lth layer. L This represents the deviation of the Lth layer.

[0040] Furthermore, step five includes:

[0041] Through formula

[0042]

[0043] Calculate the predicted value The comparison loss with the true value R, where M represents the number of all users, N represents the number of all services, and R is the loss between the true value and the true value. ij This represents the actual QoS value. This represents the QoS value predicted by this model.

[0044] The present invention also provides a service quality prediction device based on adaptive feature learning, the device comprising:

[0045] The input feature acquisition module is used to obtain the context information of users and services, the interaction matrix of users and services, and to obtain the input features of the model.

[0046] The multi-source feature acquisition module is used to extract multi-source features of users and services based on input features.

[0047] The interaction feature acquisition module is used to encode multi-source features to obtain the interaction features of users and services;

[0048] The input module is used to input interactive features into a deep neural network to obtain a predicted quality of service value.

[0049] The loss calculation module is used to calculate the comparative loss;

[0050] The training module is used to modify the parameters of the deep neural network, repeatedly execute the multi-source feature acquisition module to the loss calculation module until the contrast loss no longer decreases, and save the parameters of the deep neural network to obtain the trained model.

[0051] The service quality prediction module is used to input real-time acquired input features into the trained model to obtain the service quality prediction results.

[0052] Furthermore, the input feature acquisition module is also used for:

[0053] The input features of the model are represented as

[0054] Xe=(U-ID,U-AS,U-Region,S-ID,S-AS,S-Region,S-WSDL,S-ServiceProvider)

[0055] Wherein, U-ID and S-ID represent the IDs of the user and the service, respectively; U-AS and S-AS represent the Autonomous System Numbers of the user and the service, respectively; U-Region and S-Region represent the regions where the user and the service are located, respectively; S-WSDL represents the WSDL address of the service; and S-ServiceProvider represents the service provider.

[0056] Furthermore, the multi-source feature acquisition module is also used for:

[0057] Step 201: Perform one-hot encoding on the input features Xe of the model to obtain...

[0058] Xe=(0100,010,100,1000,100,001,0010,0100)

[0059] Step 202: Map each one-hot encoding to a dense vector of fixed length d. For each domain's text information, obtain an embedding matrix. To represent all textual features in this domain;

[0060] Step 203: Extract the embedding representation of the vectors with a value of 1 in each domain, and remove the embedding representation of the vectors with a value of 0 in each domain to obtain the multi-source feature matrix of the service call. Where f represents the number of files containing all text information.

[0061] Furthermore, the interaction feature acquisition module is also used for:

[0062] Step 301: Encode the multi-source feature matrix E using an encoder:

[0063]

[0064] a ij The weight representing the relationship between context feature i and context feature j is calculated as follows:

[0065] a′ ij =relu(W(w) i ⊙e j )+b)

[0066]

[0067] Where ReLU represents the activation function, W represents the weight matrix, b represents the bias, and the ⊙ symbol indicates that corresponding positions of two vectors are multiplied, and the result is still a vector rather than a number.

[0068] Step 302: For each row vector p of matrix P i Normalization operation yields:

[0069]

[0070] Where, p i Let represent the i-th row vector in matrix P, where Represents each row vector p i The mean of ε is a minimum value that prevents the denominator of the equation from being zero, while The variance of the row vectors is represented by γ, and β are the parameters that the model can be trained on.

[0071] Furthermore, the input module is also used for:

[0072] The normalized matrix P is input into the deep neural network to calculate the predicted service quality value.

[0073] Q = flatten(P)

[0074] y1 = relu(W1Q + b1),

[0075] y2=relu(W2y1+b2),…,

[0076] y L =relu(W L y L-1 +b L )

[0077] Here, `flatten()` represents the operation of converting a two-dimensional matrix into a one-dimensional vector, y L The output of the Lth layer of the neural network, where ReLU represents the activation function, W L Let b represent the weight matrix of the Lth layer. L This represents the deviation of the Lth layer.

[0078] Furthermore, the loss calculation module is also used for:

[0079] Through formula

[0080]

[0081] Calculate the predicted value The comparison loss with the true value R, where M represents the number of all users, N represents the number of all services, and R is the loss between the true value and the true value. ij This represents the actual QoS value. This represents the QoS value predicted by this model.

[0082] The advantages of this invention are as follows: This invention obtains the input features of the model based on the contextual information of users and services and the interaction matrix between users and services, trains the model, and uses the trained model to predict service quality. It fully considers the contextual information and the interaction between users and services, thus taking into account the relationship between users and services, and the service quality prediction is relatively accurate. Attached Figure Description

[0083] Figure 1 This is a flowchart of a service quality prediction method based on adaptive feature learning disclosed in Embodiment 1 of the present invention. Detailed Implementation

[0084] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below in conjunction with the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0085] Example 1

[0086] Before service consumers choose a service, the service market typically needs to predict the quality of services they haven't used before and recommend suitable services based on its own experience. Considering that different service consumers may have significantly different perceptions of the service quality of the same service, this invention proposes an adaptive feature learning-based service quality prediction method. This method utilizes contextual information about users and services to assist in service quality prediction, while also considering the importance of different contextual information for service quality prediction, assigning different weights to different contextual information, and using this to predict the QoS of services that consumers have not used before. Figure 1 As shown, the specific method and process include:

[0087] Step 1: Obtain the contextual information of users and services, and the interaction matrix between users and services, to obtain the input features of the model; the specific process is as follows:

[0088] The input features of the model are represented as

[0089] Xe=(U-ID,U-AS,U-Region,S-ID,S-AS,S-Region,S-WSDL,S-SetviceProvider)

[0090] Wherein, U-ID and S-ID represent the IDs of the user and the service, respectively; U-AS and S-AS represent the Autonomous System Numbers of the user and the service, respectively; U-Region and S-Region represent the regions where the user and the service are located, respectively; S-WSDL represents the WSDL address of the service; and S-ServiceProvider represents the service provider.

[0091] In practical applications, when a user chooses a payment service from four options—WeChat Pay, Alipay, Amazon Pay, and Apple Pay—the service marketplace first collects contextual information about the user (including their geographical location, IP address, etc.) and information about the four services. This information is then input into the model. The model predicts the service quality score for the user and the four services. Based on these service quality scores, the service marketplace recommends the most suitable service to the user.

[0092] The user's context information includes: the user's ID (U-ID), the user's Autonomous System Number (U-AS), and the user's region (U-Region). The context information for the four services includes: the service's ID (S-ID), the service's Autonomous System Number (S-Region), the service's WSDL address (S-WSDL), and the service provider (S-ServiceProvider). These pieces of information are concatenated to form the input features.

[0093] Xe=(U-ID,U-AS,U-Region,S-ID,S-AS,S-Region,S-WSDL,S-ServiceProvider).

[0094] Step 2: Extract multi-source features for users and services based on input features; the specific process is as follows:

[0095] Step 201: Perform one-hot encoding on the input features Xe of the model to obtain...

[0096] Xe=(0100,010,100,1000,100,001,0010,0100)

[0097] Step 202: Map each one-hot encoding to a dense vector of fixed length d. For each domain's text information, obtain an embedding matrix. To represent all textual features in this domain;

[0098] Step 203: Extract the embedding representation of the vectors with a value of 1 in each domain, and remove the embedding representation of the vectors with a value of 0 in each domain to obtain the multi-source feature matrix of the service call. Where f represents the number of files containing all text information.

[0099] Step 3: Encode the multi-source features to obtain the interaction features between users and services; the specific process is as follows:

[0100] Step 301: Encode the multi-source feature matrix E using an encoder:

[0101]

[0102] a ij The weight representing the relationship between context feature i and context feature j is calculated as follows:

[0103] a′ ij =relu(W(e i ⊙ej )+b)

[0104]

[0105] Where ReLU represents the activation function, W represents the weight matrix, b represents the bias, and the ⊙ symbol indicates that corresponding positions of two vectors are multiplied, and the result is still a vector rather than a number.

[0106] Step 302: For each row vector p of matrix P i Normalization operation yields:

[0107]

[0108] Where, p i Let represent the i-th row vector in matrix P, where Represents each row vector p i The mean of ε is a minimum value that prevents the denominator of the equation from being zero, while The variance of the row vectors is represented by γ, and β are the parameters that the model can be trained on.

[0109] Step 4: Input the interaction features into the deep neural network to obtain the predicted quality of service value; the specific process is as follows:

[0110] The normalized matrix P is input into the deep neural network to calculate the predicted service quality value.

[0111] Q = flatten(P)

[0112] y1 = relu(W1Q + b1),

[0113] y2=relu(W2y1+b2),…,

[0114] y L =relu(W L y L-1 +b L )

[0115] Here, `flatten()` represents the operation of converting a two-dimensional matrix into a one-dimensional vector, y L The output of the Lth layer of the neural network, where ReLU represents the activation function, W L Let b represent the weight matrix of the Lth layer. L This represents the deviation of the Lth layer.

[0116] Step 5: Calculate the contrast loss; the specific process is as follows:

[0117] Through formula

[0118]

[0119] Calculate the predicted value The comparison loss with the true value R, where M represents the number of all users, N represents the number of all services, and R is the loss between the true value and the true value. ij This represents the actual QoS value. This represents the QoS value predicted by this model.

[0120] Step 6: Modify the parameters of the deep neural network, repeat steps 2 to 5 until the contrastive loss no longer decreases, and save the parameters of the deep neural network to obtain the trained model;

[0121] Step 7: Input the real-time acquired input features into the trained model to obtain the service quality prediction results.

[0122] Through the above technical solutions, this invention obtains the input features of the model based on the contextual information of users and services and the interaction matrix between users and services, trains the model, and uses the trained model to predict service quality. It fully considers the contextual information and the interaction between users and services, thus taking into account the relationship between users and services, and the service quality prediction is relatively accurate.

[0123] Example 2

[0124] Based on Embodiment 1, Embodiment 2 of the present invention also provides a service quality prediction device with adaptive feature learning, the device comprising:

[0125] The input feature acquisition module is used to obtain the context information of users and services, the interaction matrix of users and services, and to obtain the input features of the model.

[0126] The multi-source feature acquisition module is used to extract multi-source features of users and services based on input features.

[0127] The interaction feature acquisition module is used to encode multi-source features to obtain the interaction features of users and services;

[0128] The input module is used to input interactive features into a deep neural network to obtain a predicted quality of service value.

[0129] The loss calculation module is used to calculate the comparative loss;

[0130] The training module is used to modify the parameters of the deep neural network, repeatedly execute the multi-source feature acquisition module to the loss calculation module until the contrast loss no longer decreases, and save the parameters of the deep neural network to obtain the trained model.

[0131] The service quality prediction module is used to input real-time acquired input features into the trained model to obtain the service quality prediction results.

[0132] Specifically, the input feature acquisition module is also used for:

[0133] The input features of the model are represented as

[0134] Xe=(U-ID,U-AS,U-Region,S-ID,S-AS,S-Region,S-WSDL,S-ServiceProvider)

[0135] Wherein, U-ID and S-ID represent the IDs of the user and the service, respectively; U-AS and S-AS represent the Autonomous System Numbers of the user and the service, respectively; U-Region and S-Region represent the regions where the user and the service are located, respectively; S-WSDL represents the WSDL address of the service; and S-ServiceProvider represents the service provider.

[0136] More specifically, the multi-source feature acquisition module is also used for:

[0137] Step 201: Perform one-hot encoding on the input features Xe of the model to obtain...

[0138] Xe=(0100,010,100,1000,100,001,0010,0100)

[0139] Step 202: Map each one-hot encoding to a dense vector of fixed length d. For each domain's text information, obtain an embedding matrix. To represent all textual features in this domain;

[0140] Step 203: Extract the embedding representation of the vectors with a value of 1 in each domain, and remove the embedding representation of the vectors with a value of 0 in each domain to obtain the multi-source feature matrix of the service call. Where f represents the number of files containing all text information.

[0141] More specifically, the interaction feature acquisition module is also used for:

[0142] Step 301: Encode the multi-source feature matrix E using an encoder:

[0143]

[0144] a ij The weight representing the relationship between context feature i and context feature j is calculated as follows:

[0145] a′ ij =relu(W(e i ⊙e j )+b)

[0146]

[0147] Where ReLU represents the activation function, W represents the weight matrix, b represents the bias, and the ⊙ symbol indicates that corresponding positions of two vectors are multiplied, and the result is still a vector rather than a number.

[0148] Step 302: For each row vector p of matrix P i Normalization operation yields:

[0149]

[0150] Where, p i Let represent the i-th row vector in matrix P, where Represents each row vector p i The mean of ε is a minimum value that prevents the denominator of the equation from being zero, while The variance of the row vectors is represented by γ, and β are the parameters that the model can be trained on.

[0151] More specifically, the input module is also used for:

[0152] The normalized matrix P is input into the deep neural network to calculate the predicted service quality value.

[0153] Q = flatten(P)

[0154] y1 = relu(W1Q + b1),

[0155] y2=relu(W2y1+b2),…,

[0156] y L =relu(W L y L-1 +b L )

[0157] Here, `flatten()` represents the operation of converting a two-dimensional matrix into a one-dimensional vector, y L The output of the Lth layer of the neural network, where ReLU represents the activation function, W L Let b represent the weight matrix of the Lth layer. L This represents the deviation of the Lth layer.

[0158] Furthermore, the loss calculation module is also used for:

[0159] Through formula

[0160]

[0161] Calculate the predicted value The comparison loss with the true value R, where M represents the number of all users, N represents the number of all services, and R is the loss between the true value and the true value. ij This represents the actual QoS value. This represents the QoS value predicted by this model.

[0162] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A service quality prediction method based on adaptive feature learning, characterized in that, The method includes: Step 1: Obtain the contextual information of users and services, and the interaction matrix of users and services to obtain the input features of the model; Step 2: Extract multi-source features for users and services based on input features; Step 201: Input features for the model Perform one-hot encoding to obtain Step 202: Map each one-hot code to a fixed-length... In the dense vector, for the text information of each domain, an embedding matrix is ​​obtained. To represent all textual features in this domain; Step 203: Extract the embedding representation of the vectors with a value of 1 in each domain, and remove the embedding representation of the vectors with a value of 0 in each domain to obtain the multi-source feature matrix of the service call. ,in This indicates the number of files containing all text information. Step 3: Encode the multi-source features to obtain the interaction features of users and services; Step 301: Use the encoder to process the multi-source feature matrix Encode: The weight representing the relationship between the i-th context feature and the j-th context feature is calculated as follows: Where relu represents the activation function. Let b represent the weight matrix, and b represent the bias. The symbol indicates that corresponding positions of two vectors are multiplied; Step 302: For the matrix Each row vector Normalization operation yields: in, Representation matrix The first in There are n row vectors, where Represent each row vector The mean, It is a constant, while The variance of the row vector is represented by γ and β, which are two adjustable parameters. Step 4: Input the interaction features into the deep neural network to obtain the predicted quality of service value; The matrix after normalization The input is fed into a deep neural network to calculate the predicted quality of service value. in, This represents the operation of converting a two-dimensional matrix into a one-dimensional vector. No. The output of the layered neural network, This represents the activation function. Indicates the first The weight matrix of the layer, Indicates the first Layer deviation; Step 5: Calculate the contrast loss; Through formula Calculate the predicted value and the true value The comparative loss, of which, This represents the total number of users. This indicates the number of all services. This represents the actual QoS value. This represents the QoS value predicted by this model; Step 6: Modify the parameters of the deep neural network, repeat steps 2 to 5 until the contrastive loss no longer decreases, and save the parameters of the deep neural network to obtain the trained model; Step 7: Input the real-time acquired input features into the trained model to obtain the service quality prediction results.

2. The service quality prediction method based on adaptive feature learning according to claim 1, characterized in that, Step one includes: The input features of the model are represented as in, These represent the user's and service's IDs, respectively. These represent the autonomous system numbers of users and services, respectively. These represent the regions where the user and service are located, respectively. Indicates the WSDL address of the service. This refers to the service provider.

3. A service quality prediction device based on adaptive feature learning, characterized in that, The device includes: The input feature acquisition module is used to obtain the context information of users and services, the interaction matrix of users and services, and to obtain the input features of the model. The multi-source feature acquisition module is used to extract multi-source features of users and services based on input features; the multi-source feature acquisition module is also used for: Step 201: Input features for the model Perform one-hot encoding to obtain Step 202: Map each one-hot code to a fixed-length... In the dense vector, for the text information of each domain, an embedding matrix is ​​obtained. To represent all textual features in this domain; Step 203: Extract the embedding representation of the vectors with a value of 1 in each domain, and remove the embedding representation of the vectors with a value of 0 in each domain to obtain the multi-source feature matrix of the service call. ,in This indicates the number of files containing all text information. The interaction feature acquisition module is used to encode multi-source features to obtain the interaction features of users and services; the interaction feature acquisition module is also used for: Step 301: Use the encoder to process the multi-source feature matrix Encode: The weight representing the relationship between the i-th context feature and the j-th context feature is calculated as follows: Where relu represents the activation function. Let b represent the weight matrix, and b represent the bias. The symbol indicates that corresponding positions of two vectors are multiplied; Step 302: For the matrix Each row vector Normalization operation yields: in, Representation matrix The first in n row vectors, where Represent each row vector The mean, It is a constant, while The variance of the row vector is represented by γ and β, which are two adjustable parameters. The input module is used to input interaction features into the deep neural network to obtain the predicted quality of service value; the input module is also used for: The matrix after normalization The input is fed into a deep neural network to calculate the predicted quality of service value. in, This represents the operation of converting a two-dimensional matrix into a one-dimensional vector. No. The output of the layered neural network, This represents the activation function. Indicates the first The weight matrix of the layer, Indicates the first Layer deviation; The loss calculation module is used to calculate the comparative loss; Through formula Calculate the predicted value and the true value The comparative loss, of which, This represents the total number of users. This indicates the number of all services. This represents the actual QoS value. This represents the QoS value predicted by this model; The training module is used to modify the parameters of the deep neural network, repeatedly execute the multi-source feature acquisition module to the loss calculation module until the contrast loss no longer decreases, and save the parameters of the deep neural network to obtain the trained model. The service quality prediction module is used to input real-time acquired input features into the trained model to obtain the service quality prediction results.

4. The service quality prediction device for adaptive feature learning according to claim 3, characterized in that, The input feature acquisition module is also used for: The input features of the model are represented as in, These represent the user's and service's IDs, respectively. These represent the autonomous system numbers of users and services, respectively. These represent the regions where the user and service are located, respectively. Indicates the WSDL address of the service. This refers to the service provider.

Images