QoS prediction model construction method, device, equipment and readable storage medium

Abstract

The present application relates to a QoS prediction model construction method, apparatus, device and readable storage medium, comprising performing edge loss processing on a user service interaction graph to obtain multiple user service interaction subgraphs; calculating the call similarity and geographic location similarity of each node in each user service interaction subgraph based on QoS, historical calls and latitude and longitude information; constructing a user adjacency graph and a service adjacency graph according to the call similarity and geographic location similarity of each node; embedding information in the user adjacency graph and the service adjacency graph based on ID information and geographic location information to obtain a user initial embedding expression and a service initial embedding expression; performing feature aggregation and learning on the user initial embedding expression and the service initial embedding expression based on a graph attention network to obtain a user final embedding expression and a service final embedding expression; and training a multi-layer perception network based on the final embedding expression of the user and the service to obtain a QoS prediction model, thereby accurately predicting QoS.

Description

Technical Field

[0001] The present application relates to the technical field of software engineering service recommendation, and in particular to a QoS prediction model construction method, apparatus, device and readable storage medium. Background Art

[0002] With the rise of online service technologies, the number of services published on the cloud has rapidly increased, providing internet users with a vast array of cloud services and bringing significant convenience to people's lives. Faced with this growing number of services, finding and selecting the most suitable one from among numerous services offering similar functionality has become a research hotspot in the service computing community. Because non-functional attributes (also known as quality of service (QoS)) measure a service provider's ability to provide services to clients and encompass many dimensions of these non-functional attributes, such as response time and throughput, QoS has become a primary consideration in service selection decisions. QoS attributes are determined by a variety of factors, such as the user's network environment and server configuration. Therefore, improving QoS prediction accuracy is crucial. It can help identify suitable services for users, significantly improving the user experience.

[0003] However, since QoS values ​​cannot be obtained without invoking a service, the only way to predict the QoS value for each service invoked by the target user is based on a small amount of data, and then select the service with the best performance for invocation. However, existing QoS prediction methods are not only unable to mine and learn good user and service embedding representations from a small amount of interaction data in the case of low sparsity, resulting in poor prediction accuracy, but also fail to alleviate the cold start problem. That is, when predictions are made for users who have not yet invoked the service, the prediction effect is poor, resulting in a failure to provide high-quality services to users and fail to meet user needs. Summary of the Invention

[0004] The present application provides a QoS prediction model construction method, apparatus, device and readable storage medium to solve the problem in the related art that QoS cannot be accurately predicted when data has low sparsity and cold start defects exist.

[0005] In a first aspect, a method for constructing a QoS prediction model is provided, comprising the following steps:

[0006] The user service interaction graph generated based on the original data set containing QoS information, historical call information and location information is processed by edge loss to obtain multiple user service interaction subgraphs;

[0007] Calculate the call similarity and geographic location similarity of each node in each user-service interaction subgraph based on QoS information, historical call information, and latitude and longitude information, where the nodes include user nodes and service nodes;

[0008] Build multiple user adjacency graphs and multiple service adjacency graphs based on the call similarity and geographic location similarity of each node;

[0009] Based on the ID information and geographic location information, the user adjacency graph and the service adjacency graph are embedded respectively to obtain the user initial embedding expression and the service initial embedding expression;

[0010] Based on the graph attention network, the initial user embedding expression in each user adjacency graph and the initial service embedding expression in each service adjacency graph are aggregated and learned to obtain the final user embedding expression and the final service embedding expression.

[0011] The multi-layer perception network is trained based on the final embedded expression of users and services to obtain the QoS prediction model.

[0012] In a second aspect, a QoS prediction model construction device is provided, comprising:

[0013] A graph generation module is used to perform edge loss processing on a user-service interaction graph generated based on an original data set containing QoS information, historical call information, and location information to obtain multiple user-service interaction subgraphs; calculate the call similarity and geographic location similarity of each node in each user-service interaction subgraph based on the QoS information, historical call information, and latitude and longitude information, where the nodes include user nodes and service nodes; and construct multiple user adjacency graphs and multiple service adjacency graphs based on the call similarity and geographic location similarity of each node;

[0014] The graph learning module is used to embed information in the user adjacency graph and the service adjacency graph based on ID information and geographic location information, respectively, to obtain the user initial embedding expression and the service initial embedding expression. Based on the graph attention network, the initial user embedding expression in each user adjacency graph and the initial service embedding expression in each service adjacency graph are aggregated and learned to obtain the final user embedding expression and the final service embedding expression.

[0015] The construction module is used to train the multi-layer perception network based on the user's final embedded expression and the service's final embedded expression to obtain a QoS prediction model.

[0016] In a third aspect, a QoS prediction model construction device is provided, comprising: a memory and a processor, wherein the memory stores at least one instruction, and the at least one instruction is loaded and executed by the processor to implement the aforementioned QoS prediction model construction method.

[0017] In a fourth aspect, a computer-readable storage medium is provided, wherein the computer storage medium stores a computer program, and when the computer program is executed by a processor, the aforementioned QoS prediction model construction method is implemented.

[0018] The beneficial effects of the technical solution provided by this application include: being able to predict accurate QoS prediction values ​​in cold start and low sparsity conditions.

[0019] The present application provides a QoS prediction model construction method, apparatus, device and readable storage medium, comprising performing edge loss processing on a user service interaction graph generated based on an original data set containing QoS information, historical call information and location information to obtain multiple user service interaction subgraphs; calculating the call similarity and geographic location similarity of each node in each user service interaction subgraph based on the QoS information, historical call information and latitude and longitude information, wherein the nodes include user nodes and service nodes; constructing multiple user adjacency graphs and multiple service adjacency graphs according to the call similarity and geographic location similarity of each node; embedding information in the user adjacency graph and the service adjacency graph based on ID information and geographic location information to obtain a user initial embedding expression and a service initial embedding expression; performing feature aggregation and learning on the user initial embedding expression in each user adjacency graph and the service initial embedding expression in each service adjacency graph based on a graph attention network to obtain a user final embedding expression and a service final embedding expression; and training a multi-layer perception network based on the user final embedding expression and the service final embedding expression to obtain a QoS prediction model. Through this application, high-quality user embedding expressions and service embedding expressions can be extracted based on the graph structure formed by historical call data, so that the QoS prediction model can predict accurate QoS prediction values ​​under cold start and low sparsity conditions based on the user embedding expressions and service embedding expressions. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without any creative work.

[0021] Figure 1 A flowchart of a QoS prediction model construction method provided in this application;

[0022] Figure 2 Schematic diagram of the impact of parameter loss probability on mean absolute error provided by this application;

[0023] Figure 3 Schematic diagram of the impact of parameter loss probability on root mean square error provided by this application;

[0024] Figure 4 Schematic diagram of the effect of parameter embedding dimension on mean absolute error provided by this application;

[0025] Figure 5 Schematic diagram of the effect of parameter embedding dimension on root mean square error provided by this application;

[0026] Figure 6 Schematic diagram of the effect of the number of neighbor parameters on the mean absolute error provided by this application;

[0027] Figure 7 Schematic diagram of the effect of the number of parameter neighbors on the root mean square error provided by this application;

[0028] Figure 8 Schematic diagram of the impact of parameter similarity weight on mean absolute error provided by this application;

[0029] Figure 9 Schematic diagram of the effect of parameter similarity weight on root mean square error provided by this application;

[0030] Figure 10 A schematic diagram of the structure of a QoS prediction model construction device provided in this application. DETAILED DESCRIPTION

[0031] To make the purpose, technical solutions, and advantages of the embodiments of this application more clear, the technical solutions in the embodiments of this application will be clearly and completely described below in conjunction with the drawings in the embodiments of this application. Obviously, the described embodiments are part of the embodiments of this application, not all of the embodiments. Based on the embodiments in this application, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of this application.

[0032] The embodiments of the present application provide a QoS prediction model construction method, apparatus, device and readable storage medium, which can solve the problem in related technologies that QoS cannot be accurately predicted when data has low sparsity and cold start defects exist.

[0033] See also Figure 1 As shown, the embodiment of the present application provides a method for constructing a QoS prediction model, including the following steps:

[0034] Step S10: performing edge loss processing on the user service interaction graph generated based on the original data set including QoS information, historical call information, and location information to obtain multiple user service interaction subgraphs;

[0035] Exemplarily, in this embodiment, an original data set is obtained, which includes but is not limited to the QoS information of each service, the historical call information of each service called by the user, and the location information of each user; and a specified number of data are randomly selected from the obtained original data set as training data, thereby simulating the sparse call scenario in reality. Specifically, a portion of the call records of the original data set are discarded with a certain probability, thereby generating user-service interaction subgraphs from different perspectives to achieve the purpose of data enhancement. For example, a portion of the call records are discarded with a first probability to generate user-service interaction subgraph A; and a portion of the call records are discarded with a second probability to generate user-service interaction subgraph B. It should be noted that the specific values ​​of the first probability and the second probability can be determined according to actual needs, and the values ​​of the two can be the same or different, and the number of user-service interaction subgraphs generated can also be determined according to actual needs, which is not limited here. This embodiment uses the generation of user-service interaction subgraph A and user-service interaction subgraph B as an example to illustrate how to construct a QoS prediction model.

[0036] The construction process of user-service interaction subgraphs A and B is as follows:

[0037]

[0038]

[0039] Where, Represents the user-service interaction subgraph A, represents the user-service interaction subgraph B, represents the user-service interaction graph, U represents the user set, S represents the service set, M1,M2∈{0,1} |ε| is the masking vector on the two edge sets ε, and p represents the edge loss probability.

[0040] Step S20: Calculate the call similarity and geographic location similarity of each node in each user-service interaction subgraph based on the QoS information, historical call information, and latitude and longitude information. The nodes include user nodes and service nodes.

[0041] Exemplarily, in this embodiment, the generated user-service interaction subgraph is processed as follows: the call similarity is calculated based on the historical call information, and then the address location similarity is calculated based on the known geographic location information (i.e., longitude and latitude information), and then the two similarities are weighted to obtain the final similarity matrix of the user and the service.

[0042] Furthermore, the call similarity and geographic location similarity of each node in each user service interaction subgraph are calculated based on QoS information, historical call information, and latitude and longitude information, including:

[0043] Based on QoS information and historical call information, the user call similarity corresponding to each user in each user-service interaction subgraph and the service call similarity corresponding to each service are calculated respectively;

[0044] Based on the longitude and latitude information, the distances between users and services in each user-service interaction subgraph are calculated respectively. Then, the user geographic location similarity corresponding to each user in each user-service interaction subgraph and the service geographic location similarity corresponding to each service are calculated based on the distances.

[0045] Specifically, based on the QoS information and historical call information, the user call similarity corresponding to each user in each user-service interaction subgraph, as well as the service call similarity and service geographic location similarity corresponding to each service are calculated, including:

[0046] Substitute the QoS information and historical call information into the first calculation formula and the second calculation formula respectively to obtain the user call similarity corresponding to each user in each user-service interaction subgraph and the service call similarity corresponding to each service; wherein the first calculation formula is:

[0047]

[0048] Where S inv (u i ,u j ) represents the user call similarity of user i based on the corresponding user j, represents the set of services called by user i and user j, Indicates the QoS value of user i calling service m, represents the average QoS value of the service called by user i; represents the QoS value of user j calling service m, represents the average QoS value of the service called by user j;

[0049] The second calculation formula is:

[0050]

[0051] Where S inv (s m ,s n ) represents the service call similarity of service m based on service n. Represents the set of users who have called service m and service n. represents the QoS value of service m called by user i, represents the average QoS value of service m; represents the QoS value of service n called by user j, Indicates the average QoS value of service n.

[0052] Specifically, based on the location information, the distances between users and between services in each user-service interaction subgraph are calculated respectively, and the user geographic location similarity corresponding to each user in each user-service interaction subgraph and the service geographic location similarity corresponding to each service are calculated based on the distances, including:

[0053] Substituting each user's location information into the third calculation formula to obtain the distance between users, and substituting each service's location information into the fourth calculation formula to obtain the distance between services;

[0054] Substitute the distance between users into the fifth calculation formula to obtain the user geographic location similarity corresponding to each user in each user-service interaction subgraph, and substitute the distance between services into the sixth calculation formula to obtain the service geographic location similarity corresponding to each service in each user-service interaction subgraph; wherein, the third calculation formula is:

[0055]

[0056] Where, D(u i ,u j ) represents the distance between user i and user j, represents the longitude of user i, represents the latitude of user i, represents the longitude of user j, represents the latitude of user j;

[0057] The fourth calculation formula is:

[0058]

[0059] Where, D(s m ,s n ) represents the distance between service m and service n, represents the longitude of service m, represents the latitude of service m, represents the longitude of service n, Indicates the latitude of service n;

[0060] The fifth calculation formula is:

[0061]

[0062] Where S loc (u i ,u j ) represents the user geographic location similarity between user i and user j, δ urepresents the standard deviation of the user distance matrix;

[0063] The sixth calculation formula is:

[0064]

[0065] Where S loc (s m ,s n ) represents the geographical location similarity between service m and service n, δ s Represents the standard deviation of the service distance matrix.

[0066] Exemplarily, for the user-service interaction subgraph constructed in step S10, the user call similarity of each user-service interaction subgraph is calculated using the following formula:

[0067]

[0068] Where S inv (u i ,u j ) represents the user call similarity of user i based on the corresponding user j, represents the set of services called by user i and user j, Indicates the QoS value of user i calling service m, represents the average QoS value of the service called by user i; represents the QoS value of user j calling service m, represents the average QoS of the service called by user j.

[0069] For the user-service interaction subgraph constructed in step S10, the service call similarity of each user-service interaction subgraph is calculated using the following formula:

[0070]

[0071] Where S inv (s m ,s n ) represents the service call similarity of service m based on service n. Represents the set of users who have called service m and service n. represents the QoS value of service m called by user i, represents the average QoS value of service m; represents the QoS value of service n called by user j, Indicates the average QoS value of service n.

[0072] For the user-service interaction subgraph constructed in step S10, to calculate the user geographic location similarity of each user-service interaction subgraph, it is necessary to first calculate the distance between users. The calculation formula for calculating the distance between two users is as follows:

[0073]

[0074] Where, D(u i ,u j ) represents the distance between user i and user j, represents the longitude of user i, represents the latitude of user i, represents the longitude of user j, represents the latitude of user j;

[0075] Then, based on the calculated distance between users, the geographic location similarity between users is calculated. The calculation formula is as follows:

[0076]

[0077] Where S loc (u i ,u j ) represents the user geographic location similarity between user i and user j, δ u Represents the standard deviation of the user distance matrix.

[0078] Similarly, for the user-service interaction subgraph constructed in step S10, to calculate the service geographic location similarity of each user-service interaction subgraph, it is first necessary to calculate the distance between services. The calculation formula for calculating the distance between two services is as follows:

[0079]

[0080] Where, D(s m ,s n ) represents the distance between service m and service n, represents the longitude of service m, represents the latitude of service m, represents the longitude of service n, Indicates the latitude of service n;

[0081] Then, based on the calculated distance between services, the geographic location similarity between services is calculated using the following formula:

[0082]

[0083] Where S loc (s m ,sn ) represents the geographical location similarity between service m and service n, δ s Represents the standard deviation of the service distance matrix.

[0084] Step S30: construct multiple user adjacency graphs and multiple service adjacency graphs based on the call similarity and geographic location similarity of each node;

[0085] Furthermore, multiple user adjacency graphs and multiple service adjacency graphs are constructed based on the call similarity and geographic location similarity of each node, including:

[0086] Perform weighted calculation on each user's user call similarity and user geographic location similarity to obtain the user similarity corresponding to each user;

[0087] Sort the users in each user-service interaction subgraph in descending order of user similarity to obtain a sorted set of users corresponding to each user-service interaction subgraph;

[0088] Select the top H users from the first user sorting queue corresponding to the first user service interaction subgraph to construct a first user adjacency graph;

[0089] Perform weighted calculation on the service call similarity and service location similarity of each service to obtain the service similarity corresponding to each service;

[0090] Sort the services in each user-service interaction subgraph in descending order of service similarity to obtain a sorted set of services corresponding to each user-service interaction subgraph;

[0091] The first Q services are selected from the first service sorting queue corresponding to the first user service interaction subgraph to construct the first service adjacency graph, where H and Q are both positive integers. The specific values ​​of the two can be determined according to actual needs and are not limited here.

[0092] For example, in this embodiment, taking the user-service interaction subgraph A as an example, the user similarity of each user is calculated according to the following formula:

[0093]

[0094] Where, represents the user similarity of user i, The weight of the user call similarity, where the specific value of the weight can be determined according to actual needs and is not limited here;

[0095] Then, sort the users in the user service interaction subgraph A in descending order of user similarity to obtain the final similarity ranking of each user, that is, the user sorting set corresponding to the user service interaction subgraph A is obtained; then, the top H users in the user service interaction subgraph A are intercepted as neighbors to construct the user adjacency graph. (At this time, the user adjacency graph only contains user nodes); Similarly, another user adjacency graph can be constructed based on the user service interaction subgraph B

[0096] At the same time, the service similarity of each service is calculated according to the following formula:

[0097]

[0098] Where, represents the service similarity of service m, The weight representing the similarity of service calls. The specific value of the weight can be determined according to actual needs and is not limited here.

[0099] Then, sort the services in the user-service interaction subgraph A in descending order of service similarity to obtain the final similarity ranking of each service, that is, the service sorting set corresponding to the user-service interaction subgraph A; then, from the service sorting set corresponding to the service-service interaction subgraph A, intercept the top Q services as neighbors to construct the service adjacency graph. (At this time, the service adjacency graph only contains service nodes); Similarly, another service adjacency graph can be constructed based on the user service interaction subgraph B

[0100] Step S40: Based on the ID information and the geographic location information, information is embedded in the user adjacency graph and the service adjacency graph respectively to obtain the user initial embedding expression and the service initial embedding expression;

[0101] Exemplarily, in this embodiment, user high-dimensional information is embedded, that is, information embedding is performed on the user to obtain high-dimensional features: the user's unique identification ID and the user's geographic location information are selected as initial features, and the user's unique identification ID is embedded in the K-dimensional feature space to obtain a high-dimensional vector, and then the user's geographic location information is embedded in another K-dimensional feature space to obtain another high-dimensional vector, and finally the two high-dimensional vectors are spliced ​​to obtain the high-dimensional embedding expression of the user in the user adjacency graph (that is, the user's initial embedding expression); at the same time, service high-dimensional information is embedded, that is, information embedding is performed on the service to obtain high-dimensional features: the service's unique identification ID and the service's geographic location information are selected as initial features, and the user's unique identification ID is embedded in the K-dimensional feature space to obtain a high-dimensional vector, and then the service's geographic location information is embedded in another K-dimensional feature space to obtain another high-dimensional vector, and finally the two high-dimensional vectors are spliced ​​to obtain the high-dimensional embedding expression of the service in the service adjacency graph (that is, the service's initial embedding expression).

[0102] Step S50: Based on the graph attention network, feature aggregation and learning are performed on the user initial embedding expression in each user adjacency graph and the service initial embedding expression in each service adjacency graph to obtain the user final embedding expression and the service final embedding expression;

[0103] Furthermore, based on the graph attention network, the initial user embedding expression in each user adjacency graph and the initial service embedding expression in each service adjacency graph are aggregated and learned to obtain the final user embedding expression and the final service embedding expression, including:

[0104] Based on the graph attention network, feature aggregation and learning are performed on the user initial embedding expression in each user adjacency graph and the service initial embedding expression in each service adjacency graph, respectively, to obtain the first user embedding expression of each user in each user adjacency graph and the first service embedding expression of each service;

[0105] Perform weighted summation of the first user embedding expression of each user in each user adjacency graph to obtain the final user embedding expression corresponding to each user;

[0106] The first service embedding expression of each service in each service adjacency graph is weightedly summed to obtain the final service embedding expression corresponding to each service.

[0107] Exemplarily, in this embodiment, a graph attention network is used in the adjacency graph for feature aggregation to learn high-dimensional features of users and services, and then the embeddings of multiple views are weighted by attention to obtain the final high-dimensional features of users and services.

[0108] Specifically, in order to enable nodes to learn well-characterized embedded expressions from neighbor information, this embodiment introduces a graph attention network for graph representation learning. Graph attention networks are different from graph convolutional networks. Graph convolutional networks use average weights to aggregate node neighbor information to learn representations, while graph attention networks assign different weights to different neighbors for weighted aggregation, thereby learning better feature expressions for more closely linked neighbor nodes. Taking the user adjacency graph as an example, for each node in the graph, the calculation formula for the first user embedding expression corresponding to each node is as follows:

[0109]

[0110]

[0111] Where, represents the user adjacency graph k, represents the importance of user node j to user node i in user adjacency graph k, represents the initial user embedding expression of user i in the user adjacency graph k, represents the initial embedding expression of user j in the user adjacency graph k, || is the concatenation operation, represents the transposed matrix of user adjacency graph k, σ is the softmax activation function, represents the first user embedding expression of user k in the user adjacency graph, represents the neighbors of user i in the user adjacency graph k.

[0112] As can be seen, this embodiment learns different first-user embedding expressions for each user from different user adjacency graphs. Therefore, this embodiment performs multi-graph aggregation to obtain the final features: the first-user embedding expressions of users in different user adjacency graphs are combined to obtain the final well-characterized embedding expression for each user. That is, the attention mechanism is used to perform weighted summation on different user adjacency graphs to obtain the final embedding expression of the user; the final embedding expression of the user is as follows:

[0113]

[0114]

[0115] Where, and Respectively represent the importance of user i in the user adjacency graph k and user adjacency graph o, and denote the first user embedding expression of user i in user adjacency graph k and user adjacency graph o, respectively. represents the final embedding expression of user i.

[0116] Similarly, taking the service adjacency graph as an example, for each node in the graph, the calculation formula for the first service embedding expression corresponding to each node is as follows:

[0117]

[0118]

[0119] Where, represents the service adjacency graph x, represents the importance of service node n to service node m in the service adjacency graph x, represents the initial embedding expression of service m in the service adjacency graph x, represents the initial embedding expression of service n in the service adjacency graph x, || is the concatenation operation, represents the transposed matrix of the service adjacency graph x, σ is the softmax activation function, represents the first service embedding expression of the service in the service adjacency graph x, represents the neighbors of service m in the service adjacency graph x.

[0120] As can be seen, this embodiment learns different first service embedding expressions for each service from different service adjacency graphs. Therefore, this embodiment performs multi-graph aggregation to obtain the final features: the first service embedding expressions of services in different service adjacency graphs are integrated to obtain the final well-characterized embedding expression of each service. That is, the attention mechanism is used to perform weighted summation on different service adjacency graphs to obtain the final embedding expression of the service. The final embedding expression of the service is as follows:

[0121]

[0122]

[0123] Where, and Respectively represent the importance of service m in the service adjacency graph x and service adjacency graph y, and denote the first service embedding expression of service m in the service adjacency graph x and the user adjacency graph y, respectively. Represents the final embedded expression of service m.

[0124] Step S60: training the multi-layer perception network based on the final embedded expression of the user and the final embedded expression of the service to obtain a QoS prediction model.

[0125] For example, in this embodiment, a decoder is constructed. The decoder inputs the high-dimensional features (final embedded expressions) of users and services into a multi-layer perceptron for training to obtain a QoS prediction model. Therefore, when QoS prediction is required for a target user, only the final embedded expression corresponding to each target user and the final embedded expressions of all services need to be input into the QoS prediction model. The multi-layer perceptron in the QoS prediction model can then perform QoS prediction for the target user. Specifically, after obtaining the final embedded expression of the user, the final embedded expression of the user is transcoded and output. The specific conversion is as follows:

[0126]

[0127] Where g1 represents the output after transcoding the user's final embedded expression, σ represents the softmax activation function, and W1 represents the weight, which can be determined according to actual needs. represents the final user embedding expression of target user i, b1 represents the bias, which is a constant value;

[0128] At the same time, after obtaining the final embedded expression of the service, the final embedded expression of the service is transcoded and output. The specific conversion is as follows:

[0129]

[0130] Where g2 represents the output of the final embedded expression of the service after transcoding, σ represents the softmax activation function, and W2 represents the weight, which can be determined according to actual needs. represents the final service embedding expression of the target service m, b2 represents the bias, which is a constant value;

[0131] Finally, g1 and g2 are concatenated and the concatenated result is input into a neural network layer for calculation to obtain the final predicted QoS value.

[0132]

[0133] Where, represents the final predicted QoS value, σ represents the softmax activation function, W3 represents the weight, which can be determined according to actual needs, and b3 represents the bias, which is a constant value.

[0134] In summary, the purpose of this application is to alleviate the data sparsity and cold start issues in service QoS prediction. That is, under the premise of simulating real-world situations, high-dimensional features of users and services are mined from extremely sparse interaction data to perform QoS numerical prediction. Therefore, through this application, high-quality user embedding expressions and service embedding expressions can be extracted based on the graph structure formed by historical call data, so that the QoS prediction model can predict accurate QoS prediction values ​​based on these user embedding expressions and service embedding expressions in cold start and low sparsity conditions.

[0135] Furthermore, after the step of training the multi-layer perception network based on the user final embedded expression and the service final embedded expression to obtain the QoS prediction model, the method further includes:

[0136] The variance between the QoS prediction value output by the QoS prediction model and the QoS true value is used as the prediction loss function;

[0137] Backpropagation is performed based on the prediction loss function and the contrast loss function to update the user final embedding expression and the service final embedding expression to obtain a new user final embedding expression and a new service final embedding expression;

[0138] The QoS prediction model is trained based on the new user final embedding expression and the new service final embedding expression to obtain a new QoS prediction model;

[0139] The contrast loss function is:

[0140]

[0141]

[0142]

[0143] Where, L ssl represents the contrast loss function, represents the user contrast loss function, represents the service comparison loss function, represents the embedding expression of user k in the user adjacency graph, represents the embedded expression of the user in the user adjacency graph o, represents the embedded expression of the service in the service adjacency graph x, represents the embedding expression of the service in the service adjacency graph y, cos(·) represents the similarity between two embedding expressions, and τ represents the temperature parameter in the normalization function.

[0144] For demonstration, in this embodiment, multi-graph comparative learning will also be performed, that is, the same node in different adjacency graphs will be regarded as a positive pair, and different nodes will be regarded as a negative pair. It is assumed that some high-dimensional features of the nodes are fixed, such as the location information of the user and the network environment, the functions provided by the service, and the server configuration. Therefore, they should be as close as possible to different views to indicate the high-level abstract characteristics of the user and the service, which means that the embedding representations of the positive pairs should be as close as possible, while the features of the negative pairs should be as different as possible. Therefore, in this embodiment, the following loss function will be used to maximize the similarity of the positive pairs and minimize the similarity of the negative pairs. Among them, the user comparative learning loss function and the service comparative loss function are as follows:

[0145]

[0146]

[0147] Where, represents the user contrast loss function, represents the service comparison loss function, represents the embedding expression of user k in the user adjacency graph, represents the embedded expression of the user in the user adjacency graph o, represents the embedded expression of the service in the service adjacency graph x, represents the embedding expression of the service in the service adjacency graph y, cos(·) represents the similarity between two embedding expressions (i.e., two vectors), which can be set to cosine similarity, and τ represents the temperature parameter in the normalization function (i.e., softmax).

[0148] Then, the comprehensive user and service contrast loss function is used as the final overall contrast loss function L ssl ,Right now:

[0149]

[0150] At the same time, the variance between the QoS prediction value output by the QoS prediction model and the QoS true value is used as the prediction loss function:

[0151]

[0152] Where, L main represents the prediction loss function, represents the QoS prediction value, r u,s Indicates the actual value of QoS.

[0153] Finally, the prediction loss function and the comparison loss function are combined as the overall loss function L of the model to perform back propagation training, that is:

[0154] L=L main +Lssl

[0155] The model is trained by back propagation through the loss function L, and then the various parameters in the model are adjusted to make the various parameters in the model reach the optimal value.

[0156] The following is a specific example of using the method provided by this application to perform QoS prediction. Taking QoS history records collected from WSDream as an example, and combined with the accompanying drawings, the implementation process of this application is described in detail. The model involved in this application is an encoder-decoder structure, where the encoder and decoder are connected via an attention mechanism network, and the decoder generates a diverse API sequence through a diversified beam search.

[0157] In order to simulate the real situation, this embodiment will randomly delete a specified number of observation data to retain a few observation data as training samples to simulate the low sparse data in the real situation, and use the data that were not selected as training samples in the original data as prediction samples to test the model prediction effect.

[0158] Specifically, the graph generation module first drops edges on the training samples (i.e., the user-service interaction graph) with a specific probability to construct user-service interaction subgraphs from different perspectives. Then, for each user-service interaction subgraph, the corresponding call similarity and geographic location similarity are calculated and weighted summed to obtain the final similarity ranking. A specified number of top-K similar nodes are selected as neighbors to construct the user adjacency graph and service adjacency graph.

[0159] Then, a graph network coding model is built through the graph learning module to learn high-dimensional embedding expressions of users and services: for the user adjacency graph, information is embedded based on the user's unique identifier and geographic location information to obtain a high-dimensional representation vector, and the graph attention network is used to propagate information in the user adjacency graph to learn the user representation. Then, the user adjacency graph from multiple perspectives is integrated, and the idea of ​​graph comparative learning is used to maximize the similarity of positive pairs and minimize the similarity of negative pairs, so as to learn better user representation, and the final user embedding expression is obtained by integrating multiple views. For the service adjacency graph, information is embedded based on the service's unique identifier and geographic location information to obtain a high-dimensional representation vector, and the graph attention network is used to propagate information in the service adjacency graph to learn the service representation. Then, the service adjacency graph from multiple perspectives is integrated, and the idea of ​​graph comparative learning is used to maximize the similarity of positive pairs and minimize the similarity of negative pairs, so as to learn better service representation, and the final service embedding expression is obtained by integrating multiple views.

[0160] Finally, a decoder model is built through the construction module, and the high-dimensional features of users and services are input into the multi-layer perceptron (i.e., QoS prediction module) to predict the QoS of the target user.

[0161] The following parameters will be compared based on the above experimental results:

[0162] The parameters in the model of this embodiment are compared to obtain the best parameters to obtain the best effect of the model.

[0163] Among them, MAE (mean absolute error) shows the average of the absolute values ​​of the deviations of all individual observations from the arithmetic mean. The smaller the MAE, the better the prediction result. The calculation formula of MAE is:

[0164]

[0165] RMSE (root mean square error): represents the deviation between the observed value and the true value. The smaller the RMSE, the better the prediction result. The calculation formula of RMSE is:

[0166]

[0167] Where Q u,s Indicates the actual QoS value of user u calling service s, It represents the predicted QoS value of user u calling service s, and N represents the number of QoS values.

[0168] First, we compare the parameter dropout_rate (i.e., edge dropout probability), which determines the probability of retaining edges when constructing the user-service interaction subgraph. Figure 2 and Figure 3 As shown in the figure (D1 to D5 in the figure represent different data densities), when the dropout_rate decreases, the prediction accuracy improves, but after reaching the threshold of 0.8, the prediction results decrease as the dropout_rate decreases. Therefore, in this embodiment, the dropout_rate parameter of the model can be selected as 0.8 to enable the model to obtain the best prediction effect.

[0169] Secondly, the parameter similarity_weight (i.e., embedding dimension) is compared. The parameter similarity_weight determines the proportion of call similarity in the weighted sum of similarities. The final similarity ranking is obtained by combining call similarity and location similarity. Therefore, considering that call similarity is affected by geographic location, a weight mechanism is introduced to balance the two similarities. Figure 4 and Figure 5 As shown in Figure 2, when the weight is set to 0.7, the prediction effect is the best, indicating that the model has a good balance between the two similarities.

[0170] Then compare the parameter num_neighbor (number of neighbors), which determines the weight of the similarity when the similarity is weighted. Figure 6 and Figure 7 As shown, the MAE and RMSE values ​​decrease as the number of nodes increases until a certain threshold is reached. This indicates that as the number of neighbors increases, nodes aggregate more information from their neighbors and mine the graph structure to learn better representations, resulting in more accurate QoS predictions. However, after reaching the threshold, QoS prediction accuracy improves slightly and even decreases as the number of neighbors increases. This may be because node embeddings may be affected by weakly correlated neighbors, resulting in poorly learned representations. Therefore, in this embodiment, the model parameter num_neighbor can be set to 0.7 to achieve the best prediction results.

[0171] Finally, the parameter embedding_dimension (i.e. similarity weight) is compared. The parameter embedding_dimension determines the proportion of similarity when the similarity is weighted and summed. Figure 8 and Figure 9 As shown in the figure, MAE and RMSE decrease as the embedding size increases until a certain threshold is reached. Therefore, in this embodiment, the embedding_dimension parameter of the model can be selected as 0.7 so that the model can obtain the best prediction effect.

[0172] In addition, the following will compare the performance of the benchmark method MGCF (multi-component graph convolutional collaborative filtering and deep factorization machine) and the LGC (location-aware graph contrastive learning) provided by this embodiment using different data densities (for example, density of 0.5% to 4%) and based on the two evaluation indicators of MAE and RMSE. Among them, MGCF uses a graph network to learn representations of users and services to predict QoS. As shown in Table 1, the graph network of MGCF has difficulty learning good representations when the data sparsity is low, while the prediction accuracy of this embodiment is significantly improved compared with the baseline method.

[0173] Table 1 Performance comparison table

[0174]

[0175] The present application also provides a QoS prediction model construction device, including:

[0176] The graph generation module is used to perform edge loss processing on a user-service interaction graph generated from an original dataset containing QoS information, historical call information, and location information to obtain multiple user-service interaction subgraphs. The module calculates the call similarity and geographic location similarity of each node in each user-service interaction subgraph based on the QoS information, historical call information, and latitude and longitude information. The nodes include user nodes and service nodes. Multiple user adjacency graphs and multiple service adjacency graphs are constructed based on the call similarity and geographic location similarity of each node.

[0177] The graph learning module is used to embed information in the user adjacency graph and the service adjacency graph based on ID information and geographic location information, respectively, to obtain the user initial embedding expression and the service initial embedding expression. Based on the graph attention network, the initial user embedding expression in each user adjacency graph and the initial service embedding expression in each service adjacency graph are aggregated and learned to obtain the final user embedding expression and the final service embedding expression.

[0178] The construction module is used to train the multi-layer perception network based on the user's final embedded expression and the service's final embedded expression to obtain a QoS prediction model.

[0179] Furthermore, the graph generation module is specifically used to:

[0180] Based on QoS information and historical call information, the user call similarity corresponding to each user in each user-service interaction subgraph and the service call similarity corresponding to each service are calculated respectively;

[0181] Based on the longitude and latitude information, the distances between users and services in each user-service interaction subgraph are calculated respectively. Then, the user geographic location similarity corresponding to each user in each user-service interaction subgraph and the service geographic location similarity corresponding to each service are calculated based on the distances.

[0182] Furthermore, the graph generation module is specifically used to:

[0183] Based on QoS information and historical call information, the user call similarity corresponding to each user in each user-service interaction subgraph, as well as the service call similarity and service geographic location similarity corresponding to each service are calculated respectively, including:

[0184] Substituting the QoS information and the historical call information into the first calculation formula and the second calculation formula respectively to obtain the user call similarity corresponding to each user in each user-service interaction subgraph and the service call similarity corresponding to each service;

[0185] Among them, the first calculation formula is:

[0186]

[0187] Where S inv (u i ,u j ) represents the user call similarity of user i based on the corresponding user j, s∈ represents the set of services called by user i and user j, Indicates the QoS value of user i calling service m, represents the average QoS value of the service called by user i; represents the QoS value of user j calling service m, represents the average QoS value of the service called by user j;

[0188] The second calculation formula is:

[0189]

[0190] Where S inv (s m ,s n ) represents the service call similarity of service m based on service n. Represents the set of users who have called service m and service n. represents the QoS value of service m called by user i, represents the average QoS value of service m; represents the QoS value of service n called by user j, Indicates the average QoS value of service n.

[0191] Furthermore, the graph generation module is specifically used to:

[0192] Substituting each user's location information into the third calculation formula to obtain the distance between users, and substituting each service's location information into the fourth calculation formula to obtain the distance between services;

[0193] Substitute the distance between users into the fifth calculation formula to obtain the user geographic location similarity corresponding to each user in each user-service interaction subgraph, and substitute the distance between services into the sixth calculation formula to obtain the service geographic location similarity corresponding to each service in each user-service interaction subgraph;

[0194] Among them, the third calculation formula is:

[0195]

[0196] Where, D(u i ,u j) represents the distance between user i and user j, represents the longitude of user i, represents the latitude of user i, represents the longitude of user j, represents the latitude of user j;

[0197] The fourth calculation formula is:

[0198]

[0199] Where, D(s m ,s n ) represents the distance between service m and service n, represents the longitude of service m, represents the latitude of service m, represents the longitude of service n, Indicates the latitude of service n;

[0200] The fifth calculation formula is:

[0201]

[0202] Where S loc (u i ,u j ) represents the user geographic location similarity between user i and user j, δ u represents the standard deviation of the user distance matrix;

[0203] The sixth calculation formula is:

[0204]

[0205] Where S loc (s m ,s n ) represents the geographical location similarity between service m and service n, δ s Represents the standard deviation of the service distance matrix.

[0206] Furthermore, the graph generation module is specifically used to:

[0207] Perform weighted calculation on each user's user call similarity and user geographic location similarity to obtain the user similarity corresponding to each user;

[0208] Sort the users in each user-service interaction subgraph in descending order of user similarity to obtain a sorted set of users corresponding to each user-service interaction subgraph;

[0209] Select the top H users from the first user sorting queue corresponding to the first user service interaction subgraph to construct a first user adjacency graph;

[0210] Perform weighted calculation on the service call similarity and service location similarity of each service to obtain the service similarity corresponding to each service;

[0211] Sort the services in each user-service interaction subgraph in descending order of service similarity to obtain a sorted set of services corresponding to each user-service interaction subgraph;

[0212] The first Q services are selected from the first service sorting queue corresponding to the first user-service interaction subgraph to construct a first service adjacency graph, where H and Q are both positive integers.

[0213] Furthermore, the graph learning module is specifically used to:

[0214] Based on the graph attention network, feature aggregation and learning are performed on the user initial embedding expression in each user adjacency graph and the service initial embedding expression in each service adjacency graph, respectively, to obtain the first user embedding expression of each user in each user adjacency graph and the first service embedding expression of each service;

[0215] Perform weighted summation of the first user embedding expression of each user in each user adjacency graph to obtain the final user embedding expression corresponding to each user;

[0216] The first service embedding expression of each service in each service adjacency graph is weightedly summed to obtain the final service embedding expression corresponding to each service.

[0217] Furthermore, the building blocks are used to:

[0218] The variance between the QoS prediction value output by the QoS prediction model and the QoS true value is used as the prediction loss function;

[0219] Backpropagation is performed based on the prediction loss function and the contrast loss function to update the user final embedding expression and the service final embedding expression to obtain a new user final embedding expression and a new service final embedding expression;

[0220] The QoS prediction model is trained based on the new user final embedding expression and the new service final embedding expression to obtain a new QoS prediction model;

[0221] The contrast loss function is:

[0222]

[0223]

[0224]

[0225] Where, Lssl represents the contrast loss function, represents the user contrast loss function, represents the service comparison loss function, represents the embedding expression of user k in the user adjacency graph, represents the embedded expression of the user in the user adjacency graph o, represents the embedded expression of the service in the service adjacency graph x, represents the embedding expression of the service in the service adjacency graph y, cos(·) represents the similarity between two embedding expressions, and τ represents the temperature parameter in the normalization function.

[0226] It should be noted that those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working processes of the above-described devices and units can refer to the corresponding processes in the aforementioned QoS prediction model construction method embodiment, and will not be repeated here.

[0227] An embodiment of the present application also provides a QoS prediction model construction device, comprising: a memory, a processor, and a network interface connected via a system bus, wherein at least one instruction is stored in the memory, and at least one instruction is loaded and executed by the processor to implement all or part of the steps of the aforementioned QoS prediction model construction method.

[0228] Among them, the network interface is used for network communication, such as sending assigned tasks, etc. Those skilled in the art will understand that Figure 10 The structure shown in the figure is only a block diagram of a part of the structure related to the solution of the present application, and does not constitute a limitation on the computer device to which the solution of the present application is applied. The specific computer device may include more or fewer components than shown in the figure, or combine certain components, or have a different component arrangement.

[0229] A processor can be a CPU, other general-purpose processors, application-specific integrated circuits, field-programmable gate arrays, or other programmable logic devices. A general-purpose processor can be a microprocessor or any conventional processor. The processor is the control center of a computer device, connecting all parts of the computer device using various interfaces and circuits.

[0230] Memory can be used to store computer programs and / or modules. The processor implements various functions of the computer device by running or executing the computer programs and / or modules stored in the memory and accessing the data stored in the memory. Memory can primarily include a program storage area and a data storage area. The program storage area can store the operating system and at least one application required for a function, while the data storage area can store data generated based on the use of the mobile phone, etc. In addition, memory can include high-speed random access memory and non-volatile memory, such as hard disks, internal memory, smart memory cards, flash memory devices, or other volatile solid-state memory devices.

[0231] An embodiment of the present application also provides a computer-readable storage medium on which a computer program is stored. When the computer program is executed by a processor, all or part of the steps of the aforementioned QoS prediction model construction method are implemented.

[0232] The embodiments of the present application implement all or part of the aforementioned processes, and may also be completed by instructing related hardware through a computer program. The computer program may be stored in a computer-readable storage medium, and the computer program, when executed by a processor, may implement the steps of each of the above methods. Among them, the computer program includes computer program code, and the computer program code may be in the form of source code, object code, executable file or some intermediate form, etc. Computer-readable media may include: any entity or device capable of carrying computer program code, a USB flash drive, a mobile hard disk, an optical disc, a computer memory, a read-only memory, a random access memory, a telecommunications signal and a software distribution medium, etc. It should be noted that the content contained in the computer-readable medium may be appropriately increased or decreased according to the requirements of legislation and patent practice in the jurisdiction. For example, in some jurisdictions, according to legislation and patent practice, computer-readable media do not include electric carrier signals and telecommunications signals.

[0233] Those skilled in the art will appreciate that the embodiments of the present application may be provided as methods, systems, servers, or computer program products. Therefore, the present application may adopt the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Furthermore, the present application may adopt the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to magnetic disk storage and optical storage, etc.) containing computer-usable program code.

[0234] It should be noted that the terms "comprises," "includes," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or system that includes a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements that are inherent to such process, method, article, or system.

[0235] The above description is only a specific implementation method of the present application, which enables those skilled in the art to understand or implement the present application. It is not limited to the embodiments shown in this article, but must comply with the widest scope consistent with the principles and novel features of the application in this article.

Claims

1. A QoS prediction model construction method, characterized in that: The following steps are involved: The user service interaction graph generated based on the original data set containing QoS information, historical call information and location information is processed by edge loss to obtain multiple user service interaction subgraphs; Calculate the call similarity and geographic location similarity of each node in each user-service interaction subgraph based on QoS information, historical call information, and latitude and longitude information, where the nodes include user nodes and service nodes; Build multiple user adjacency graphs and multiple service adjacency graphs based on the call similarity and geographic location similarity of each node; Based on the ID information and geographic location information, the user adjacency graph and the service adjacency graph are embedded respectively to obtain the user initial embedding expression and the service initial embedding expression; Based on the graph attention network, the initial user embedding expression in each user adjacency graph and the initial service embedding expression in each service adjacency graph are aggregated and learned to obtain the final user embedding expression and the final service embedding expression. Based on the final embedded expression of users and services, a multi-layer perception network is trained to obtain a QoS prediction model. The calculation of the call similarity and geographic location similarity of each node in each user service interaction subgraph based on the QoS information, historical call information, and latitude and longitude information includes: Based on QoS information and historical call information, the user call similarity corresponding to each user in each user-service interaction subgraph and the service call similarity corresponding to each service are calculated respectively; Based on the longitude and latitude information, the distances between users and services in each user-service interaction subgraph are calculated respectively, and the user geographic location similarity corresponding to each user and the service geographic location similarity corresponding to each service in each user-service interaction subgraph are calculated based on the distances; The method of constructing multiple user adjacency graphs and multiple service adjacency graphs based on the call similarity and geographic location similarity of each node includes: Perform weighted calculation on each user's user call similarity and user geographic location similarity to obtain the user similarity corresponding to each user; Sort the users in each user-service interaction subgraph in descending order of user similarity to obtain a sorted set of users corresponding to each user-service interaction subgraph; Select the top H users from the first user sorting queue corresponding to the first user service interaction subgraph to construct a first user adjacency graph; Perform weighted calculation on the service call similarity and service location similarity of each service to obtain the service similarity corresponding to each service; Sort the services in each user-service interaction subgraph in descending order of service similarity to obtain a sorted set of services corresponding to each user-service interaction subgraph; The first Q services are selected from the first service sorting queue corresponding to the first user-service interaction subgraph to construct a first service adjacency graph, where H and Q are both positive integers.

2. The QoS prediction model construction method according to claim 1, characterized in that: The method of calculating the user call similarity corresponding to each user in each user-service interaction subgraph and the service call similarity and service geographic location similarity corresponding to each service based on the QoS information and the historical call information includes: Substituting the QoS information and the historical call information into the first calculation formula and the second calculation formula respectively to obtain the user call similarity corresponding to each user in each user-service interaction subgraph and the service call similarity corresponding to each service; Among them, the first calculation formula is: Where, Represents a user User-based The corresponding user call similarity, Represents a user and users The set of services called, Represents a user Calling the service QoS value, Represents a user The average QoS value of the service being called; Represents a user j Calling the service QoS value, Represents a user j The average QoS value of the service being called; The second calculation formula is: Where, Representation Service Service-based The corresponding service call similarity, Indicates that the service has been called and services The user collection, Represents a user Called service QoS value, Representation Service Average QoS value; Represents a user j Called service QoS value, Representation Service n The average QoS value.

3. The QoS prediction model construction method according to claim 1, characterized in that: The method of calculating the distances between users and between services in each user-service interaction subgraph based on the location information, and calculating the user geographic location similarity corresponding to each user and the service geographic location similarity corresponding to each service in each user-service interaction subgraph based on the distances, includes: Substituting each user's location information into the third calculation formula to obtain the distance between users, and substituting each service's location information into the fourth calculation formula to obtain the distance between services; Substitute the distance between users into the fifth calculation formula to obtain the user geographic location similarity corresponding to each user in each user-service interaction subgraph, and substitute the distance between services into the sixth calculation formula to obtain the service geographic location similarity corresponding to each service in each user-service interaction subgraph; Among them, the third calculation formula is: Where, Represents a user With users The distance between Represents a user Longitude, Represents a user Latitude, Represents a user j Longitude, Represents a user Latitude; The fourth calculation formula is: Where, Representation Service and Services The distance between Representation Service Longitude, Representation Service Latitude, Representation Service Longitude, Representation Service Latitude; The fifth calculation formula is: Where, Represents a user With users The similarity of user geographic locations between represents the standard deviation of the user distance matrix; The sixth calculation formula is: Where, Representation Service and Services The similarity of the geographical locations of services between Represents the standard deviation of the service distance matrix.

4. The QoS prediction model construction method according to claim 1, wherein: The graph attention network performs feature aggregation and learning on the user initial embedding expression in each user adjacency graph and the service initial embedding expression in each service adjacency graph, respectively, to obtain the user final embedding expression and the service final embedding expression, including: Based on the graph attention network, feature aggregation and learning are performed on the user initial embedding expression in each user adjacency graph and the service initial embedding expression in each service adjacency graph, respectively, to obtain the first user embedding expression of each user in each user adjacency graph and the first service embedding expression of each service; Perform weighted summation of the first user embedding expression of each user in each user adjacency graph to obtain the final user embedding expression corresponding to each user; The first service embedding expression of each service in each service adjacency graph is weightedly summed to obtain the final service embedding expression corresponding to each service.

5. The QoS prediction model construction method according to claim 1, wherein: After the step of training the multi-layer perception network based on the user's final embedded expression and the service's final embedded expression to obtain the QoS prediction model, the method further includes: The variance between the QoS prediction value output by the QoS prediction model and the QoS true value is used as the prediction loss function; Backpropagation is performed based on the prediction loss function and the contrast loss function to update the user final embedding expression and the service final embedding expression to obtain a new user final embedding expression and a new service final embedding expression; The QoS prediction model is trained based on the new user final embedding expression and the new service final embedding expression to obtain a new QoS prediction model; The contrast loss function is: Where, represents the contrast loss function, represents the user contrast loss function, represents the service comparison loss function, Represents the user in the user adjacency graph The embedded expression in Represents the user in the user adjacency graph The embedded expression in Represents a service in the service adjacency graph x The embedded expression in Represents a service in the service adjacency graph The embedded expression in represents the similarity between two embedded expressions, represents the temperature parameter in the normalization function.

6. A QoS prediction model construction device, characterized in that: include: A graph generation module is used to perform edge loss processing on a user service interaction graph generated based on an original data set containing QoS information, historical call information, and location information, to obtain multiple user service interaction subgraphs; Calculate the call similarity and geographic location similarity of each node in each user-service interaction subgraph based on QoS information, historical call information, and latitude and longitude information, where the nodes include user nodes and service nodes; construct multiple user adjacency graphs and multiple service adjacency graphs based on the call similarity and geographic location similarity of each node; The graph learning module is used to embed information in the user adjacency graph and the service adjacency graph based on ID information and geographic location information, respectively, to obtain the user initial embedding expression and the service initial embedding expression; Based on the graph attention network, the initial user embedding expression in each user adjacency graph and the initial service embedding expression in each service adjacency graph are aggregated and learned to obtain the final user embedding expression and the final service embedding expression. A construction module is used to train a multi-layer perception network based on the final embedding expression of the user and the final embedding expression of the service to obtain a QoS prediction model; The graph generation module is specifically used for: Based on QoS information and historical call information, the user call similarity corresponding to each user in each user-service interaction subgraph and the service call similarity corresponding to each service are calculated respectively; Based on the longitude and latitude information, the distances between users and services in each user-service interaction subgraph are calculated respectively, and the user geographic location similarity corresponding to each user and the service geographic location similarity corresponding to each service in each user-service interaction subgraph are calculated based on the distances; The graph generation module is further specifically configured to: Perform weighted calculation on each user's user call similarity and user geographic location similarity to obtain the user similarity corresponding to each user; Sort the users in each user-service interaction subgraph in descending order of user similarity to obtain a sorted set of users corresponding to each user-service interaction subgraph; Select the top H users from the first user sorting queue corresponding to the first user service interaction subgraph to construct a first user adjacency graph; Perform weighted calculation on the service call similarity and service location similarity of each service to obtain the service similarity corresponding to each service; Sort the services in each user-service interaction subgraph in descending order of service similarity to obtain a sorted set of services corresponding to each user-service interaction subgraph; The first Q services are selected from the first service sorting queue corresponding to the first user-service interaction subgraph to construct a first service adjacency graph, where H and Q are both positive integers.

7. A QoS prediction model building device, characterized in that: include: A memory and a processor, wherein the memory stores at least one instruction, and the at least one instruction is loaded and executed by the processor to implement the QoS prediction model construction method according to any one of claims 1 to 5.

8. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the method for constructing a QoS prediction model according to any one of claims 1 to 5 is implemented.

Images