A service trust evaluation method for cross-border service networks

Abstract

The present invention discloses a service trust evaluation method for a cross-border service network, comprising: outputting service trust to the user, the service trust comprising direct service trust and indirect service trust; fusing the average value of each QoS indicator within a set time period and the service integrity, assigning a combined weight to the fusion result, and then summing up each QoS indicator assigned a combined weight to obtain the direct service trust; introducing user social relationships based on the LightGCN model, aggregating multi-layer user embedding vectors with user embedding vectors of corresponding layers based on the user-service interaction network to obtain a final user embedding vector representation; and obtaining the indirect service trust based on the initial user embedding vector, the service consumer embedding vector, and the service embedding vector using the STLightGCN model. This method can ensure the fairness and impartiality of the evaluation method and meet the personalized needs of users.

Description

Technical Field

[0001] The present invention belongs to the field of computer software, and in particular relates to a service trust evaluation method for a cross-border service network. Background Art

[0002] With the evolving global digital economy and the advancement of technologies like "cloud, big data, the Internet of Things, mobile computing, and artificial intelligence," isolated, single digital services are no longer sufficient to meet the needs of daily life and production. To address the problem of "service silos," researchers in the service computing field have proposed a series of theories and technologies for the convergence of digital services. Cross-border services aim to deeply integrate and innovate services across industries, organizations, and value chains, providing users with high-quality and high-value services. By combining or aggregating services across networks, domains, and the world, a more complex service model is created to handle big data and complex, connected businesses in the real world.

[0003] Cross-border services are characterized by high-dimensional heterogeneity, complex dynamics, and open distribution during operation. Connecting cross-border service nodes in different regions forms a cross-border service network (CSN). As a key support vehicle for cross-border services, the CSN must address three core issues: cross-border service openness and access, cross-border service integration, and cross-border service quality assurance. It must also possess high availability, scalability, and reliability.

[0004] However, as the number of CSN service accesses increases and the scale of network nodes expands, the existing CSN network architecture has gradually exposed security and trust issues. The continuous opening and integration of cross-border services has led to the blurring of service security boundaries, increased network exposure, the rapid spread of complex and covert attacks, and the difficulty of establishing trust between service nodes. Traditional network security technologies based on IP addresses and security boundaries are no longer able to meet the network security, data security, and privacy protection needs of CSN users. Therefore, research on service trust and security issues in CSN is highly relevant.

[0005] Service trust evaluation refers to the continuous assessment of service credibility during service interactions. It helps users better select trustworthy service providers, thereby improving user experience and service quality. A well-established service trust evaluation mechanism can provide users with a trustworthy service environment, enhance user trust in service providers, improve service quality, and promote the high-quality development of the digital economy. It can also foster competition among service providers, thereby stimulating innovation, improving digital productivity, and fostering a healthy competitive mechanism.

[0006] Existing research generally evaluates trust in services from two perspectives: direct evaluation based on Quality of Service (QoS) and indirect evaluation based on user feedback. Existing comprehensive service trust evaluation methods combine these two approaches, with weighted sum being the most common combination. However, existing trust evaluation methods still face the following challenges:

[0007] (1) When directly evaluating the trustworthiness of a service, only the QoS indicators of the service are considered, while the service level agreement (SLA) information of the service is ignored. To measure whether a service is trustworthy or honest, its actual QoS indicators should be compared with the SLA signed with the customer. Comparison with other services is not a reference value. For example, an AI service for image recognition will have a longer response time and latency than a weather query service, but this does not mean that the service that takes longer is untrustworthy.

[0008] (2) The SLA solutions currently adopted by the industry are all based on the average QoS over a period of time. This approach may be exploited by opportunistic SPs, which may have a certain impact on the service experience and interests of SCs.

[0009] (3) The current trust evaluation model uses the same set of standards for all users to evaluate whether the service is trustworthy, without considering the individual subjective preferences of users and ignoring the subjectivity of trust. CSN should take meeting the personalized needs of different users and improving user satisfaction as its design principles and goals. Summary of the Invention

[0010] The present invention provides a service trust evaluation method for a cross-border service network, which can ensure the fairness and impartiality of the evaluation method, meet the personalized needs of users, and effectively improve the reliability and robustness of service trust.

[0011] A service trust evaluation method for a cross-border service network, comprising:

[0012] Outputting service trust to the user so that the user can measure the service trust level of the cross-border service network based on the service trust, wherein the service trust includes direct service trust and indirect service trust;

[0013] By fusing the average value of each QoS indicator within a set time period with the service integrity, the fusion result is assigned a combined weight, and then the direct trust of the service is obtained by summing up each QoS indicator assigned a combined weight;

[0014] The service integrity is determined based on the relationship between the average value of the QoS indicator within a set time period and the corresponding SLA average value, the relationship between the maximum value of the QoS indicator and the corresponding SLA peak value, and the relationship between the ratio of the number of interaction requests corresponding to the QoS indicator value between the SLA average value and the SLA peak value to the total number of interaction requests and a set ratio threshold.

[0015] The combined weight is obtained by multiplying the subjective weight factor and the objective weight vector, and the subjective weight factor and the objective weight vector are obtained by processing the QoS indicators within the set time period using the FAHP method and the SVR method respectively;

[0016] An STLightGCN model is constructed, which introduces user social relationships on the basis of the LightGCN model. The multi-layer user embedding vectors based on the user social relationship network are aggregated with the user embedding vectors of the corresponding number of layers based on the user-service interaction network to obtain the final user embedding vector representation. The initial user embedding vector based on the user social relationship network, and the service consumer embedding vector and service embedding vector based on the user-service interaction network are used to obtain the indirect trust of the service through the STLightGCN model.

[0017] Furthermore, the service integrity of the i-th QoS indicator Hon(q i )for:

[0018]

[0019] in, is the value of the i-th QoS indicator at time t, Indicates the SLA peak value, Indicates the average SLA value. is the number of interaction requests corresponding to the QoS indicator value between SLA and SLA peak, N is the number of all interaction requests in (t0, t1), τ q is the percentage threshold.

[0020] Furthermore, the service direct trust TR D for:

[0021]

[0022] w i =μ i ·θ i

[0023] Among them, Hon(q i ) is the service integrity of the i-th QoS indicator, Set the average value of the i-th QoS indicator within the time period, β i is the importance weight of the i-th QoS indicator, w i is the combined weight assigned after the fusion of the i-th QoS indicator, μ i is the subjective weight factor of the i-th QoS indicator, θ i is the objective weight vector of the i-th QoS indicator, and n is the number of QoS indicators.

[0024] Furthermore, the FAHP method is used to process the QoS indicators within a set time period to obtain subjective weight factors, including:

[0025] The importance of each pair of QoS indicators within a set time period is evaluated. Based on the importance evaluation results, the initial pairwise comparison matrix is obtained using the FAHP relative importance scale. The consistency test of the initial indicator comparison matrix is performed to obtain the pairwise comparison matrix.

[0026] Based on the FAHP scale mapping table, each scale in the pairwise comparison matrix is mapped to a triangular fuzzy number to obtain a fuzzy pairwise comparison matrix, where the triangular fuzzy number includes the lowest value, the middle value and the highest value;

[0027] The subjective weight factor is obtained by comparing the fuzzy composite vector of each QoS indicator, where the subjective weight factor μ of the i-th QoS indicator is i for:

[0028]

[0029]

[0030] in, is the fuzzy composite vector corresponding to the i-th QoS indicator, is the fuzzy composite vector corresponding to the j-th QoS indicator, m i is the middle value of the fuzzy composite vector corresponding to the i-th QoS indicator, m j is the middle value of the fuzzy composite vector corresponding to the jth QoS indicator, l j is the lowest value in the fuzzy composite vector corresponding to the jth QoS indicator, h i is the highest value in the fuzzy composite vector corresponding to the i-th QoS indicator, and n is the number of QoS indicators.

[0031] Furthermore, the QoS indicators within the set time period are processed using the SVR method to obtain an objective weight vector, including:

[0032] The QoS indicators within a set time period are used as a training sample set, and the training sample set is labeled, wherein the labels are divided into trusted labels and untrusted labels;

[0033] The optimization objective function of SVR is determined as:

[0034]

[0035] stθ T q i +by i ≤ε+ξ i

[0036]

[0037] ξ i ≥0,

[0038] Among them, C is the penalty coefficient, which is used to control the balance between the interval and the slack variable penalty, ε is the maximum tolerance upper limit of the distance between the SVR model prediction value and the true value, that is, the size of the hard interval; ξ i and are the upper and lower slack variables of the soft interval corresponding to the i-th QoS indicator, n is the number of QoS indicators, y i is the true value of the SVR model corresponding to the i-th QoS indicator, i.e., the label, θ is the objective weight vector, q i is the i-th QoS indicator;

[0039] As shown in the following formula, the Lagrange multiplier method is used to find the optimal solution of the SVR optimization objective function:

[0040]

[0041]

[0042] in, and α i is the Lagrange multiplier.

[0043] Furthermore, we construct the STLightGCN model, including:

[0044] Obtaining a user social trust network, wherein the user social trust network includes a set of all users and a social relationship measurement matrix between users, and obtaining a social relationship measurement matrix based on social trust.

[0045] Obtaining a user-service interaction network, wherein the user-service interaction network includes interaction relationships between service consumers and services, a set of service consumers, and a set of services;

[0046] Construct the initial STLightGCN model, including the vector representation layer, message propagation layer, aggregation layer and prediction output layer, where:

[0047] The vector representation layer obtains the initial service consumer embedding vector and the initial service embedding vector based on the user-service interaction network, and also obtains the initial user embedding vector based on the user social trust network;

[0048] Through the message propagation layer, the LightGCN message propagation method is used to obtain the multi-layer service consumer embedding vector, the multi-layer service embedding vector and the multi-layer user embedding vector based on the initial service consumer embedding vector, the initial service embedding vector and the initial user embedding vector respectively;

[0049] The service consumer embedding vector of each layer is fused with the user embedding vector of the corresponding layer through the aggregation layer. The fusion results of each layer are summed to obtain the final user embedding vector representation. At the same time, the service embedding vector of each layer is summed to obtain the final service embedding vector representation.

[0050] Multiply the transpose of the final user embedding vector representation with the final service embedding vector representation to obtain the predicted service indirect trust;

[0051] The initial STLightGCN model is trained based on positive samples and negative samples using the Bayesian personalized ranking loss function to obtain the STLightGCN model. The positive samples are a set of trusted services that have interacted with the user, and the negative samples are a set of untrusted services that have interacted with the user or a set of services that have never interacted with the user. The positive samples and negative samples are respectively input into the initial STLightGCN model to obtain positive and negative prediction scores. The loss value is calculated using the Bayesian personalized ranking loss function to maximize the distance between the positive and negative prediction scores. When the loss value reaches convergence, the training of the initial STLightGCN model is completed to obtain the STLightGCN model.

[0052] Furthermore, a social relationship measurement matrix is obtained based on the social trust, wherein the social trust T ef for:

[0053]

[0054] Among them, r ef is the social trust metric between the e-th user and the f-th neighbor user in the user social trust network, r init For r ef The initial assignment of |S e | is the size of the service set of the e-th user, |S f | is the size of the service set of the fth neighbor user, d max is the maximum trust propagation distance between the e-th user and the f-th neighbor user, d ef is the shortest trust propagation distance between the e-th user and the f-th neighbor user.

[0055] Furthermore, the message propagation method of LightGCN is used to obtain multi-layer service consumer embedding vectors based on the initial service consumer embedding vector, where the service consumer embedding vector of the l+1th layer is for:

[0056]

[0057] in, is the set of all service consumers that interact with service s in the user-service interaction network, is the set of all services that interact with service consumer u in the user-service interaction network;

[0058] The message propagation method of LightGCN is used to obtain the multi-layer service consumer embedding vector based on the initial service embedding vector, where the service embedding vector of the l+1 layer is for:

[0059]

[0060] The message propagation method of LightGCN is used to obtain multi-layer user embedding vectors based on the initial user embedding vector, where the user embedding vector of the l+1 layer is for:

[0061]

[0062] in, is the set of neighbor nodes of user u based on the user’s social trust network when the social trust is greater than the social trust threshold, is the set of users whose neighbor node v has a social trust degree greater than the social trust degree threshold in the user's social trust network.

[0063] Compared with the prior art, the present invention has the following beneficial effects:

[0064] A set of service credibility evaluation index systems based on dynamic SLA for cross-border services is proposed, which solves the problems of single evaluation indicators and insufficient fairness in existing solutions. Secondly, based on this index system, this embodiment proposes a direct service trust evaluation method that integrates subjective and objective factors. The user's subjective trust preference is introduced into the direct service evaluation process through fuzzy theory, which not only ensures the fairness of the evaluation method, but also meets the user's personalized needs to a certain extent. Finally, this embodiment proposes an indirect service trust evaluation method based on collaborative filtering of graph neural networks (GNN). By constructing a user-service interaction network and a user social trust network, and improving on the LightGCN model, the STLightGCN model is proposed, which integrates the user's social trust into the service indirect trust evaluation process, effectively improving the reliability and robustness of the service indirect trust. BRIEF DESCRIPTION OF THE DRAWINGS

[0065] Figure 1 A framework diagram of a service trust evaluation method for a cross-border service network provided by a specific embodiment of the present invention;

[0066] Figure 2 This is a graph showing the final prediction performance of SVR provided in the first embodiment of the present invention;

[0067] Figure 3 This is a performance comparison chart of different regression models provided in specific embodiment 1 of the present invention;

[0068] Figure 4 A comparison chart of user satisfaction provided in specific embodiment 1 of the present invention;

[0069] Figure 5 The STLightGCN model training Loss curve diagram provided in specific embodiment 2 of the present invention;

[0070] Figure 6 A schematic diagram of a user trust social network provided in specific embodiment 2 of the present invention;

[0071] Figure 7 This is a comparison chart of model performance under different numbers of graph convolution layers provided in specific embodiment 2 of the present invention;

[0072] Figure 8 This is a performance comparison chart of different models provided in specific embodiment 2 of the present invention. DETAILED DESCRIPTION

[0073] The present invention provides a service trust evaluation method for a cross-border service network, comprising:

[0074] Output service trust to users so that users can measure the service trust level of the cross-border service network based on service trust. Service trust includes direct service trust and indirect service trust. Service trust TR is:

[0075] TR=ρ·TR D +(1-ρ)·TR R

[0076] Among them, TR D represents the direct trust of the service, which is determined by a direct trust evaluation method based on the fusion of subjective and objective factors using the Fuzzy Analytic Hierarchy Process (FAHP) and Support Vector Regression (SVR); TRR represents the indirect trust based on user feedback and user trust social networks, which is determined by a service indirect trust evaluation method based on GNN collaborative filtering; ρ∈[0,1] represents the weight factor of the direct trust of the service, which can be manually adjusted by users or decision makers according to actual conditions.

[0077] By fusing the average value of each QoS indicator and the service integrity within a set time period, the fusion result is assigned a combined weight, and then the sum of each QoS indicator assigned a combined weight is added to obtain the direct trust degree TR of the service. D for:

[0078]

[0079] w i =μ i ·θ i

[0080] Among them, Hon(q i ) is the service integrity of the i-th QoS indicator, Set the average value of the i-th QoS indicator within the time, β i is the importance weight of the i-th QoS indicator, w i is the combined weight assigned after the fusion of the i-th QoS indicator, μ i is the subjective weight factor of the i-th QoS indicator, θ i is the objective weight vector of the i-th QoS indicator, and n is the number of QoS indicators.

[0081] The service integrity Hon(·) is constructed based on the relationship between the average value of the QoS indicator within a set time period and the corresponding SLA average value, the relationship between the maximum value of the QoS indicator and the corresponding SLA peak value, and the relationship between the ratio of the number of interaction requests corresponding to the QoS indicator value between the SLA average value and the SLA peak value to the total number of interaction requests and a set ratio threshold.

[0082] In a specific embodiment, the service integrity of the i-th QoS indicator Hon(q i )for:

[0083]

[0084] in, is the value of the i-th QoS indicator at time t, Indicates the SLA peak value, Indicates the average SLA value. is the number of interaction requests corresponding to the QoS indicator value between SLA and SLA peak, N is the number of all interaction requests in (t0, t1), τ q is the percentage threshold.

[0085] The combined weight is obtained by multiplying the subjective weight factor and the objective weight vector, and the subjective weight factor and the objective weight vector are obtained by processing the QoS indicators within the set time period using the FAHP method and the SVR method respectively;

[0086] In a specific embodiment, the QoS indicators within a set time period are processed using the FAHP method to obtain a subjective weight factor, including:

[0087] The importance of each pair of QoS indicators within the set time period is evaluated. Based on the importance evaluation results, the FAHP relative importance scale table is used, as shown in Table 1, to obtain the initial pairwise comparison matrix A:

[0088]

[0089] Among them, aij represents the scale corresponding to the importance of the i-th indicator relative to the j-th indicator.

[0090] A consistency check is performed on the initial indicator comparison matrix to obtain a pairwise comparison matrix. The consistency index (CI) is calculated based on the eigenvalues of the pairwise comparison matrix to verify the legitimacy of the initial pairwise comparison matrix. For example, if a loop appears, it is illegal. If the consistency requirements are not met, the initial pairwise comparison matrix is rebuilt until the consistency check is legal.

[0091] Table 1 FAHP relative importance scale

[0092]

[0093] Based on the FAHP scale mapping table, as shown in Table 2, each scale a in the pairwise comparison matrix is ij Mapping to triangular fuzzy numbers (L, M, H) yields a fuzzy pairwise comparison matrix, where the triangular fuzzy numbers include the lowest value L, the middle value M, and the highest value H. Fuzzy pairwise comparison matrix for:

[0094]

[0095] Table 2 FAHP scale mapping table

[0096]

[0097] The sum of the triangular fuzzy numbers corresponding to each QoS indicator, that is, the ratio of the sum of the triangular fuzzy numbers of each row in the fuzzy pairwise comparison matrix to the sum of all triangular fuzzy numbers, is used as the fuzzy composite vector of each QoS indicator, where the fuzzy composite vector of the i-th QoS indicator is for:

[0098]

[0099] in, is the triangular fuzzy number in the ath row and bth column, and the subjective weight factor μ of the i-th QoS indicator is obtained by the following formula i for:

[0100]

[0101]

[0102] in, is the fuzzy composite vector corresponding to the i-th QoS indicator, is the fuzzy composite vector corresponding to the j-th QoS indicator, m i is the middle value of the fuzzy composite vector corresponding to the i-th QoS indicator, m j is the middle value of the fuzzy composite vector corresponding to the jth QoS indicator, l j is the lowest value in the fuzzy composite vector corresponding to the jth QoS indicator, h i is the highest value in the fuzzy composite vector corresponding to the i-th QoS indicator, and n is the number of QoS indicators.

[0103] In a specific embodiment, the QoS indicators within a set time period are processed using an SVR method to obtain an objective weight vector, including:

[0104] The SVR algorithm is based on the Support Vector Machine (SVM)

[14] Evolving from the SVM, the concept behind the SVM is to find a hyperplane in n-dimensional space that effectively separates different data categories. During training, SVR introduces a symmetric slack variable that increases the penalty for points that are far from the hyperplane, while not penalizing points within a threshold. This more tolerant approach further improves the model's generalization. The SVR algorithm's time complexity is independent of the input space's dimensionality, and it offers high prediction accuracy even with a small number of features. Therefore, the SVR algorithm was chosen to establish the objective weighting factor.

[0105] (1) Collect training data: The QoS indicators within a set time period are used as training sample sets, and the training sample sets are labeled. The labels are divided into trusted labels and untrusted labels. In order to ensure the consistency of the collected QoS data and expert opinions, the reliability of the training samples can be improved by taking the average value of multiple repetitions or using statistical consistency tests.

[0106] (2) Data preprocessing: First, missing values or invalid values in the collected training data can be filled with the average value of the QoS indicator of all services; second, outlier processing. Visual analysis methods such as box plots can be used to analyze the size distribution of each indicator, and data containing outliers can be removed from the training samples; finally, data normalization. Different QoS indicators have different meanings and different measurement units. Data normalization can be used to convert dimensional data into dimensionless data, which facilitates comparison and weighting between indicators of different dimensions, and can eliminate the impact of singular samples on training to a certain extent. In the experiment of this embodiment, this embodiment adopts the maximum-minimum normalization method.

[0107] (3) Optimization problem formulation: f(x) = θ T x+b is the fitting function that the regression model ultimately requires, where θ is the objective weight vector required by this embodiment, and x is the service QoS indicator vector, that is, x=(q1,q2,…,q n ), b is the offset constant. The optimization objective function of SVR is as follows: The optimization objective function of SVR is determined as:

[0108]

[0109] stθ T q i +by i ≤ε+ξ i

[0110]

[0111] ξ i ≥0,

[0112] Among them, C is the penalty coefficient, which is used to control the balance between the interval and the slack variable penalty, ε is the maximum tolerance upper limit of the distance between the SVR model prediction value and the true value, that is, the size of the hard interval; ξ i and are the upper and lower slack variables of the soft interval corresponding to the i-th QoS indicator, n is the number of QoS indicators, y i is the true value of the SVR model corresponding to the i-th QoS indicator, i.e., the label, θ is the objective weight vector, q i is the i-th QoS indicator;

[0113] As shown in the following formula, the Lagrange multiplier method is used to find the optimal solution of the SVR optimization objective function:

[0114]

[0115]

[0116] in, and α i is the Lagrange multiplier.

[0117] A specific embodiment of the present invention constructs an STLightGCN model, which introduces user social relationships on the basis of the LightGCN model, aggregates the multi-layer user embedding vectors based on the user social relationship network with the user embedding vectors of the corresponding layers based on the user-service interaction network to obtain the final user embedding vector representation, the initial user embedding vector based on the user social relationship network, and the service consumer embedding vector and service embedding vector based on the user-service interaction network are obtained through the STLightGCN model to obtain the service indirect trust TR R .

[0118] The embodiment of the present invention will establish a user trust social network in a cross-border service network. Based on the similarity of user service preferences and the trust relationship between users, the LightGCN model is improved and optimized. The STLightGCN model is proposed to integrate the user social trust network and the user-service interaction network, filter out invalid and unreliable service evaluations or feedback, and ultimately obtain a reliable indirect service trust. The specific process is as follows:

[0119] (1) Build a user social trust network. Humans are social animals. In any system, whether in the real world or online, a person will have more or less connections with others at all times. If the social attributes between people (such as friendship and trust) can be introduced into the trust model and invalid recommendations and reviews are filtered based on these attributes, the occurrence of trust attacks will be effectively reduced, thereby improving the accuracy of the trust model.

[0120] User social trust network is composed of G t =(U,T) binary definition, where U represents the set of all users, including service consumers and service providers; T represents the measurement matrix of users' social relationships (such as friendship, trust, etc.), and only the social trust (or recommendation) between users is considered here.

[0121] A social relationship measurement matrix is obtained based on the social trust, wherein the social trust T ef for:

[0122]

[0123] Among them, r ef is the social trust metric between the e-th user and the f-th neighbor user in the user social trust network, r init For r ef The initial assignment of |S e | is the size of the service set of the e-th user, |S f | is the size of the service set of the fth neighbor user, d max is the maximum trust propagation distance between the e-th user and the f-th neighbor user, d ef is the shortest trust propagation distance between the e-th user and the f-th neighbor user. The shortest trust propagation distance can be calculated using the Dijkstra shortest path algorithm.

[0124] Various situations when building trust between users: Situation 1 is under normal circumstances, when a trust relationship has been directly established between users; Situations 2 and 3 are both estimates and predictions of social trust when a trust relationship has not been directly established between users. Situation 2 makes an estimate based on the similarity of users' service preferences. Based on the basic assumption that users with the same service preferences are more likely to establish trust, it is believed that the higher the similarity of service preferences, the higher the social trust between users; Situation 3 takes into account the situation where new users have not used any services, and estimates the user trust propagation distance based on the user's social trust network, which to a certain extent solves the problem of cold start for new users.

[0125] (2) Constructing a user-service interaction network. In a cross-border service network, services can be regarded as a special commodity. Therefore, the user-service interaction network is similar to the traditional user-commodity interaction network and can be represented by a bipartite graph structure. The user-service interaction network can be represented by G s =(U, S, E) triple definition, where U represents the set of all service consumers in the cross-border service network, S represents the set of all services in the cross-border service network, and E represents the current interaction relationship between service consumers and services, which is the set of undirected edges between set U and set S.

[0126] (3) STLightGCN model construction. The LightGCN model demonstrates that the GCN model does not need to be so complex in the collaborative filtering scenario. The two standard operations of feature transformation and nonlinear activation can be removed from the GCN model, which increases the efficiency of the model while also improving the model prediction accuracy. However, the LightGCN model is still designed based on the user-item interaction network, and ignores the social relationships between users, so the model may still face some trust-related attacks. In order to solve the above problems, this embodiment makes certain improvements based on the LightGCN model, and proposes a collaborative filtering model based on the fusion of user social trust network and user-service interaction network, called the STLightGCN model. Its basic framework is as follows: Figure 1 shown.

[0127] STLightGCN consists of four layers, from bottom to top: vector representation layer, message propagation layer, aggregation layer and prediction output layer. The following will introduce them one by one:

[0128] (1) Vector representation layer: randomly initialize the embedding vector representations of users and services, denoted as d represents the dimension of the embedding vector.

[0129] (2) Message propagation layer: In order to aggregate and update the features of user nodes and service nodes in the graph, the traditional GCN model chooses to stack the convolution layers by feature transformation and nonlinear activation, so that the output of the higher-level nodes contains more information; while the specific embodiment of the present invention adopts the message propagation method of LightGCN and extracts features from the graph convolution layer by linear weighted summation. For the traditional user-service interaction network, the message propagation method of LightGCN is used to obtain multi-layer service consumer embedding vectors based on the initial service consumer embedding vector, where the service consumer embedding vector of the l+1 layer is for:

[0130]

[0131] in, is the set of all service consumers that interact with service s in the user-service interaction network, is the set of all services that interact with service consumer u in the user-service interaction network;

[0132] The message propagation method of LightGCN is used to obtain the multi-layer service consumer embedding vector based on the initial service embedding vector, where the service embedding vector of the l+1 layer is for:

[0133]

[0134] The message propagation method of LightGCN is used to obtain multi-layer user embedding vectors based on the initial user embedding vector, where the user embedding vector of the l+1 layer is for:

[0135]

[0136] in, is the set of neighbor nodes of user u based on the user’s social trust network when the social trust is greater than the social trust threshold, is the set of users whose neighbor node v has a social trust degree greater than the social trust degree threshold in the user's social trust network.

[0137] (3) Aggregation layer: The aggregation layer aggregates the vector representations based on the user social trust network and the vector representations based on the user-service interaction network obtained by the message delivery layer. The aggregated vectors have stronger semantic expression capabilities. The aggregation process is shown in the following formula:

[0138]

[0139] in, represents the aggregation vector of the first layer; sum(·) represents the sum of the corresponding elements of two vectors; norm(·) represents the norm of the vector. The norm is a function that can represent the length of the vector in the vector space. Common norm functions include L1 norm and L2 norm. In the calculation process, this embodiment makes Through the aggregation layer, we can finally get the vector representation of user u and service s, which are and The calculation is as follows:

[0140]

[0141]

[0142] Among them, α lis an adjustable parameter, according to the experiments of He et al.

[15] ,when When , the model prediction performance is the best, and L is the number of graph convolutional layers.

[0143] (4) Prediction output layer: As the last layer of STLightGCN, it predicts the score of users and services based on the vector representation of the user and service generated by integrating the social trust network and the user-service interaction network. The size of the prediction score determines the probability that the cross-border service network will recommend service s to user u, that is, the user's indirect trust in the service. The calculation method is shown in the following formula:

[0144]

[0145] During the training process, this embodiment uses the Bayesian Personalized Ranking (BPR) loss function to train the model. BPR is a personalized ranking recommendation algorithm whose core idea is to make the distance between positive and negative samples in the training sample as large as possible. Positive samples are selected from services that the user has interacted with, while negative samples are selected from services that the user has not interacted with. Generally, the number of negative samples should not be too large. In this paper, the number of negative samples is 8. The BPR loss function is defined as follows:

[0146]

[0147] Among them, Θ is the model parameter, mainly including the initialization vector in the vector representation layer, such as λ is a non-negative hyperparameter, is a positive prediction score, is the negative prediction score, |U| is the user, the purpose is to avoid overfitting, |U| is the number of users, The set of services that are positive samples.

[0148] This patent studies the trustworthiness of cross-border services in cross-border service networks and clarifies the definition of service trust in cross-border service networks, including direct trust based on QoS and indirect trust based on user feedback. It also models and measures the non-functional and functional attributes of cross-border services. The specific benefits of this patent include:

[0149] First, this patent proposes a service credibility evaluation index system based on dynamic SLA for cross-border services. Compared with the current mainstream static SLA, dynamic SLA takes into account the fluctuations between service consumers and services during the exchange process, which not only provides a certain fault tolerance space for service providers, but also avoids the unfairness brought by the mean effect of static SLA for service consumers.

[0150] Secondly, this patent proposes a direct service trust evaluation method based on the subjective and objective fusion of FAHP and SVR. This method not only ensures the objective authenticity of service quality, but also takes into account the subjective needs of users by introducing user trust preferences, thereby improving user satisfaction to a certain extent. The comparative experimental results show that the prediction performance of the SVR model is good, and the predicted Kendall coefficient τ = 0.80289 on the QWS dataset is far superior to other regression models. This shows that SVR is more suitable for scenarios with small feature dimensions and few training samples, and is more suitable for the determination process of objective weights of direct service trust; in addition, this embodiment compares user satisfaction before and after the introduction of subjective weights determined based on the FAHP method, and finds that after the introduction of user subjective trust preferences, user satisfaction is improved by an average of 5%-8%.

[0151] Finally, this patent proposes a service indirect trust evaluation method based on GNN collaborative filtering. By establishing a user-service interaction network and a user trust social network, and leveraging the powerful performance of the STLightGCN model, untrustworthy user reviews are filtered out, thereby obtaining an accurate and unbiased service indirect trust. Comparative experiments have shown that the performance of this model is far superior to the MF and TrustSVD models. In addition, this embodiment trains and tests the proposed model on the public dataset Epinions and finds that when the number of graph convolutional layers is equal to 3, the model prediction performance is optimal.

[0152] Example 1

[0153] The goal of this example is to verify the effectiveness of the proposed direct service trust evaluation method based on the fusion of FAHP and SVR using the QWS dataset. The experiment can be divided into three steps: data preprocessing, verifying the accuracy of the SVR model, and verifying the benefits of subjective trust preferences.

[0154] (1) Data preprocessing

[0155] The direct trust evaluation model proposed in this paper includes two major components: traditional QoS data and service integrity based on dynamic SLAs. However, the QWS dataset only contains service QoS data and no data related to service integrity. To address this issue, this embodiment clusters the services in the QWS dataset based on their semantic information (such as service name and service description information) and estimates service integrity based on the distribution of service QoS data within each category. The specific process is as follows:

[0156] (a) The semantic similarity between different services is calculated based on the pre-trained Sentence-BERT model, and the services are clustered based on their semantic similarity. After clustering, the 2872 services in the QWS dataset can be divided into 251 categories.

[0157] (b) Estimate the service integrity based on the mean and variance of the QoS indicators of each service category. If the mean value of a QoS indicator q is μ q , with variance σ q ,This embodiment assumes that the distribution of service QoS indicators is normally distributed, and the ,service integrity is estimated using the following formula.

[0158]

[0159] (2) Verify the accuracy of the SVR model:

[0160] After data preprocessing, we obtain training data that meets the model input requirements. Since only the QWS 1.0 dataset contains expert ratings and scores on services, this example uses only QWS 1.0 as a training sample. The training set and test set are randomly divided in an 8:2 ratio, and the optimal solution is found according to the objective weight determination method described above.

[0161] The kernel function for the SVR model uses a linear kernel function with a penalty coefficient of C = 5. In this scenario, the SVR model's predicted scores are not required to be identical to those in the training set; instead, the model is required to rank services relatively accurately. Therefore, this example uses the Kendall rank correlation coefficient τ to evaluate the SVR model's prediction accuracy. The Kendall rank correlation coefficient can effectively measure the ordinal correlation between two sets of data series.

[0162] Figure 2 This demonstrates the final prediction performance of SVR. Although at some points the SVR model's predicted scores differ significantly from the true values, the overall trend of the prediction curve and the relative ranking of each point are close to the real data. The Kendall coefficient, τ, is 0.80289, indicating a significant correlation between the model's predicted values and the actual values.

[0163] Figure 3A comparison chart of the prediction performance of the SVR model and other regression models is shown. As can be seen from the chart, the linear regression model has the worst prediction effect due to its simplicity; the logistic regression model has a slightly better prediction effect than the linear regression model; the random forest regression model is the most stable of all models, but its prediction performance is slightly inferior to the multilayer perceptron (MLP) regression model; and the SVR model has the best average prediction performance among these models, indicating that the SVR is more suitable for scenarios with small feature dimensions and few training samples, and is also more suitable for the process of determining the objective weight of direct service trust.

[0164] This example verifies the beneficial effect of introducing subjective trust by comparing the changes in user satisfaction before and after the introduction of subjective trust preferences. The specific process is as follows:

[0165] (a) Determine subjective weights using the FAHP method. This example assumes that the user prioritizes each QoS indicator in the order: Ava > Rel > Lat, and that service integrity and QoS have the same priority for the same indicator. Table 1 provides a pairwise comparison matrix A for each indicator.

[0166]

[0167] According to the process of the FAHP method introduced in this patent, this embodiment can finally obtain the user's subjective trust weight vector as: (0.18468, 0.16533, 0.14999, 0.18468, 0.16533, 0.14999).

[0168] (b) The ranking results of the user's subjective trust weight are used as a benchmark and compared with the ranking results of the objective weight and the weight after the subjective and objective fusion. Here, this embodiment still uses the Kendall coefficient to measure user satisfaction. Figure 4 This example shows the changes in user satisfaction with 10 weather-related services in the QWS dataset after introducing user subjective trust preference vectors. Experimental verification shows that the introduction of subjective trust preference vectors increases user satisfaction by an average of 5%-8%.

[0169] Example 2

[0170] The goal of this experiment is to train the STLightGCN model based on the user-service interaction network and user social network proposed in this chapter, compare and verify the model effect, and explore the impact of different numbers of graph convolutional layers l on the model prediction results.

[0171] Due to the lack of service-related datasets, this embodiment uses the public dataset Epinions for training. This dataset originates from the consumer review website Epinions, where users can review a product and can choose to establish trust relationships with other users. The website will sort and filter product reviews based on the user's trust social network. This dataset consists of two parts: (1) the trust social network formed between different users; (2) the user's rating of the product, which has a total of 5 levels. Here, this embodiment uses the products in Epinions to analogize the services in the cross-border service network. The dataset contains 16,809 users, 47,856 products, 374,151 valid reviews, and 380,303 sets of trust relationships.

[0172] The algorithm of this experiment is implemented in Python 3.6 using the PyTorch deep learning framework and the DGL graph neural network framework. The operating system is CentOS 7, the graphics processor is Nvidia Tesla P40, and the Cuda version is 10.2.

[0173] This experiment uses the hit ratio (HR) and normalized discounted cumulative gain (NDCG), which are commonly used in recommendation systems, to evaluate the model performance.

[0174] During the experiment, this example randomly partitioned the training, validation, and test sets in a ratio of 8:1:1. The model's graph embedding layer size was 32, and training was performed with l = 2, 3, 4, and 5, respectively. Each batch size was 1024, and training was repeated for 20 epochs. Early termination was allowed, and the BPR loss function was used. A grid search strategy was used for other model training parameters (such as the learning rate), and the model parameters with the highest NDCG were retained.

[0175] The loss curve during training is as follows Figure 5 As shown in the figure, at the beginning of training, the larger the number of graph convolutional layers, l, the faster the model converges, indicating that the output of the message propagation layer contains more information from the graph. As training progresses, all loss curves tend to flatten out. Notably, the loss curve for l = 5 becomes significantly unstable in the later stages of training.

[0176] Figure 6This figure shows a trust social network constructed based on the trust relationships between the top 50 users in the Epinions dataset. The vertices in the graph represent users, the directed edges represent trust relationships, the head vertex represents the trust giver, and the tail vertex represents the trustee. The figure clearly shows that the number of directed edges to the right-hand vertices (such as nodes 1 and 23) is significantly higher than to the left-hand vertices, indicating that the right-hand vertices contain more information and are more valuable for reference.

[0177] Figure 7 The performance of the model is shown with different numbers of graph convolutional layers, l. This example shows that a larger value of l does not necessarily improve the model performance. The model achieves optimal results in both HR@10 and NDCG@10 when l = 3. This experimental result can also be verified in daily life. Taking trust relationships as an example, this example will trust one's friends and may also choose to trust friends of friends, but the trust in friends who are farther away is not as high. When l is further increased, the model performance decreases. Combined with the loss curve during training, it can be analyzed that increasing the number of graph convolutional layers does not necessarily bring a significant improvement in model performance. It is also easy to cause overfitting, resulting in a decrease in model performance.

[0178] In addition, the STLightGCN model is combined with the Matrix Factorization (MF) model

[20] and TrustSVD model

[21] A comparison was made, from Figure 8 It can be clearly seen that the STLightGCN model outperforms the other two models in both HR and NDCG evaluation indicators. Both MF and TrustSVD models are CF algorithms based on matrix decomposition. Their main purpose is to solve the problem that the CF algorithm has weak ability to handle sparse matrices and difficult to maintain similarity matrices. However, these two methods do not take into account the transitivity of social relationships between users. STLightGCN completes the feature transfer of such social relationships through linear superposition of graph convolution layers, and avoids the computational complexity brought by feature transformation and nonlinear activation. However, as the network scale expands, the STLightGCN model may also face the problems of data sparsity and difficulty in maintaining graph matrices.

Claims

1. A service trust evaluation method for a cross-border service network, characterized in that: include: Outputting service trust to the user so that the user can measure the service trust level of the cross-border service network based on the service trust, wherein the service trust includes direct service trust and indirect service trust; The average value of each QoS indicator within a set time period and the service integrity are integrated, the integrated result is assigned a combined weight, and then the sum of each QoS indicator assigned a combined weight is added to obtain the direct trust of the service; The service integrity is determined based on the relationship between the average value of the QoS indicator within a set time period and the corresponding SLA average value, the relationship between the maximum value of the QoS indicator and the corresponding SLA peak value, and the relationship between the ratio of the number of interaction requests corresponding to the QoS indicator value between the SLA average value and the SLA peak value to the total number of interaction requests and a set ratio threshold. The combined weight is obtained by multiplying the subjective weight factor and the objective weight vector, and the subjective weight factor and the objective weight vector are obtained by processing the QoS indicators within the set time period using the FAHP method and the SVR method respectively; An STLightGCN model is constructed, wherein the STLightGCN model introduces user social relationships on the basis of the LightGCN model, wherein the multi-layer user embedding vectors based on the user social relationship network are aggregated with the user embedding vectors of the corresponding number of layers based on the user-service interaction network to obtain the final user embedding vector representation, the user embedding vector based on the user social relationship network, and the service consumer embedding vector and service embedding vector based on the user-service interaction network are used to obtain the indirect trust of the service through the STLightGCN model.

2. The service trust evaluation method for a cross-border service network according to claim 1, characterized in that: The service integrity of the i-th QoS indicator Hon(q i )for: in, is the value of the i-th QoS indicator at time t, Indicates the SLA peak value, Indicates the average SLA value. is the number of interaction requests corresponding to the QoS indicator value between SLA and SLA peak, N is the number of all interaction requests in (t0, t1), τ q is the percentage threshold.

3. The service trust evaluation method for a cross-border service network according to claim 1, characterized in that: The direct trust degree TR of the service D for: w i =μ i ·i i Among them, Hon(q i ) is the service integrity of the i-th QoS indicator, is the average value of the i-th QoS indicator in the set time period, β i is the importance weight of the i-th QoS indicator, w i is the combined weight assigned after the fusion of the i-th QoS indicator, μ i is the subjective weight factor of the i-th QoS indicator, θ i is the objective weight vector of the i-th QoS indicator, and n is the number of QoS indicators.

4. The service trust evaluation method for a cross-border service network according to claim 1, characterized in that: The FAHP method is used to process the QoS indicators within a set time period to obtain subjective weight factors, including: The importance of each pair of QoS indicators within a set time period is evaluated. Based on the importance evaluation results, the initial pairwise comparison matrix is obtained using the FAHP relative importance scale. The consistency test of the initial indicator comparison matrix is performed to obtain the pairwise comparison matrix. Based on the FAHP scale mapping table, each scale in the pairwise comparison matrix is mapped to a triangular fuzzy number to obtain a fuzzy pairwise comparison matrix, where the triangular fuzzy number includes the lowest value, the middle value and the highest value; The sum of the triangular fuzzy numbers corresponding to each QoS indicator, that is, the ratio of the sum of the triangular fuzzy numbers of each row in the fuzzy pairwise comparison matrix to the sum of all triangular fuzzy numbers, is used as the fuzzy composite vector of each QoS indicator; The subjective weight factor is obtained by comparing the fuzzy composite vector of each QoS indicator, where the subjective weight factor μ of the i-th QoS indicator is i for: in, is the fuzzy composite vector corresponding to the i-th QoS indicator, is the fuzzy composite vector corresponding to the j-th QoS indicator, m i is the middle value of the fuzzy composite vector corresponding to the i-th QoS indicator, m j is the middle value of the fuzzy composite vector corresponding to the jth QoS indicator, l j is the lowest value in the fuzzy composite vector corresponding to the jth QoS indicator, h i is the highest value in the fuzzy composite vector corresponding to the i-th QoS indicator, and n is the number of QoS indicators.

5. The service trust evaluation method for a cross-border service network according to claim 1, characterized in that: The SVR method is used to process the QoS indicators within a set time period to obtain an objective weight vector, including: The QoS indicators within a set time period are used as a training sample set, and the training sample set is labeled, wherein the labels are divided into trusted labels and untrusted labels; The optimization objective function of SVR is determined as: stθ T q i +by i ≤e+ξ i Among them, C is the penalty coefficient, which is used to control the balance between the interval and the slack variable penalty, ε is the maximum tolerance upper limit of the distance between the SVR model prediction value and the true value, that is, the size of the hard interval; ξ i and are the upper and lower slack variables of the soft interval corresponding to the i-th QoS indicator, n is the number of QoS indicators, y i is the true value of the SVR model corresponding to the i-th QoS indicator, i.e., the label, θ is the objective weight vector, q i is the i-th QoS indicator; As shown in the following formula, the Lagrange multiplier method is used to find the optimal solution of the SVR optimization objective function: in, and α i is the Lagrange multiplier.

6. The service trust evaluation method for a cross-border service network according to claim 1, characterized in that: Build the STLightGCN model, including: Obtaining a user social trust network, wherein the user social trust network includes a set of all users and a social relationship measurement matrix between users, and obtaining a social relationship measurement matrix based on social trust. Obtaining a user-service interaction network, wherein the user-service interaction network includes interaction relationships between service consumers and services, a set of service consumers, and a set of services; Construct the initial STLightGCN model, including the vector representation layer, message propagation layer, aggregation layer and prediction output layer, where: The vector representation layer obtains the initial service consumer embedding vector and the initial service embedding vector based on the user-service interaction network, and also obtains the initial user embedding vector based on the user social trust network; Through the message propagation layer, the LightGCN message propagation method is used to obtain the multi-layer service consumer embedding vector, the multi-layer service embedding vector and the multi-layer user embedding vector based on the initial service consumer embedding vector, the initial service embedding vector and the initial user embedding vector respectively; The service consumer embedding vector of each layer is fused with the user embedding vector of the corresponding layer through the aggregation layer. The fusion results of each layer are summed to obtain the final user embedding vector representation. At the same time, the service embedding vector of each layer is summed to obtain the final service embedding vector representation. Multiply the transpose of the final user embedding vector representation with the final service embedding vector representation to obtain the predicted service indirect trust; The initial STLightGCN model is trained using the Bayesian personalized ranking loss function based on positive and negative samples to obtain the STLightGCN model.

7. The service trust evaluation method for a cross-border service network according to claim 6, characterized in that: A social relationship measurement matrix is obtained based on the social trust, wherein the social trust T ef for: Among them, r ef is the social trust metric between the e-th user and the f-th neighbor user in the user social trust network, r init For r ef The initial assignment of |S e | is the size of the service set of the e-th user, |S f | is the size of the service set of the fth neighbor user, d max is the maximum trust propagation distance between the e-th user and the f-th neighbor user, d ef is the shortest trust propagation distance between the e-th user and the f-th neighbor user.

8. The service trust evaluation method for a cross-border service network according to claim 6, characterized in that: The message propagation method of LightGCN is used to obtain the multi-layer service consumer embedding vector based on the initial service consumer embedding vector, where the service consumer embedding vector of the l+1 layer is for: in, is the set of all service consumers that interact with service s in the user-service interaction network, is the set of all services that interact with service consumer u in the user-service interaction network; The message propagation method of LightGCN is used to obtain the multi-layer service consumer embedding vector based on the initial service embedding vector, where the service embedding vector of the l+1 layer is for: The message propagation method of LightGCN is used to obtain multi-layer user embedding vectors based on the initial user embedding vector, where the user embedding vector of the l+1 layer is for: in, is the set of neighbor nodes of user u based on the user’s social trust network when the social trust is greater than the social trust threshold, is the set of users whose neighbor node v has a social trust degree greater than the social trust degree threshold in the user's social trust network.

Images