Social network semantic alignment trust evaluation system based on graph convolutional network

Abstract

The invention relates to the technical field of social network trust evaluation, in particular to a social network semantic alignment trust evaluation system based on a graph convolutional network. According to the system, initial features of user nodes and weighted directed edges are constructed through a trust information coding module; a trust semantic calibration module generates calibration factors using trust and trusted neighborhoods, adjusts ratings to align heterogeneous scoring semantics; fusing individual statistical characteristics and edge graph convolution through a trust semantic interpretation module, and depicting trust role expression; the trust semantic gain module is combined with edge graph convolution and global similarity sorting, sparse semantic information is enhanced, and final trust features of double-role fusion are generated; and finally, predicting a missing trust relationship by a trust evaluation module. According to the system, semantic deviation caused by subjectivity can be effectively relieved, and the trust prediction accuracy and robustness are improved.

Description

A Semantic Alignment Trust Evaluation System for Social Networks Based on Graph Convolutional Networks Technical Field

[0001] This invention relates to the field of social network trust evaluation technology, and specifically to a social network semantic alignment trust evaluation system based on graph convolutional networks. Background Technology

[0002] Online social networks (OSNs) have become core platforms for information acquisition, social interaction, and resource sharing. Among these, directed trust relationships between users serve as a key structure, widely used in scenarios such as identity authentication, capability endorsement, content recommendation, digital signatures, and transaction security. Trust relationships extend from the "truster" to the "trusted," and their strength is often quantified by trust ratings (scalar values)—the higher the value, the stronger the trust.

[0003] However, in real-world systems, due to limitations in privacy protection, data collection costs, and platform visibility, only a portion of trust relationships are known, while a large number of potential relationships remain unknown. Therefore, accurately inferring unknown trust relationships and their ratings based on limited known trust information becomes a core challenge for improving recommendation accuracy, refining user profiles, and enhancing system security. This task faces three fundamental problems stemming from the subjectivity of user trust: Inconsistent semantic scales: Different users interpret the same rating value significantly differently. For example, a user accustomed to giving high scores might give only a "3" rating indicating general trust, while a user accustomed to giving low scores might give a "3" rating indicating high trust. If such individualized rating biases are not calibrated, they will be amplified during trust propagation, leading to distorted assessments.

[0004] Semantic expression imbalance: Users exhibit significant heterogeneity in trust expression—some users have stable behavior and clear preferences, providing clear semantic signals; while others have sparse interactions, large rating fluctuations, and no obvious bias, resulting in ambiguous trust semantics, making it difficult for the model to effectively characterize their personalized patterns.

[0005] Semantic information scarcity: Social networks are inherently sparse, and most users lack sufficient direct interaction, resulting in a severe shortage of explicit semantic information available for learning. Especially for low-activity nodes, their local neighborhood information is insufficient to support reliable reasoning, leading to a dual dilemma of "data sparsity + semantic ambiguity".

[0006] To address these challenges, the academic community has proposed various trust assessment methods. Early works such as Matri use matrix factorization to model multipath propagation, but it is difficult to capture nonlinear dependencies; OpinionWalk introduces a subjective logic model, but it relies on manually set discount and merging rules, which limits its generalization ability; NeuralWalk, although it automatically learns trust rules through neural networks, is difficult to scale to large-scale networks due to its high computational complexity.

[0007] In recent years, Graph Neural Networks (GNNs) have been introduced into this field due to their natural fit with the propagation (path-based transmission) and aggregation (multi-path fusion) of trust. For example, Guardian was the first to use directed graph convolution to model the asymmetry of trust relationships, pointing out that the subjectivity of users as trust subjects is a key factor affecting the accuracy of assessment. Subsequent research has further improved personalized modeling capabilities by focusing on empirical subjectivity (such as TrustGNN, GBTrust, TrustGuard) and preference subjectivity (such as GATrust, KGTrust) and modeling path weights, multi-dimensional features, or temporal dynamics through attention mechanisms.

[0008] However, existing methods generally focus on behavioral performance during the trust propagation process, failing to address the semantic output of trust rating itself and systematically resolve the semantic bias caused by subjectivity. Specifically, they have not yet effectively achieved: adaptive calibration of heterogeneous rating scales; accurate interpretation of imbalanced expression patterns; and structured gains for sparse semantic information.

[0009] Therefore, there is an urgent need for a new trust assessment framework that can align, decouple, and enhance user trust expressions from the semantic source, so as to systematically overcome the assessment inaccuracy caused by subjectivity and improve the accuracy and robustness of the model in real sparse and heterogeneous social environments. Summary of the Invention

[0010] To address the technical problem of low accuracy in semantic alignment trust evaluation in social networks, this invention aims to provide a semantic alignment trust evaluation system for social networks based on graph convolutional networks. The specific technical solution is as follows: One embodiment of this invention provides a semantic alignment trust evaluation system for social networks based on graph convolutional networks. The online social network includes multiple user nodes and some known trust relationships. Each trust relationship is represented as a weighted directed edge from the trustor to the trusted party, with the weight representing the trust rating. The system includes: a trust information encoding module, used to encode the graph structure information of the online social network into initial features of each user node, and to encode the ratings of known trust relationships into initial rating features of the corresponding directed edges; and a trust semantic calibration module, used to generate calibration factors for each user node based on the initial features, the user node's trust neighborhood, and the trusted neighborhood, and to use the calibration factors to adjust the edges. The initial rating features are adjusted to obtain calibrated rating semantic features; the trust semantic interpretation module is used to combine the user's individual rating statistical features with the neighborhood information aggregated by the out-edge graph convolution to generate trustor role features; the trustor role features are used to at least characterize the user's personalized trust expression features as a trustor; the trust semantic gain module is used to generate trusted party role features based on the calibrated rating semantic features, and combine the trustor role features to form a dual-role trust feature; the trusted party role features are used to at least characterize the user's personalized trust expression features as a trusted party; the dual-role trust features are subjected to convolutional enhancement processing based on global semantic ranking to generate the final trust features of user nodes; the trust evaluation module is used to, for any pair of users to be predicted, concatenate the final trust features and input them into a multilayer perceptron, outputting the predicted trust rating to complete the missing trust relationships in the online social network.

[0011] Further, obtaining the calibration factor includes: obtaining the out-degree neighbor set of the trustee node and its initial features and initial rating features, and calculating the rating strictness vector through out-edge directed graph convolution and nonlinear transformation; the trustee node is a user node within the trust neighborhood, and the trust neighborhood consists of the user's trusted neighbors and the trust relationships pointing to the neighbors; obtaining the in-degree neighbor set of the trusted node and its initial features and initial rating features, and calculating the credibility vector through in-edge directed graph convolution and nonlinear transformation; the trusted node is a user node within the trusted neighborhood, and the trusted neighborhood consists of the trusted user's neighbors and the trust relationships pointing to the user; concatenating the rating strictness vector and the credibility vector, and inputting them into a multilayer perceptron to generate the calibration factor.

[0012] Further, obtaining the trustee role features includes: calculating the mean and rating fluctuation of all out-degree neighbor ratings of the user node; combining the mean rating, rating fluctuation, and node out-degree; and mapping them through a multilayer perceptron to form individual semantic expression features; obtaining the social semantic expression features of the user node as a trustee by performing edge graph convolution aggregation on the user node and aggregating neighbor information; and using a gated feature fusion mechanism to determine the weights of the social semantic expression features and the individual semantic expression features respectively, and performing weighted fusion to form the trustee role features of the user node.

[0013] Further, obtaining the dual-role trust features includes: performing an in-edge graph convolution operation on the calibrated rating semantic features of the user node to generate a trusted party role feature; and concatenating the trusted party role feature with the trustor role feature to obtain the dual-role trust features.

[0014] Further, the dual-role trust features are subjected to convolutional enhancement processing based on global semantic ranking to generate the final trust features of user nodes. This includes: linearly mapping the dual-role trust features to obtain the semantic ranking score of each node, and globally ranking all nodes according to the semantic ranking score to obtain a ranked trust feature vector sequence and a corresponding global ranking index; arranging the feature vectors of each node in order based on the global ranking index to construct a global node representation matrix; applying a one-dimensional convolution kernel to perform a sliding window operation on the ranking dimension of the global node representation matrix to extract and aggregate node features with similar trust patterns; using the inverse mapping of the global ranking index, restoring the features output by the convolution operation to the original node order to obtain trust semantic enhancement features; and performing a residual connection between the trust semantic enhancement features of the same user node and the dual-role trust features to generate the final trust features.

[0015] This invention offers the following advantages: A social network semantic alignment trust evaluation system based on graph convolutional networks systematically solves the core problems of existing methods in trust rating modeling, such as semantic bias, expression imbalance, and information sparsity caused by user subjectivity, through the synergistic effect of its modules. Specifically, the trust information encoding module encodes the graph structure and trust rating as initial features of nodes and edges, respectively, laying a structured data foundation for subsequent semantic modeling. The trust semantic calibration module generates personalized calibration factors by fusing the initial features of user nodes and their trust / trusted neighborhood context, adaptively adjusting the original ratings and effectively mitigating semantic distortion caused by heterogeneous rating scales among different users, achieving source alignment of trust rating semantics. The trust semantic interpretation module combines individual historical rating statistical features with neighborhood behavior information aggregated by out-edge graph convolution to accurately characterize the personalized expression patterns of users as "trusters," thereby... By decoupling and explicitly modeling users' subjective expression styles, this system overcomes the limitations of traditional methods that treat users as homogeneous entities. The trust semantic gain module further captures the semantic meaning of users as "trusted parties" and constructs dual-role trust features. Based on this, it introduces global ranking and one-dimensional convolution mechanisms to mine group structures with similar trust patterns from users across the entire network. This semantic enhancement of individual features under sparse or noisy interference significantly improves feature representation capabilities and robustness in low-interaction-density scenarios. Finally, the trust evaluation module concatenates the final trust features of target user pairs and inputs them into a multilayer perceptron to achieve high-precision prediction of missing trust relationships. Overall, this system not only achieves calibration, interpretation, and amplification of trust expressions at the semantic level but also significantly improves the accuracy, interpretability, and generalization ability of trust prediction in real, sparse, and heterogeneous social environments, providing reliable technical support for applications such as social recommendation, fake account identification, and trusted community construction. Attached Figure Description

[0016] To more clearly illustrate the technical solutions and advantages in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0017] Figure 1 is a module structure diagram of a social network semantic alignment trust evaluation system based on graph convolutional networks according to an embodiment of the present invention; Figure 2 is a schematic diagram of a social trust evaluation problem scenario according to an embodiment of the present invention; Figure 3 is a schematic diagram of the trust semantic calibration process according to an embodiment of the present invention; Figure 4 is a schematic diagram of the trust semantic interpretation process according to an embodiment of the present invention; Figure 5 is a schematic diagram of the trust semantic gain process according to an embodiment of the present invention; Figure 6 is a distribution diagram of trust rating bias of the Advogato dataset according to an embodiment of the present invention; Figure 7 is a distribution diagram of trust rating bias of the PGP dataset according to an embodiment of the present invention; Figure 8 is a scatter plot of the mean and variance of trust rating of the Advogato dataset according to an embodiment of the present invention; Figure 9 is a scatter plot of the mean and variance of trust rating of the PGP dataset according to an embodiment of the present invention; Figure 10 is a comparison diagram of the rating bias distribution of the Advogato dataset according to an embodiment of the present invention; Figure 11 is a comparison diagram of the rating bias distribution of the PGP dataset according to an embodiment of the present invention; Figure 12 is a trend curve of the model's various performance indicators according to an embodiment of the present invention; Figure 13 is a performance variation curve of SA-GCN under different convolutional kernel sizes according to an embodiment of the present invention; Figure 14 is a framework diagram of an SA-GCN model for implementing social network trust evaluation according to an embodiment of the present invention. Detailed Implementation

[0018] To further illustrate the technical means and effects adopted by the present invention to achieve its intended purpose, the specific implementation methods, structures, features, and effects of the technical solution proposed according to the present invention are described in detail below with reference to the accompanying drawings and preferred embodiments. In the following description, different "one embodiment" or "another embodiment" do not necessarily refer to the same embodiment. Furthermore, specific features, structures, or characteristics in one or more embodiments can be combined in any suitable form.

[0019] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0020] Online social networks contain a large number of users who can establish trust relationships. Within these relationships, each user may play a dual role: both the trustor and the trusted party. The party initiating the trust is called the "truster," and the party receiving the trust is called the "trustee." Trust relationships have a clear direction, always pointing from the trustor (the one issuing the trust) to the trustee (the one receiving the trust). Furthermore, the strength of a trust relationship is quantified by a "trust rating," with higher ratings indicating stronger trust and vice versa. It is worth noting that only a portion of trust relationships within social networks are known; a large number of unknown trust relationships remain to be discovered.

[0021] When there is no direct trust relationship between two users, it is necessary to predict the trust relationship and trust rating between them. Existing trust ratings are essentially subjective expressions by users, and the semantic differences between different users are significant. These differences can be summarized in three aspects: inconsistent semantic scales, unbalanced semantic expressions, and insufficient semantic information. Therefore, the core task of trust assessment is to extract stable and unified semantic expressions from the significantly different trust behaviors of the trustor and the trusted party in a social environment characterized by "prominent subjectivity and significant data sparsity," ultimately achieving accurate and robust predictions of unknown trust relationships.

[0022] Accordingly, one embodiment of the present invention proposes a semantic alignment trust evaluation system for social networks based on graph convolutional networks. The online social network includes multiple user nodes and some known trust relationships. Each trust relationship is represented as a weighted directed edge from the trustor to the trusted party, with the weight representing the trust rating. This semantic alignment trust evaluation system for social networks can be applied to trust evaluation in online social networks, transaction decisions in e-commerce platforms, content filtering in knowledge-sharing communities, recommendation systems, and the prevention and control of misinformation, among other application scenarios.

[0023] The module structure diagram of the system is shown in Figure 1, including: a trust information encoding module 101, which is used to encode the graph structure information of the online social network into the initial features of each user node, and to encode the rating of known trust relationships into the initial rating features of the corresponding directed edges.

[0024] Here, online social networks are composed of different user nodes.

[0025] In one embodiment of the present invention, the graph structure construction process based on an online social network includes: abstracting each user as a node in the graph, and modeling the relationships of following, friendship, or interaction between users as directed edges, thereby forming a directed graph. Where V represents the set of user nodes, Let R represent the set of directed edges, i.e., the set of known direct trust relationships, and let R represent the set of observed trust rating records. Represents an unknown set of trust relationships. Let represent an unknown set of trust ratings. If user u trusts user v, that is, user u is the trustor and user v is the trusted party, then a direct trust relationship exists between the two users, denoted as . And has a trust rating If the trust state of user u towards user v is unknown, then the direct trust relationship between the two users is denoted as... The corresponding unknown trust rating is recorded as It is worth noting that the trust assessment problem in social networks is based on known direct trust information to predict unknown direct trust relationships and trust ratings.

[0026] Furthermore, encoding the graph structure information into initial node features specifically includes: generating an initial feature vector for each node based on its local topological attributes, including but not limited to at least one of in-degree, out-degree, number of neighboring nodes, clustering coefficient, or PageRank value; the initial feature vector is used to characterize the user's structural role in the social network (such as a central user, a peripheral user, or a bridging user).

[0027] Meanwhile, encoding the trust rating between users as an initial rating feature of weighted directed edges specifically includes: if user u has an explicit or implicit trust rating for user v. (For example, numerical trust levels obtained by aggregating likes, comments, and shares, or ratings given by users), then in directed edges Assign weights The weight is used as the initial rating feature of the edge; if there is no trust relationship, the corresponding edge does not exist or the weight is set to zero.

[0028] Therefore, the constructed graph not only contains the structural embedding information of the nodes, but also integrates the directional and quantitative trust semantics between users, providing a multi-dimensional feature foundation for subsequent trust propagation, role recognition, or recommendation decisions.

[0029] The trust semantic calibration module 102 is used to generate calibration factors for any user node based on the trusted and trusted neighborhoods of the user node, in combination with the initial features, and to adjust the initial rating features of the edges using the calibration factors to obtain the calibrated rating semantic features.

[0030] Figure 2 illustrates a scenario for social trust assessment. In Figure 2, user b typically tends to give neighbors higher trust ratings, such as 3 or 4 points; while user f rates user a 1 point and user d 4 points. Compared to user b, user f's ratings are more dispersed and their trust inclination is less clear. Even if both user b and user f rate user d 4 points, the actual trust strength represented by that rating differs within their respective rating contexts. Therefore, trust ratings implicitly contain individualized trust semantics, and these semantic differences are easily amplified in trust propagation, thus affecting the accuracy of the overall assessment.

[0031] To address the aforementioned deficiency of inconsistent semantic scales in trust rating, this invention designs a trust semantic calibration module to align the rating features of different users at the semantic level. A trustor's ratings of different objects reflect their subjective rating benchmark; the performance of a trusted party under multiple ratings reflects their overall trustworthiness. Therefore, the model aggregates trust information in two directions based on directed trust relationships between users, focusing on both the user's trust neighborhood and the trusted party's neighborhood: in the trust neighborhood, out-edge graph convolution integrates the same user's trust information for different trusted parties to capture the user's rating strictness as a trustor; in the trusted neighborhood, in-edge graph convolution integrates the trust information of different trustors for the same user to learn the user's trustworthiness as a trusted party. Subsequently, an interaction mapping network nonlinearly combines these two types of semantic vectors to learn the semantic scale shift from their interaction relationships, generating a calibration factor to calibrate the rating features and unify the semantic scale. As an exemplary implementation, the execution steps of the aforementioned trust semantic calibration module 102 include: The core objective of the trust semantic calibration module is to resolve trust semantic biases caused by different users' subjective rating habits and calibrate them to a unified expression scale. In actual trust rating, the rating given by the user is essentially a "projection" of the trustee's objective credibility onto the trustor's personal rating rigor scale. Influenced by factors such as the trustor's subjective cognition, rating habits, situational awareness, and social relationships, this "projection" often undergoes semantic shifts. Therefore, the calibration process needs to combine the characteristics of both the trustee and the trustor to jointly correct such biases.

[0032] The calibration characteristic of the trusted individual is called rating credibility, which is the degree of objective recognition the trusted individual receives in community feedback, reflecting their actual credibility. The calibration characteristic of the trustor is called rating strictness, which is the subjective strictness exhibited by the trustor in historical rating behavior, reflecting the "tightness" or "looseness" tendency of their personal rating scale.

[0033] Figure 3 illustrates the flowchart of trust semantic calibration. As shown in Figure 3, the calibration process consists of two steps: First, decoupling and characterizing features: extracting the credibility of the trusted party's rating and the strictness of the trustor's rating; second, deep fusion and calibration: nonlinearly and deeply fusing the two types of features to learn the calibration factor of the trust relationship, aligning the initial rating features, and generating standardized rating semantic calibration features. Since the influence of subjective rating and objective credibility on the final rating semantics exhibits a nonlinear composite relationship, a nonlinear interactive mapping network is used here to fuse the two, in order to more accurately capture and correct semantic biases in the rating.

[0034] Specifically, the steps for determining the calibration factor include: First, obtaining the out-degree neighbor set of the trustee node and its initial features and initial rating features, and calculating the rating strictness vector through out-edge directed graph convolution and nonlinear transformation.

[0035] Here, the trustor node is the user node within the trust neighborhood, and the trust neighborhood consists of the user's trusted neighbors and the trust relationships pointing to those neighbors. The rating strictness reflects the severity of the trust rating given by the trustor and can be quantified from the trustor's rating behavior towards users within their trust neighborhood. Based on the trustor's trust neighborhood, the model uses directed graph convolution to aggregate the node features of trusted neighbors and trust rating information to generate the trustor's rating strictness.

[0036] As an example, the rating strictness vector of trustor (user) u. The expression can be: In the formula, Let represent the learnable linear projection parameters of user u. This represents the set of users within user u's trusted neighborhood. This represents the number of users in the user's trusted neighborhood. Represents the initial characteristics of user v. This represents the initial rating feature of user u for user v. This represents the bias value for user u. This indicates a feature splicing operation.

[0037] The second step is to obtain the set of in-degree neighbors of the trusted node and its initial features and initial rating features, and then calculate the rating credibility vector through in-edge directed graph convolution and nonlinear transformation.

[0038] Here, the trusted node refers to the user node within the trusted neighborhood, which consists of the trusted user's neighbors and the trust relationships pointing to the user. The rating credibility reflects the objective credibility of the trusted node, which can be inferred from the collective ratings of the trusted node by the trusted neighborhood. Based on the trusted node's trusted neighborhood, the model uses directed graph convolution to aggregate the node features and trust rating features of the trusted neighbors to generate the trusted node's credibility.

[0039] As an example, the rating credibility vector of the trusted party (user) v. The expression can be: In the formula, Let represent the learnable linear projection parameters of user v. This represents the set of users within user v's trusted neighborhood. Represents the initial characteristics of user u. This represents the initial rating feature of user u for user v. This represents the bias value of user v. This indicates a feature splicing operation.

[0040] The third step involves concatenating the rating strictness vector and the confidence vector, and then inputting them into a multilayer perceptron to generate calibration factors.

[0041] As an example, the expression for the calibration factor can be: In the formula, The calibration factor is represented by tanh, and the hyperbolic tangent activation function is represented by tanh. This represents a multilayer perceptron. This represents the rating strictness vector of the trustor (user) u. This represents the credibility vector of the rating of the trusted user (v).

[0042] The user's rating semantic calibration features are obtained by calibrating their initial rating features using calibration factors, let Let the initial rating feature of user u to user v be represented. Then, the expression for its calibrated rating semantic feature can be: In the formula, This represents the calibrated rating semantic features of user u for user v. This indicates an element-wise multiplication operation.

[0043] Thus, this embodiment has obtained the calibrated rating semantic features corresponding to the user.

[0044] The trust semantic interpretation module 103 is used to combine the user's individual rating statistical features with the neighborhood information aggregated by the out-edge graph convolution to generate trustee role features.

[0045] Here, the trustee role feature is used at least to characterize the personalized trust expression features of the user as a trustee.

[0046] In Figure 2, some users (such as users a and b, which are darker in color) have high social frequency, resulting in a rich number of trust edges and more stable trust behavior; while other users (such as user f, which is lighter in color) have sparse interactions and limited available trust data, making it difficult to directly observe their true trust tendencies. This imbalance makes it difficult for the model to accurately capture individualized trust patterns and reduces the interpretability of trust semantics.

[0047] To address the aforementioned imbalance in trust semantic expression, this invention designs a trust semantic interpretation module, which reveals users' subjective preferences in trust expression, enabling the model to understand the semantic tendencies of different users. Based on the trust relationship established by the ratings given by trusters, the module interprets the frequency, average, and stable patterns of user trust ratings by statistically analyzing the out-degree, mean, and variance of ratings, embedding these patterns as individual semantic expression features. Simultaneously, the model captures local social influences through out-edge graph convolution based on the features of trust neighbors and the corresponding trust relationship strengths within the trust neighborhood, interpreting user social patterns and obtaining social semantic expression features. Finally, a gated feature fusion mechanism is employed to integrate social and individual semantic expression features, fully combining individual frequency, stability, average, and social patterns to interpret user trust rating habits and generate truster role features.

[0048] The core objective of the trust semantic interpretation module is to analyze users' trust behavior and interpret their rating habits as trusters. Users' rating habits are primarily shaped by two factors: firstly, the individual's inherent rating behavior patterns (referred to as individual semantic expression), and secondly, their social environment (referred to as social semantic expression). Individual semantic expression reflects the user's inherent preferences and stable patterns during the rating process; social semantic expression depicts the social cues and collective tendencies conveyed by their trust neighborhood (including trusted neighbor nodes, interaction methods, and feedback value orientations). These social tendencies continuously influence an individual's trust judgments and may lead to the formation of a rating pattern with a certain degree of bias.

[0049] Figure 4 shows a flowchart illustrating the trust semantic interpretation process. The entire interpretation process can be divided into three steps: Step 1 is individual semantic interpretation: based on the user's historical rating behavior, analyze their trust rating semantics at the individual level. Step 2 is social semantic interpretation: utilizing the user's trust neighborhood (i.e., their trusted neighbor nodes and related interaction information), interpret the influence of the social environment on their trust concepts, and characterize the shaping effect of others' trust attitudes and value orientations on the user. Step 3 is feature fusion and role characterization: based on obtaining individual semantic expressions and social semantic expressions, an adaptive fusion mechanism is used to weigh the contributions of both, ultimately forming an overall semantic characterization of the user's trustee role.

[0050] Specifically, obtaining the trustee role characteristics includes: First, calculating the mean and rating fluctuation of all out-degree neighbors of the user node, combining the rating mean, rating fluctuation with the node out-degree, and mapping them through a multilayer perceptron to form individual semantic expression features.

[0051] Individual semantic expression represents a user's inherent preferences and judgment criteria for different levels of trust. It can be extracted from a user's trust rating behavior and includes three dimensions: activity, tendency, and stability. These dimensions characterize a user's rating habits from different perspectives. Activity reflects the user's level of participation in social networks; tendency describes whether the user's overall trust level leans towards lenient or strict; and stability measures the degree of fluctuation in their rating behavior. The model extracts these dimensions using frequency, average, and stable patterns respectively, and encodes them uniformly into the same trust feature space, constituting the user's individual semantic expression in specific contexts.

[0052] Frequency pattern (node ​​out-degree) This can reflect a user's activity level on a social network, usually measured by the number of ratings actively issued by the user. Its expression can be: In the formula, This represents the out-degree of the node representing user v. This represents the number of trust ratings that user v sends to users within its trusted neighborhood.

[0053] Average pattern (rating mean) This can reflect a user's overall rating tendency and the stringency of their overall rating standards. It is quantified by the average rating value of all objects within their trust neighborhood, and its expression can be: In the formula, This represents the average rating of user v. This represents user v's trust rating of user u.

[0054] Stable mode (rating volatility) This can represent the degree of volatility in user ratings, measured by the standard deviation or variance of their ratings. Its expression can be: In the formula, This indicates fluctuations in the user node's rating.

[0055] Therefore, the personal semantic expression features of user v can be represented as: In the formula, This represents the personal semantic expression characteristics of user v; This represents a multilayer perceptron that learns the influence of various factors and their higher-order interactions on trust expression through nonlinear transformations, and finally outputs the individual semantic expression features of the user. This indicates an embedding operation. The feature concatenation operation can embed three factors with different dimensions into the same low-dimensional continuous space.

[0056] The second step involves performing edge graph convolution aggregation on the user node and aggregating neighbor information to obtain the social semantic expression features of the user node as a trustee.

[0057] Social semantic expression aims to reveal the deep socialization characteristics of users within social networks. According to sociological theories of social influence, a user's trust concepts are not isolated but deeply embedded within their social circle. A user's trust preferences often show a high correlation with their trusted neighbors—that is, the user tends to adopt the views of their trusted circle or is subtly influenced by their neighbors. Specifically, the influence of neighbors on users stems primarily from two dimensions: first, the inherent attributes of the neighbors themselves, representing the "quality" of the social circle; and second, the user's trust behavior towards their trusted neighbors, i.e., the trust ratings the user assigns to their neighbors, which reflect the user's expression of trust and value orientation within their trust circle. These characteristics collectively constitute the "social semantic field" surrounding the target user.

[0058] Based on the above analysis, a user's social semantic expression needs to comprehensively consider the node features of all neighbors in their trust neighborhood, as well as the trust rating information given by the user to each trusted neighbor, thereby quantifying the comprehensive impact of the local social environment on the user's trust perception. Therefore, the model uses directed graph convolution operations to jointly aggregate the user's initial features and corresponding trust rating features within the user's trust neighborhood only along the outgoing edges of the trust relationship, generating the user's social semantic expression.

[0059] As an example, the social semantic expression features of user v The expression can be: In the formula, Represents the initial characteristics of user u. This represents the initial rating feature of user v for user u. This represents the set of users within user v's trusted neighborhood. This indicates a feature splicing operation.

[0060] The third step involves using a gating feature fusion mechanism to determine the weights of the social semantic expression features and the individual semantic expression features, and then performing weighted fusion to form the trustee role features of the user node.

[0061] Users' trust role characteristics are used to express their rating habits. In the trust decision-making process, users exhibit significant individual differences in their reliance on "internal behavioral patterns" and "external social environment": some users tend to adhere to their own trust benchmarks, while others place greater emphasis on the influence of the social environment. Therefore, it is necessary to distinguish the different impacts of individual semantics and social semantics on trust role characteristics during the modeling process.

[0062] Based on the above analysis, in order to adaptively adjust the contribution of the two types of semantics, this module adopts a gated feature fusion mechanism to adaptively weight and fuse individual semantic expression and social semantic expression, thereby quantifying the characteristics of the trustee role.

[0063] As an example, the trustee role characteristics of user v The expression can be: In the formula, The weight coefficients represent the social semantic expression features. sigmoid represents the sigmoid activation function. This represents the first learnable matrix. This represents the social semantic expression features of user v. This represents the first learnable bias value. This indicates a feature concatenation operation. Weight coefficients representing the semantic expression features of an individual , This represents the second learnable matrix. This represents the personal semantic expression features of user v. This represents the second learnable bias value.

[0064] The larger the value, the more significant the influence of user v's external social environment. The larger the value, the more significant the influence of the individual behavioral patterns of user v.

[0065] Thus, this embodiment has obtained the user's trustee role characteristics.

[0066] The trust semantic gain module 104 is used to generate the trusted party role features based on the calibrated rating semantic features, and combine them with the trustor role features to form a dual-role trust feature; the dual-role trust feature is then subjected to convolutional enhancement processing based on global semantic ranking to generate the final trust feature of the user node.

[0067] Here, the characteristics of the trusted party role are used at least to characterize the personalized trust expression characteristics of the user as the trusted party role.

[0068] In social trust graphs, most users lack direct trust relationships, resulting in insufficient semantic information to support the model's learning of rating patterns. For these information-scarce nodes, semantic differences and data sparsity combine, causing them to fall into a "semantic ambiguity and insufficient data" dilemma in global trust propagation, further weakening the stability of overall predictions. For example, in Figure 2, user f, due to the limited number of trust edges, lacks learnable semantic references, making it difficult for their trust expressions to be accurately understood and effectively utilized globally.

[0069] To address the insufficient trust semantic information caused by sparse trust relationships, this invention designs a trust semantic gain module to improve the completeness of node semantics. The trust semantic gain module consists of two parts: local gain and global gain. Local gain, based on rating semantic calibration features and the user's trust neighborhood, generates trusted party role features, which are then concatenated with the trustor role features to form the user's dual-role trust features. Global gain, based on the dual-role trust features of user nodes, proposes a "ranking-convolution" mechanism to overcome the limitation of traditional GCNs (Graph Convolutional Networks) that rely solely on local neighborhoods. In the global trust feature space, implicit neighborhoods are constructed based on semantic similarity, and then features of globally similar nodes are extracted through one-dimensional convolution, achieving cross-structure semantic information supplementation.

[0070] The trust semantic enhancement module aims to address the problem of insufficient semantic information caused by structural sparsity in social trust networks. Trust semantic enhancement information typically originates from the network structure, including both local and global network structures. Local network structures usually include user profiles with dual roles as both the trustor and the trusted party, while the global network structure contains similar semantic information from the trust feature space. By fusing these two types of information, a final user trust feature can be formed, based on local semantic information and supplemented by global semantic information.

[0071] Figure 5 illustrates the flowchart of trust semantic gain. The trust semantic gain process includes the following three steps: First, local gain is achieved by aggregating trust information based on the user's trusted neighborhood, constructing the user's trusted role features, and fusing them with the user's trustor role features to obtain the user's dual-role trust features; Second, global gain is achieved by constructing an implicit semantic neighborhood (a group of user nodes semantically similar to the target user) based on trust semantics using a ranking method, and aggregating trust information in this neighborhood to obtain cross-local domain trust semantic enhancement features; Third, the dual-role trust features and trust semantic enhancement features are residually joined to obtain the user's trust features.

[0072] Specifically, acquiring dual-role trust features includes: Local gain, used to characterize a user's complete trust profile within a local network, requires integrating semantic information from both the user's roles as a trustor and a trusted party. The trustor role features reflect the user's rating pattern when placing trust in others; while the trusted party role features reflect the group trust and recognition the user receives within the social network. Local gain treats these two types of information as two complementary perspectives on the same user's local trust state, integrating them within a unified semantic space: on the one hand, preserving the differences inherent in each role; on the other hand, through their collaborative characterization, forming a holistic description of "how I see others" and "how others see me," thus obtaining a more complete expression of local trust.

[0073] The first step is to perform an in-edge graph convolution operation on the calibrated rating semantic features of the user node to generate the trusted party role features.

[0074] Information about the trusted party's role characteristics is primarily contained within the trust neighbor of the trusting user. Since different trustors have varying rating standards and habits, directly aggregating initial rating features from their trust neighbor to characterize the trusted party's role characteristics would amplify these subjective differences. The calibrated rating semantic features already constructed above align to some extent with different user rating scales and better reflect the trusted party's stable reputation level. Based on this consideration, the calibrated rating features are used when modeling the trusted party's role characteristics, together with the initial features of neighboring nodes in the trust domain, to represent the comprehensive rating signal from the group. The trusted party's role characteristics are obtained by referring to the user's social trust expression capture method.

[0075] As an example, the trusted person role characteristics of user v The expression can be: In the formula, This represents the set of users within user v's trusted neighborhood. Indicates the user's initial characteristics, This indicates a feature concatenation operation. This represents the rating semantic calibration feature, i.e., the calibrated rating semantic feature.

[0076] The second step is to concatenate the trusted party's role characteristics with the trustor's role characteristics to obtain dual-role trust characteristics.

[0077] Dual-role trust features are used to represent a user's overall trust status, and their formation depends on the interaction and synergy between the trustor's role features and the trusted party's role features. To characterize this inter-role relationship, the model jointly maps the two types of role features and introduces a non-linear activation function to obtain the user's dual-role trust features.

[0078] As an example, user v's dual-role trust characteristics The expression can be: In the formula, Let W represent the nonlinear activation function, and W represent the learnable matrix. This indicates the characteristics of user v's trustee role. b represents the trusted user role characteristic of user v, and b represents the bias value. This indicates a feature splicing operation.

[0079] Specifically, the dual-role trust features are enhanced through convolutional processing based on global semantic ranking to generate the final trust features of user nodes. This includes: the goal of global gain is to supplement and enhance local features by leveraging semantically similar node information in the overall network. Homogeneity theory in sociology states that individuals with similar attributes and attitudes tend to form similar behavioral patterns and cluster in similar social positions, even if there is no direct relationship between them. This means that in a trust network, users in similar semantic positions often have similar trust preferences and decision-making patterns. Therefore, even if two users do not have explicit edges on the original trust graph, as long as they are sufficiently close in the semantic space learned by the model, they can be considered to have an "implicit trust relationship," suitable for complementing each other's information. This invention further constructs implicit semantic proximity neighborhoods for the users to be predicted based on local social semantics, enhancing the semantic expression of users by mining numerous similar implicit semantics.

[0080] Based on the above analysis, the basic idea of ​​quantifying global gain is as follows: In the learned semantic representation space, a comparable "semantic position" is characterized for each user, and based on this, a group of candidate neighbors with similar semantics are organized and selected globally. The trust information of these global semantic neighbors is then integrated with the local trust information of the target user to form a richer and smoother trust semantic expression. To achieve this goal, this invention sequentially constructs semantic ranking scores, semantic ranking vectors and indices, and a trust feature matrix for all users in the semantic space to characterize the user's semantic position and organize their global semantic neighbors.

[0081] The first step is to perform a linear mapping on the dual-role trust features to obtain the semantic ranking score of each node, and then perform a global ranking of all nodes based on the semantic ranking score to obtain the ranked trust feature vector sequence and the corresponding global ranking index.

[0082] In this embodiment, the semantic ranking score is used to represent the user's position in the semantic space on a uniform scale. The model maps the high-dimensional dual-role trust features to one-dimensional real numbers using a learnable linear projection function, letting... The semantic ranking score of user v can be expressed as: In the formula, Represents the learnable projection matrix. This represents the dual-role trust characteristic of user v.

[0083] In this embodiment, the semantic ranking vector and ranking index are used to rank and retrieve a user's semantic nearest neighbors globally. Let S denote the semantic ranking vector. The sorted index representing all users can be expressed as: In the formula, This indicates an ascending order operation, which sorts all users' semantic ranking scores in ascending numerical order.

[0084] The second step is to arrange the feature vectors of each node in order based on the global sorting index, and construct a global node representation matrix.

[0085] In this embodiment, the overall user trust feature matrix is ​​used to align the sorted semantic neighbor information with user features and explicitly model the global semantic neighborhood in sequence form. The model rearranges the dual-role trust features of all users according to the sorting index and weights them with the semantic sorting vector to obtain the overall user trust feature matrix (i.e., the global node representation matrix), whose expression can be: In the formula, Y represents the trust feature matrix of all users. This indicates the dual-role trust characteristic of user v. This represents the matrix obtained by rearranging the dual-role features according to the sorting index, and S represents the semantic sorting vector.

[0086] The third step involves applying a one-dimensional convolution kernel to the sorting dimension of the global node representation matrix to perform a sliding window operation, in order to extract and aggregate node features with similar trust patterns.

[0087] In this embodiment, the global semantic representation is a user feature representation that integrates implicit semantic neighborhood information. In the overall user trust feature matrix, adjacent user nodes exhibit high similarity in the semantic space, forming implicit semantic neighborhoods. Based on this observation, this embodiment employs a one-dimensional convolutional kernel for sliding window aggregation along the ranking dimension, with each convolutional window corresponding to the implicit semantic neighborhood of the user at the center position. By aggregating the trust information within each implicit semantic neighborhood, the global semantic representation of the user node is obtained.

[0088] In the formula, This represents the global semantic representation of the user node with location index i. This represents the non-activation function, and K represents the window size of the convolution kernel. This represents the weight parameters of the c-th convolutional kernel at the k-th position within the window. This represents the bias value corresponding to the c-th convolutional kernel. This represents the feature vector of the user node with index i+k in the global node representation matrix.

[0089] The fourth step involves using the inverse mapping of the global sorting index to restore the features output by the convolution operation to the original node order, thereby obtaining trust semantic enhancement features.

[0090] In this embodiment, the trust semantic enhancement feature is the trust feature after the user completes the global gain. Since the scale of the output feature changes with the neighborhood structure and user characteristics after the one-dimensional convolution aggregates the neighborhood features, in order to maintain consistency with the initial feature space, after restoring the convolution output to the original feature arrangement order through the inverse sorting index, this embodiment performs normalization (LayerNorm) processing to obtain the trust semantic enhancement feature.

[0091] In the formula, This represents the trust semantic enhancement feature for user v. Z represents the normalization function, such as layer normalization, and Z represents the set of global semantic representations of user nodes. This represents the inverse mapping based on the globally sorted index. This means extracting the feature vector corresponding to the index position of user v from the global matrix after restoring the order.

[0092] The fifth step is to perform a residual connection between the trust semantic enhancement features of the same user node and the dual-role trust features to generate the final trust features.

[0093] Here, concatenating the two feature residuals together can achieve gain fusion. The purpose of gain feature fusion is to enhance the expressive power of trust features, which satisfies both the preservation of dual-role semantics in the local network and the introduction of supplementary information from the global semantic neighborhood.

[0094] As an example, the expression for user v's final trust characteristic can be: In the formula, This represents the final trust characteristic of user v. This indicates the dual-role trust characteristic of user v. This represents the trust semantic enhancement feature for user v.

[0095] Referring to the process for determining the final trust characteristics of user v described above, the final trust characteristics of any user to be predicted can be obtained.

[0096] Thus, this embodiment has obtained the final trust characteristics of the user to be predicted.

[0097] The trust assessment module 105 is used to concatenate the final trust features for any pair of users to be predicted, input them into a multilayer perceptron, and output a predicted trust rating to supplement the missing trust relationships in the online social network.

[0098] In a preferred embodiment of the present invention, the trust assessment module is implemented by the following steps: First, for each user pair to be predicted, the corresponding final trust feature vector is obtained; second, the two feature vectors are combined in a preset manner to form a joint input feature, the organization method including but not limited to: concatenation or element-wise multiplication, preferably using concatenation to preserve the integrity of the role information of both parties; then, the joint input feature is input into a multilayer perceptron (MLP), which contains at least one hidden layer, each layer uses a nonlinear activation function, and the last layer outputs a scalar value; in order to make the output value fall within a reasonable trust score range (e.g., 0-5 points or a continuous value of 0-1), a Sigmoid or Tanh activation function is configured at the output of the MLP and mapped to the target score range through a linear transformation.

[0099] Furthermore, the parameters of the MLP are trained through supervised learning: observed trust relationships in online social networks are used as positive samples, and their true trust ratings are used as labels; for unobserved user pairs, negative sampling strategies can be used to generate negative samples, or regression training can be performed only on positive samples; the loss function uses mean squared error (MSE) or mean absolute error (MAE), and the model parameters are optimized through backpropagation and gradient descent.

[0100] Ultimately, for any user pair lacking a trust relationship, the model's output prediction is the completed trust rating, which can be used to recommend trustworthy users, enhance social interaction, or build a more complete trust graph.

[0101] Based on the same inventive concept as the graph convolutional network-based social network semantic alignment trust evaluation system, another embodiment of this invention proposes an SA-GCN model for implementing social network trust evaluation, as shown in Figure 14. The SA-GCN model includes three layers: an input layer, a semantic learning layer, and an output layer. These modules work collaboratively to form a complete "calibration-interpretation-gain" processing flow, achieving the alignment and enhancement of user trust semantics.

[0102] The first layer is the input layer, which includes a trust information encoding module. This module is responsible for mapping the discrete graph structure and trust rating information in the social network to a continuous feature space, providing initial features for subsequent semantic processing.

[0103] Trust information in social network graphs is mainly divided into two types: trust relationships and trust ratings. To achieve unified modeling of trust relationships and rating information, this invention performs feature encoding on both.

[0104] Trust relationship information is obtained using graph embedding, which maps the relationship information of user nodes in a high-dimensional discrete graph structure to a low-dimensional, continuous latent representation space, generating the user's initial features. Specifically, user... initial features D represents the size of the feature dimension. Trust rating information is obtained by one-hot encoding of specific trust ratings between users. The trust rating from user u to user v is represented as... Then, a learnable weight matrix is ​​used to transform it into a dense vector, which serves as the corresponding trust relationship edge. Initial rating features .

[0105] As an example, the expression for the initial rating feature can be: In the formula, , Let represent the learnable weight matrix, and D represent the size of the feature dimension. This represents the initial rating feature of the edge formed by users u and v.

[0106] The second layer is the semantic learning layer, which consists of a trust semantic calibration module, a trust semantic interpretation module, and a trust semantic gain module. This layer adopts a combined serial and parallel architecture. The specific process is as follows: The trust semantic calibration module takes the initial features and initial rating features of the user node as input, performs outgoing edge directed graph convolution and incoming edge directed graph convolution based on the trust neighborhood and the trusted neighborhood, respectively, to calculate the rating strictness of the trustor and the credibility of the trusted party, and generates calibration factors through a nonlinear interactive mapping network to adjust the initial rating features, outputting rating semantic calibration features.

[0107] Meanwhile, the trust semantic interpretation module takes the initial features of user nodes and the user's specific trust rating as input. On the one hand, it generates individual semantic expression features by statistically analyzing the user's rating out-degree, rating mean, and rating variance. On the other hand, it aggregates trust neighborhood information through out-edge directed graph convolution to generate social semantic expression features. Then, it uses a gating feature fusion mechanism to perform weighted fusion of the two types of semantic features and outputs the user's trustee role features.

[0108] Building upon this foundation, the trust semantic gain module takes the initial features of user nodes and rating semantic calibration features as input. It performs directed graph convolution with incoming edges based on the trusted neighborhood to generate the user's trusted role features. These features are then concatenated with the trustor role features output by the trust semantic interpretation module to obtain the user's dual-role trust features. Next, the dual-role trust features of all users are sorted to construct a trust feature matrix for all users. A one-dimensional convolution operation is then performed along the sorting dimension to aggregate trust information from semantically similar users globally. After inverse sorting and residual connections, the final trust features of the user are output. These three modules collaboratively implement a multi-level feature extraction process, from semantic scale alignment and individual expression interpretation to global information enhancement, providing high-quality user trust features for subsequent trust assessment.

[0109] It should be noted that the detailed calculation and implementation process of trust semantic calibration, trust semantic interpretation and trust semantic gain can be determined through the steps involved in the aforementioned trust semantic calibration module 102, trust semantic interpretation module 103 and trust semantic gain module 104.

[0110] The trust semantic calibration module takes the processing results of the first layer as input, including the initial features and initial rating features of the user node, and outputs the rating semantic calibration features. Simultaneously, the trust semantic interpretation module takes the initial features of the user node and the user's specific trust rating output from the first layer as input, and outputs the user's trustee role features. The trust semantic gain module consists of two parts: local gain and global gain. The local gain takes the initial features of the user node and the rating semantic calibration features obtained from the trust semantic calibration module as input, resulting in the user's trusted party role features, and then outputs the user's dual-role trust features. The global gain module takes the user's dual-role trust features as input and outputs the user's final trust features.

[0111] The third layer is the output layer, namely the trust assessment module, which evaluates the trust relationship between two users based on their final trust characteristics and outputs a rating prediction value for unknown trust relationships.

[0112] In a preferred embodiment of the present invention, for two users u and v to be assessed for trust, their corresponding final trust features are obtained respectively. and The two values ​​are then concatenated and input into a classic multi-class classification task to output the corresponding trust evaluation value. .

[0113] As an example, trust assessment value The expression can be: In the formula, Represents the normalized exponential function, This represents a multilayer perceptron. This represents the final trust characteristics of user u. This represents the final trust characteristic of user v. This indicates a feature splicing operation.

[0114] Overall, the SA-GCN described above uses a "calibration-interpretation-gain" approach to construct a semantic alignment and feature enhancement process: the trust semantic calibration module eliminates semantic scale bias, the trust semantic interpretation module models differentiated semantic expressions at the individual level, and the trust semantic gain module fully enriches trust semantic information. The synergistic effect of these three modules enables the model to learn more consistent, interpretable, and generalizable trust representations in highly subjective and sparse social environments, laying a solid foundation for subsequent trust evaluation.

[0115] In one specific embodiment of the present invention, the SA-GCN model can exhibit its forward propagation process, taking a social trust graph as input, including a set of user nodes, a set of directed edges, and a trust rating vector. The model sequentially performs steps such as trust information encoding, semantic calibration, dual-role feature generation, and trust prediction, and finally outputs a predicted trust rating value for any pair of users.

[0116] To illustrate this process more clearly, the following pseudocode describes the forward propagation logic of the SA-GCN model: The first two lines correspond to the trust information encoding module, which initializes the initial features of each node and the initial rating features of each edge.

[0117] Lines 3 to 8 correspond to the trust semantic calibration module: First, the strictness of the trustor and the credibility of the trusted party are calculated. Line 7 generates calibration factors to calibrate the original rating (line 8) and eliminate the semantic scale difference in ratings between users.

[0118] Lines 9 to 16 correspond to the trust semantic interpretation module: statistical node out-degree, rating mean and standard deviation, generating individual semantic expression features (lines 10-12). Perform out-edge convolution to collect local trust information (line 13), and use a gating feature fusion mechanism to fuse the social semantic expression features and individual semantic expression features of the trustee role to obtain the trustee role features (lines 14-16).

[0119] Lines 17 to 24 correspond to the trust semantic gain module: Based on the calibration rating features, a directed graph convolution operation with incoming edges is performed on the trusted party to obtain the trusted party role features (line 17). Then, the trustor role features and trusted party role features of the nodes are concatenated to form a dual-role comprehensive trust feature (line 18). Through trust semantic ranking, a global context matrix Y is constructed (lines 19-21). One-dimensional convolution is used to extract similar trust patterns in the ranking dimension (lines 21-22). Then, the node order is restored through an inverse ranking operation (line 23), generating trust semantic enhancement features. Residual connections are performed on the dual-role comprehensive trust features and trust semantic enhancement features to generate the final trust features (line 26).

[0120] Lines 27 to 29 correspond to the trust assessment module: for an unknown trust relationship between two users, the final trust features of the two users are concatenated and input into the MLP to predict the trust rating.

[0121] As an example, the time complexity analysis of SA-GCN is as follows: Let This represents the number of nodes in a social network graph. Let represent the number of edges in the graph, and d represent the node feature dimension. The trust rating semantic convolution kernel window size is K. In the trust semantic calibration and behavior modeling stage, a neighbor aggregation operation is performed on each node, with a time complexity of O(k). In the global similarity aggregation stage, the features are ranked as follows: Convolution is The time complexity of the evaluation phase is... The total time complexity of SA-GCN is approximately: To verify the effectiveness and practicality of the method described in this invention, a comparative experiment was conducted on a typical social network dataset. It should be noted that the following experimental setup is merely illustrative and does not constitute a limitation on the scope of protection of this invention.

[0122] Experimental Datasets: The experiments used the real-world datasets Advogato and PGP. These two datasets are widely used in the field of social network trust assessment. The Advogato dataset originates from an online software development community. When constructing the social network graph, each user is considered a node, and the "authentication" relationship between users is represented as a directed edge, with the edge pointing from the user issuing the authentication to the user being authenticated. There are three authentication levels: Apprentice, Journeyer, and Master, corresponding to trust ratings of 1, 2, and 3, respectively, with higher values ​​indicating higher trust levels.

[0123] The PGP dataset originates from a public key authentication network, where each PGP user represents a node. If user A signs user B's public key, it is considered a directed edge from A to B. The signature level reflects the strength of trust between users, and there are four different levels, corresponding to trust rating values ​​of 1, 2, 3, and 4, with higher values ​​indicating higher trust. Table 1 shows the statistics for the Advogato and PGP datasets.

[0124] Baseline Models for the Experiment: The experiment uses classic trust assessment models as the baseline model for SA-GCN, including Matri, Guardian, GATrust, TrustGuard, and TrustGNN. Matri is a matrix factorization-based trust assessment model. Guardian is a graph neural network-based trust assessment model that focuses on the asymmetry of trust relationships. GATrust is a graph attention network-based trust assessment model that focuses on the impact of user subjectivity on user trust preferences. TrustGNN, also a graph neural network-based trust assessment model, divides the trust propagation process into multiple propagation paths, focusing on the impact of user subjectivity on trust propagation patterns. TrustGuard is a dynamic graph neural network-based trust assessment model with a spatiotemporal dual-channel structure, focusing on the asymmetry of trust relationships and the impact of spatiotemporal dynamic trust relationships on trust assessment results.

[0125] Experimental Environment Configuration: The experiment uses Python 3.8 and PyTorch 2.0.0 frameworks. The initial features of SA-GCN nodes are generated using node2vec, with all feature dimensions d=128. In determining hyperparameters, this invention employs a grid search: the model training process uses the Adam optimizer, with a learning rate range of {0.0001, 0.001, 0.005, 0.01, 0.05}. The L2 regularization coefficient is searched within the range {10...}. -5 10 -4 The dropout rate is set to a range of {0.0, 0.1, 0.3, 0.5, 0.8}. The Xavier initializer is selected to initialize the model parameters. The learning rate is initialized to 0.0001 and the L2 regularization coefficient is initialized to 10. -4 Dropout was initialized to 0.0, and then other values ​​were selected sequentially within the chosen range. After multiple experiments, the optimal learning rate was found to be α=0.01, the optimal L2 regularization coefficient was 10⁻⁵, and the optimal Dropout rate was 0.0. For the trust semantic gain module, its one-dimensional convolutional kernel size was K=3. The training epochs were 200. The parameter settings used in the baseline model were the same as those in the references used.

[0126] It should be noted that all experiments were run in the same hardware and software environment, and the experimental results are the average of 20 runs.

[0127] The evaluation metrics for the experiment are as follows: Model performance evaluation metrics may include: numerical prediction accuracy, overall classification accuracy, and imbalanced data metrics.

[0128] The numerical prediction accuracy metric F1-micro is the harmonic mean of precision and recall, effectively evaluating the overall performance of a model in correctly classifying unknown trust relationships into different levels. Its calculation formula is as follows: Where TP is the number of true positives, FP is the number of false positives, and FN is the number of false negatives. A larger value indicates a more robust model.

[0129] Overall classification accuracy (MAE) is the mean absolute error, used to measure the average absolute error between the model's predicted trust rating and the true value. Its calculation formula is as follows: Where n is the number of samples, For the true value, This represents the predicted value. The smaller the MAE value, the higher the accuracy of the model evaluation.

[0130] The MCC (Mean Cross-Sectional Class Control) metric measures the reliability of a model's predictions under class imbalance conditions. Its calculation formula is as follows: Where TN is the number of true negative cases, TP is the number of true positive cases, FP is the number of false positive cases, and FN is the number of false negative cases. The closer the value is to 1, the more reliable the model evaluation results are.

[0131] The experimental design is as follows: To comprehensively verify the performance of the proposed SA-GCN model, experiments include accuracy comparison, robustness comparison, ablation experiments, semantic calibration effect analysis, and hyperparameter sensitivity analysis. Specifically, the accuracy comparison experiment compares SA-GCN with existing classic trust evaluation models (including Matri, Guardian, GATrust, TrustGNN, and TrustGuard) to verify SA-GCN's advantage in accuracy. The robustness comparison experiment compares SA-GCN with GNN-based baseline models (including Guardian, GATrust, TrustGNN, and TrustGuard), simulating changes in trust data sparsity by altering the ratio of training to test sets, thus verifying SA-GCN's advantage in data sparsity scenarios. The ablation experiment compares SA-GCN with three variant models to quantify the calibration module and the contributions of the interpretation and gain modules to the overall performance of the SA-GCN model. The trust semantic calibration effect analysis experiment compares the distribution of trustee rating bias before and after model training, thereby verifying the effectiveness of the calibration module in correcting user semantic bias. The hyperparameter sensitivity analysis experiment examines the rationality of parameter settings by changing the convolution kernel size, thereby verifying the impact of the convolution kernel size in the trust semantic gain module on model performance.

[0132] The experimental process is as follows: 1. Dataset analysis experiment: Analyze the data distribution characteristics of the Advogato and PGP datasets, aiming to demonstrate the necessity and applicability of the "calibration-interpretation-gain" three-stage processing flow proposed in this invention from the data level.

[0133] The necessity of trust semantic calibration analysis is analyzed as follows: To measure the deviation of the rating strictness of trustors from the average level of social networks, this invention introduces a rating deviation index. The calculation formula is: ;in, This represents the average of all ratings received by the trusted node v, calculated using the following formula: If the rating deviation This indicates that the user's rating is stricter than the average, and vice versa.

[0134] The trust rating bias distribution is shown in Figures 6 and 7. In the distribution, the x-axis represents the rating bias (i.e., the centralized residual mean of trusters), and the y-axis represents the corresponding probability density. The height of each bar represents the relative proportion of users falling within that bias range. The KDE (Kernel Density Estimation) curve smoothly fits the histogram data, clearly revealing the overall shape, peaks, and long tail characteristics of the data distribution. The vertical dashed lines on both sides mark the positions one standard deviation (σ) from the mean, used to measure the main distribution range of the data.

[0135] In Figures 6 and 7, the probability density distributions in both datasets exhibit a clear leptokurtic characteristic, with the peaks closely surrounding the rating bias of 0. This indicates that the trust rating standards of most users in social networks tend to be consistent, forming a "group trust semantic benchmark" for reference. Although the distribution is concentrated, "long tails" extend on both sides of the peaks. This means that although extreme users are few in number, they do exist, and their bias values ​​are widely distributed. The existence of these outliers intuitively demonstrates the widespread nature of subjective cognitive biases among users. If left unaddressed, this will introduce strong trust semantic biases into the trust assessment process.

[0136] In the Advogato and PGP datasets, approximately 80.21% and 71.69% of users, respectively, have absolute bias values ​​within one standard deviation. This finding has dual significance: first, it confirms the necessity of calibration, namely, addressing long-tail outliers; second, it also demonstrates the feasibility of calibration, as this stable distribution centered on the behavior of the majority of users provides a reliable benchmark for rating calibration to achieve semantic alignment with extreme users. This invention can utilize this "group benchmark" to correct the semantics of outliers, making all trust ratings semantically comparable.

[0137] The necessity of interpreting and analyzing trust semantics is as follows: (1) Frequency pattern analysis based on out-degree statistics: The out-degree of a node reflects the activity level of a user as a truster in issuing ratings, i.e., the frequency pattern of trust ratings. The statistical data in Table 1 shows that the maximum out-degree in the Advogato and PGP datasets is as high as 786 and 1264, respectively, while the average out-degree is only 7.82 and 8.38, and there are isolated users with an out-degree of 0 in both datasets. This huge difference indicates that the activity level of users' trust expression in social networks is extremely uneven: a small number of "active users" provide the vast majority of trust ratings, while the rating data of the vast majority of "silent users" is extremely sparse. If not processed indiscriminately, the semantics of active users will dominate the model training. Therefore, it is necessary to model the out-degree of users to interpret their frequency pattern, so as to reasonably weight the contributions of users with different activity levels during evaluation.

[0138] (2) Analysis of average and stable patterns based on mean and fluctuation: Simple frequency analysis is insufficient to characterize users' trust intentions. It is also necessary to combine the rating mean and fluctuation (variance or standard deviation) to reveal the average and stable patterns of users respectively. This invention maps the user rating mean and variance to a two-dimensional space for cluster analysis. Figures 8 and 9 are scatter plots of trust rating mean and variance. Each point in the figure represents a user. The horizontal axis is the rating variance (representing stability), and the vertical axis is the rating mean (representing average preference). The size of the point is proportional to the out-degree (frequency). Through K-means cluster analysis, the Advogato and PGP datasets show 6 and 8 significant behavioral clusters, respectively, revealing the following key findings: Differences in average patterns (vertical axis analysis): As shown in Figure 8, the users in cluster 1 (orange) in the upper right corner have a rating mean close to 3.0, belonging to the "tolerant" type, and tend to give high scores; while the users in cluster 0 (blue) in the lower left corner have a rating mean of only 1.0, belonging to the "strict" type. This demonstrates a significant baseline bias among different users, necessitating the interpretation of an "average pattern" to correct for rating biases. Differences in stable patterns (horizontal axis analysis): Clusters 1 and 0 both have a variance of 0.0, indicating highly stable rating styles. In contrast, while users in cluster 6 (cyan) have a moderate mean (approximately 2.1), their variance is as high as 0.9, indicating extremely volatile ratings and unclear trust criteria. This proves that relying solely on the mean is insufficient; a "stable pattern" must be introduced to measure the reliability of user ratings.

[0139] Generality Validation: As shown in Figure 9, in the larger and sparser PGP dataset, eight significant clusters were also observed, clearly differentiating users into different groups such as "high stability-high score," "high stability-low score," and "high volatility." In summary, the real dataset demonstrates significant individual differences among users in terms of activity (frequency), rating tendency (average), and volatility (stability). This fully demonstrates the necessity of the trust semantic interpretation module proposed in this invention; that is, users must be modeled simultaneously from three dimensions—frequency pattern, average pattern, and stable pattern—to accurately reconstruct their true trust semantics.

[0140] The necessity of trust semantic gain analysis is as follows: The sparsity in Table 1 is used to measure the relative density of available explicit trust information in a social trust graph, and is defined as the ratio of the actual number of directed trust edges in the graph to the theoretical maximum possible number of directed edges. The formula is... Where m is the number of observed directed trust edges, i.e., the number of direct trust relationships between users; n is the number of nodes, i.e., the number of users. As shown in Table 1, the sparsity of Advogato is 0.119%, and the sparsity of PGP is 0.021%. Therefore, a trust semantic gain mechanism needs to be adopted in the prediction process to enrich the expression of user node trust information.

[0141] 2. Accuracy Comparison: Table 2 shows the comparison results of the evaluation accuracy between SA-GCN and the baseline model.

[0142] As shown in Table 2, on the Advogato and PGP datasets, SA-GCN significantly outperforms all baseline models in all three key metrics, and the overall evaluation accuracy of SA-GCN, TrustGNN, GATrust, TrustGuard, Guardian, and Matri shows a downward trend. On the Advogato dataset, for the F1-micro metric, SA-GCN achieves 75.6%, a 1.2% improvement over the second-best TrustGNN (74.4%), and a 1.3% improvement over GATrust (74.3%). It also shows a 2.2% improvement over Guardian and TrustGuard (both 73.4%), and a significant 10.4% improvement over Matri (65.2%). On the MAE metric, SA-GCN performs best at 0.075, significantly outperforming TrustGNN and GATrust (both 0.081). It is 13.8% and 14.8% lower than Guardian (0.087) and TrustGuard (0.088) respectively, and nearly half the performance of Matri (0.141). On the MCC metric, SA-GCN achieves 0.611, again the best performance among all models. This demonstrates that SA-GCN offers higher reliability in predicting imbalanced data. On the PGP dataset, SA-GCN continues to hold its leading position. Its F1-micro score is 87.5%, a 0.4% improvement over TrustGNN (87.1%) and a 0.8% improvement over GATrust (86.7%). While its MAE score remains the lowest (0.082), it significantly outperforms Guardian (0.086) and TrustGuard (0.085). Its MCC score also reaches the highest value of 0.732. These data consistently demonstrate that SA-GCN achieves more accurate trust assessments in both relatively sparse and dense data environments.

[0143] In summary, the SA-GCN model, after calibration, interpretation, and enhancement of trust semantics, can more accurately represent users' trust characteristics, thereby improving the accuracy of trust assessment. In contrast, while the suboptimal TrustGNN model, with its multiple propagation modes, can capture the structural information of trust propagation, it fails to address the semantic scale bias in trust assessment. This bias amplifies during propagation, affecting the final accuracy of trust assessment. GATrust uses graph attention to distinguish users' multidimensional attribute preferences, but similarly, it does not explicitly handle the semantic differences in ratings between users, making it difficult to avoid systematic biases caused by subjectivity. Guardian, as the earliest trust model based on graph convolution, uses directed GCN to model the asymmetry of trust, but completely ignores the inconsistency of trust semantics, making it difficult to maintain stable performance under complex semantic conditions. Matri uses matrix factorization to represent multiple potential factors of the trustor and the trusted, and uses dot product to calculate the trust rating between them. However, because it cannot effectively capture the asymmetry and propagation of trust relationships, its accuracy is the lowest among all models.

[0144] It's worth noting that TrustGuard employs a dual-channel dynamic GNN with spatial and temporal layers to handle dynamic trust relationships. However, semantic misalignment means its dynamic advantages are offset by semantic biases in the input layer. The advantages of SA-GCN over TrustGuard also demonstrate that addressing inherent semantic biases between users before discussing the dynamics of trust is a key entry point for improving trust assessment accuracy. Semantic calibration reduces scale differences at the input level, while semantic interpretation models individual preferences, enabling the model to more accurately reconstruct users' true trust intentions even in the absence of temporal information.

[0145] 3. Robustness Comparison: The experiments in this invention simulate different levels of data sparsity by systematically reducing the proportion of training set data, and compare the performance of SA-GCN with baseline models based on GNN to verify the advantages of SA-GCN in generalization ability and handling data sparsity problems. In the experimental setup, the training sets on the Advogato and PGP datasets were divided into proportions of 80%, 60%, 40%, and 20%, respectively. The two models were evaluated and recorded on three core metrics: F1-micro, MAE, and MCC. Figure 12 shows the trend curves of various performance metrics, with the horizontal axis representing the proportion of the training set and the vertical axis representing the values ​​of each metric. As shown in the figure, on the Advogato and PGP datasets, the performance curves of SA-GCN consistently outperformed other baseline models in all metrics. Specifically, the F1-micro and MCC curves of SA-GCN were always above other models, while the MAE curve was always at the bottom, indicating that it maintained the highest classification accuracy and the lowest prediction error under different levels of sparsity. Crucially, as the training data decreases, the slope of SA-GCN's performance curve becomes flatter, meaning that its performance decline is significantly smaller than that of all other models, indicating that SA-GCN has stronger robustness.

[0146] Table 3 shows the changes in rating metrics for SA-GCN and the baseline model under different training set settings.

[0147] Table 3 As shown in Table 3, on the Advogato dataset, when the training set ratio is 80%, SA-GCN's F1-micro is 75.6%, MAE is 0.075, and MCC is 0.611. As the training set ratio decreases to 60%, the F1-micro only drops to 74.2%, a decrease of 1.4%, while the MAE increases to 0.079, an increase of only 0.004, and the MCC remains at 0.601, a decrease of only 0.010. When the training set ratio is further reduced to 40%, SA-GCN's F1-micro is 72.2%, a cumulative decrease of 3.4% compared to the 80% setting, while the MAE remains at 0.086, a total increase of 0.011, and the MCC remains at 0.571, a total decrease of 0.040. Even under the extremely sparse 20% training set condition, SA-GCN's F1-micro remains at 66.5%, with an MAE of 0.104 and an MCC of 0.477. In comparison, within the entire reduction range from 80% to 20%, F1-micro saw a total reduction of 9.1%, while its total MAE increased by 0.029, and its total MCC decreased by 0.134, all significantly outperforming the corresponding changes of other baseline models. For example, TrustGNN's F1-micro decreased by 8.5% in the same range (from 74.4% to 65.9%), but its MAE increased by 0.036 (from 0.081 to 0.117), significantly higher than the error accumulation of SA-GCN; GATrust's MAE increase reached 0.041 (from 0.081 to 0.122), and GuardiaN's MAE increase reached 0.048, all indicating that the error increases rapidly in sparse scenarios, while SA-GCN showed the smallest error growth.

[0148] On the PGP dataset, the model's overall performance is higher, but the impact of sparsity remains significant. With 80% training set, SA-GCN achieves an F1-micro of 87.5%, MAE of 0.082, and MCC of 0.732. When the training set is reduced to 60%, the F1-micro only decreases to 86.9%, a drop of 0.6%, while the MAE increases to 0.087, an increase of only 0.005; the MCC remains almost stable (decreasing from 0.732 to 0.730). After reducing the training set to 40%, the F1-micro remains at 86.0%, a total decrease of 1.5% compared to the 80% setting; the MAE is 0.095, a total increase of 0.013; and the MCC is 0.723, a total decrease of 0.009. When the training set is further reduced to 20%, the F1-micro remains at 83.6%, with an MAE of 0.143 and an MCC of 0.682. Compared to the baseline model, within the same 80% to 40% range, TrustGNN and GATrust both saw a 2.0% decrease in F1-micro, TrustGuard saw an even greater decrease of 3.1%, while GuardiaN saw a 2.0% decrease. Meanwhile, the MAE increases of these models were 0.014, 0.016, 0.019, and 0.020, respectively, all significantly greater than SA-GCN's 0.013.

[0149] The experimental results above demonstrate the superior robustness of the SA-GCN model, stemming from the inherent advantages of its core architecture design. When training data is sparse, traditional models (such as TrustGNN and Guardian) heavily rely on the local structure of the graph for information propagation; once structural information is insufficient, performance deteriorates significantly. SA-GCN, however, exhibits strong anti-sparseness capabilities through its unique "calibration-interpretation-gain" process. The trust semantic calibration and semantic interpretation modules ensure that the model can extract more consistent and personalized semantic information from a limited and noisy initial rating, allowing each piece of known information to be utilized more efficiently. In particular, its trust semantic gain module overcomes the limitations of physical connections, finding implicit similarities for information-sparse nodes in the global trust feature space and aggregating their semantic features, achieving effective information gain. Therefore, SA-GCN not only surpasses baseline algorithms in overall performance but also demonstrates stronger stability and generalization ability in the face of the prevalent problem of data sparsity, indicating its enormous potential and value in practical applications.

[0150] 4. Ablation experiment: Table 4 shows the results of the comparative experiment between SA-GCN and three variants.

[0151] SA-GCN_α is a variant that removes the trust semantic calibration operation, meaning it does not perform semantic calibration on the initial rating features; SA-GCN_β is a variant that removes the trust semantic interpretation operation, meaning it does not consider the user's trust behavior patterns; SA-GCN_γ is a variant that removes the global trust semantic gain operation, relying only on trust interactions within the local neighborhood and the initial rating features for node representation learning, without using any global semantic enhancement signals. As shown in Table 4, the SA-GCN model exhibits optimal performance on all metrics and both datasets; removing any module leads to a decrease in model performance.

[0152] For SA-GCN_α with the trust semantic calibration module removed: on the Advogato dataset, F1-micro decreased from 75.6% to 74.0%, MAE increased from 0.075 to 0.078, and MCC decreased from 0.611 to 0.590. Similarly, on the PGP dataset, F1-micro decreased to 86.9%, MAE increased by 0.003 to 0.085, and MCC decreased from 0.732 to 0.725. This performance decline demonstrates the importance of the trust semantic calibration module. Without this module, the same rating score given by different users may represent drastically different levels of trust. The model directly processes initial rating features full of semantic bias, and the inconsistency in trust semantics is continuously amplified during the information propagation process of graph convolution, severely interfering with the learning of node features. This makes it difficult for the model to form a consistent understanding of the user's true trust level, ultimately resulting in a significant decrease in evaluation accuracy.

[0153] For SA-GCN_β, removing the trust semantic interpretation module has a slightly smaller impact on performance than the calibration module SA-GCN_α. On the Advogato dataset, F1-micro drops to 74.1%, MAE increases to 0.078, and MCC drops to 0.591; on the PGP dataset, F1-micro drops to 87.0%, MAE increases to 0.084, and MCC drops to 0.726. This indicates that ignoring users' personalized semantic expressions weakens SA-GCN's ability to express personalized trust features. The trust semantic interpretation module provides a deeper expression of user preference information beyond a single rating value. The lack of this personalized dimension results in a flattened understanding of users, especially limiting its ability to distinguish users with similar behavioral patterns but different trust intentions, thus leading to reduced evaluation accuracy.

[0154] The performance of SA-GCN_γ, with the trust semantic gain module removed, experienced the most dramatic decline. On the Advogato dataset, F1-micro dropped to 73.5%, MAE rose to 0.079, and MCC fell to 0.588. On the PGP dataset, its F1-micro also dropped to 85.9%, MAE rose to 0.087, and MCC fell to 0.720. This phenomenon profoundly reveals the important value of the trust semantic gain module in addressing the problem of trust semantic sparsity. Social networks commonly contain a large number of nodes with sparse or even isolated trust relationships. Traditional GCN models can only rely on their limited local neighbors for information aggregation, and the learned feature representations are often insufficient and biased. SA-GCN_γ degenerates to this state, and its significant performance drop, especially on the relatively sparse Advogato dataset, directly confirms the limitations of relying solely on local structural information. In stark contrast, the complete SA-GCN, through a global similarity aggregation mechanism, can overcome the physical connectivity limitations of the graph. It searches for and aggregates information from "implicitly similar" nodes in the feature space for nodes lacking this information, achieving cross-structural information gain. This provides robust information support for the feature representation of sparse nodes, significantly improving the model's robustness and generalization ability in sparse environments. Therefore, the trust semantic gain module is key to SA-GCN's superior performance in complex social network environments.

[0155] Based on the comprehensive results of the ablation experiments, we can conclude that the three core modules of SA-GCN—trust semantic calibration, trust semantic interpretation, and trust semantic gain—are not simply a collection of functions, but rather a complementary and organically collaborative whole. Trust semantic calibration cleans up the semantic bias of input features at the source, laying a solid foundation for subsequent processing; trust semantic interpretation adds a refined characterization of user personalized preferences on this basis; and the trust semantic gain module solves the bottleneck problem of information sparsity that limits model performance. The combined effect of these three modules enables SA-GCN to systematically address the challenges of evaluation inaccuracy caused by both user subjective cognitive bias and information sparsity, thus demonstrating excellent accuracy and robustness in trust evaluation tasks.

[0156] 5. The semantic calibration effect analysis is as follows: The rating deviation is used to characterize the degree of strictness of users in the trust rating process relative to the community average level, and is an important indicator for measuring subjective differences.

[0157] Figures 10 and 11 show the comparison of rating bias distribution. Both figures compare the changes in user rating bias distribution before and after model training, aiming to demonstrate the model's trust semantic calibration effect. The blue histogram represents the distribution of users' true rating bias based on the original rating data before model training; it reflects the original state of user subjectivity (strict or lenient) before model intervention. Before training, the bias is directly calculated based on the mean residual between the original rating and the average of the trusted users, which is an observational statistic that can reflect the user's true bias behavior at the data level.

[0158] In Figures 10 and 11, before training (blue histogram), the bias distribution of both datasets showed a wide shape and a clear "long tail," confirming the widespread existence of subjective cognitive bias among users. That is, the network contains a large number of "strict" or "lenient" users whose trust rating standards deviate significantly from the average level. This individual subjective bias introduces semantic noise into trust assessment.

[0159] The red histogram represents the distribution of user rating bias after calibration by the "Trust Semantic Calibration Module". The calibration module outputs rating calibration features in vector form after training. To quantify the calibration effect in the experiment, an auxiliary regression layer is introduced to map the high-dimensional rating calibration features back to a one-dimensional numerical space. The residuals are then calculated using the same benchmark to obtain an estimate of the bias under semantic consistency. Although the two methods differ in their calculation, both can characterize the degree of user deviation in ratings, and comparing the changes in their distribution can verify the calibration effect.

[0160] After calibration (red histogram), the distribution of user rating bias changed significantly: the distribution became taller and narrower, with the probability density more concentrated around the center point where the user rating bias was zero. Specifically, on the Advogato dataset, the original rating bias before calibration was scattered over approximately... Within a wide range, it exhibits a clear long-tailed distribution; however, after calibration, the deviation value is quickly pulled back and converges to near zero. Within a narrow interval. On the PGP dataset, the bias distribution also varies from... The wide-range oscillation pattern within the range has transformed into... The distribution exhibits a peak shape closely centered around zero. This centralization indicates that the SA-GCN model successfully learned and quantified the individual rating scales of different users, and adjusted the rating expression accordingly, effectively eliminating semantic noise introduced by individual subjective preferences. By unifying these heterogeneous and subjective rating scales onto a more comparable trust benchmark, the model lays a solid foundation for subsequent graph neural network learning of high-quality, semantically consistent node representations, thus verifying the important role of trust semantic calibration in solving the problem of inaccurate trust assessment.

[0161] 6. Hyperparameter Sensitivity Analysis: In the trust semantic gain module, the kernel size K is a key parameter, determining the range of user nodes with similar trust semantics that the model aggregates in the sorted trust feature space. This aggregation is not based on physical connections, but rather on a trust semantic approximation: the model first sorts all users according to their trust semantic features, and the K value defines the "window" size for information aggregation for each user in this sorting, i.e., how many semantically similar nodes to refer to before and after. Therefore, the K value determines the granularity of "grouping approximation" and directly affects the effect of trust semantic enhancement: if the K value is too small, it means that the "grouping" is too small, and nodes cannot obtain sufficient semantic supplementation from "same kind" nodes; if the K value is too large, the "grouping" is too large, which will introduce "noise" nodes with dissimilar trust patterns, thus diluting and interfering with the effective trust information. To explore the sensitivity of model performance to the K value and to verify the rationality of K=3 selected in this invention, a set of comparative experiments were conducted. The value of K was set to {1, 3, 5, 7}, and tests were conducted on two datasets, while keeping all other parameter settings consistent with the main experiment.

[0162] Figure 13 shows the performance variation curves of SA-GCN under different convolutional kernel sizes. As can be seen from Figure 13, on both the Advogato and PGP datasets, the model performance exhibits a similar and clear pattern with the change in K value: the model performance is worst when K=1. This is because the one-dimensional convolution at this point degenerates into a linear transformation of the features of a single node, completely losing the ability to aggregate semantic information of similar neighbors. This indicates that simply ranking the features of global nodes without effective local aggregation cannot achieve information gain. When K increases from 1 to 3, the model's F1-micro and MCC improve, while MAE decreases significantly. This shows that aggregating the information of the most similar nodes is crucial. Setting K=3 allows the model to capture the common features of semantically similar groups within a small and precise range, thereby effectively enhancing the representation of sparse nodes and verifying the effectiveness of the "ranking-convolution" mechanism. When K continues to increase from 3 to 5 and 7, the model performance shows a slight decrease or tends to plateau. This indicates that an excessively large convolutional kernel leads to an overly broad aggregation range. The model then begins to include nodes whose trust patterns are irrelevant to the current node's trust pattern in the trust feature space. The features of these nodes not only fail to contribute to the trust semantic gain but may also dilute and interfere with the trust information learned from the most similar nodes, resulting in a decrease in the final evaluation accuracy. This experiment demonstrates that the performance of SA-GCN is sensitive to the kernel size K. Both excessively small and excessively large K values ​​degrade model performance. Experimental results show that choosing K=3 is a reasonable trade-off; it fully utilizes global similarity information to alleviate data sparsity while avoiding noise from irrelevant information, thereby maximizing the accuracy of the trust semantic gain module.

[0163] This invention proposes a social network semantic alignment trust evaluation system based on graph convolutional networks. By executing a "calibration-interpretation-gain" process, it systematically solves the problem of inaccurate trust evaluation caused by both user subjective cognitive bias and information sparsity. Experiments demonstrate the following significant benefits: First, it significantly improves the accuracy of trust evaluation. This invention achieves more accurate prediction of trust relationships through the synergistic effect of three core modules. First, the trust semantic calibration module models the strictness of the trustor and the credibility of the trusted party, semantically aligning the initial ratings, which are highly subjective, mitigating the impact of semantic bias and providing more consistent and reliable input for model learning. Second, the trust semantic interpretation module deeply characterizes users' personalized rating habits, enabling the model to distinguish subtle differences such as "habitual positive reviewers" and "cautious high scorers," achieving a refined understanding of users' trust intentions. Finally, the trust semantic gain module, through global information aggregation, supplements semantically deficient nodes with key trust features, avoiding prediction bias caused by insufficient information. Experimental results show that on the two public datasets, Advogato and PGP, the model of this invention outperforms all baseline models, including TrustGNN and GATrust, in all three core metrics: F1-micro, MAE, and MCC. For example, on the Advogato dataset, the F1-micro score reaches 75.6%, and the MAE is as low as 0.075, demonstrating its excellent evaluation accuracy.

[0164] Secondly, the robustness and generalization ability of the model in sparse scenarios are enhanced: online social networks generally suffer from sparse trust information, which severely degrades the performance of traditional models. The global trust semantic gain mechanism designed in this invention effectively overcomes this bottleneck. This mechanism breaks through the limitation of traditional graph neural networks that rely solely on local neighbors. Through a "sorting-convolution" operation, it finds and aggregates implicit similar information for information-sparse nodes in the global trust feature space, achieving cross-structure semantic information supplementation. Robustness comparison experiments show that when the training data decreases from 80% to 20%, the performance decline of the model of this invention is smaller than that of other baseline models. This indicates that even under the harsh conditions of limited available information, this invention can still maintain robust predictive ability and has stronger generalization and practical application value.

[0165] Finally, a deep modeling of user trust subjectivity is achieved: existing technologies are insufficient in modeling trust subjectivity, mostly ignoring the heterogeneity of rating semantics. This invention not only theoretically points out this problem but also provides an effective solution through a trust semantic calibration module. The calibration effect analysis experiment intuitively demonstrates this: after model calibration, the distribution of user rating bias becomes more concentrated, and extreme bias values ​​are effectively corrected and adjusted to the central region. For example, on the Advogato dataset, the user rating bias range is significantly reduced from before calibration. Largely converged to the calibrated On the PGP dataset, this interval also starts from... Significantly reduced to This significant convergence of distributions indicates that the model effectively eliminates semantic noise introduced by individual subjective preferences, unifies the rating scales of different users onto a more comparable benchmark, and fundamentally solves the problem of inaccurate evaluation caused by subjective cognitive biases.

[0166] In summary, through the organic synergy of three major mechanisms, this invention not only sets a new technical benchmark in terms of assessment accuracy, but also achieves substantial breakthroughs in robustness and addressing core challenges in the field of trust assessment, demonstrating its enormous application potential and value in real and complex online social network environments.

[0167] The above-described embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention, and should all be included within the protection scope of the present invention.

Claims

1. A semantic alignment trust evaluation system for social networks based on graph convolutional networks, wherein the online social network includes multiple user nodes and some known trust relationships, each trust relationship is represented as a weighted directed edge from the trustor to the trusted party, and the weight represents the trust rating, characterized in that, The system includes: a trust information encoding module, used to encode the graph structure information of the online social network into initial features of each user node, and to encode the ratings of known trust relationships into initial rating features of corresponding directed edges; a trust semantic calibration module, used to generate calibration factors for each user node based on the initial features, the user node's trust neighborhood and trusted neighborhood, and to adjust the initial rating features of edges using the calibration factors to obtain calibrated rating semantic features; and a trust semantic interpretation module, used to generate trustee role features by combining the user's individual rating statistical features with neighborhood information from out-edge graph convolution aggregation; the trustee role features are at least used in the table The system comprises: a personalized trust expression feature for users acting as trustors; a trust semantic gain module for generating trusted party role features based on calibrated rating semantic features, and combining these features with the trustor role features to form a dual-role trust feature; the trusted party role features are used to characterize at least the personalized trust expression feature of the user acting as a trusted party; convolutional enhancement processing based on global semantic ranking is applied to the dual-role trust feature to generate the final trust feature of the user node; and a trust evaluation module for inputting the concatenated final trust features into a multilayer perceptron for any pair of users to be predicted, and outputting the predicted trust rating to supplement the missing trust relationships in the online social network.

2. The social network semantic alignment trust evaluation system based on graph convolutional networks according to claim 1, characterized in that, Obtaining the calibration factor includes: obtaining the out-degree neighbor set of the trustee node and its initial features and initial rating features, and calculating the rating strictness vector through out-edge directed graph convolution and nonlinear transformation; the trustee node is a user node within the trust neighborhood, and the trust neighborhood consists of the user's trusted neighbors and the trust relationships pointing to the neighbors; obtaining the in-degree neighbor set of the trusted node and its initial features and initial rating features, and calculating the credibility vector through in-edge directed graph convolution and nonlinear transformation; the trusted node is a user node within the trusted neighborhood, and the trusted neighborhood consists of the trusted user's neighbors and the trust relationships pointing to the user; concatenating the rating strictness vector and the credibility vector, and inputting them into a multilayer perceptron to generate the calibration factor.

3. The social network semantic alignment trust evaluation system based on graph convolutional networks according to claim 1, characterized in that, Obtaining the trustee role features includes: calculating the mean and rating fluctuation of all out-degree neighbors of the user node; combining the mean rating, rating fluctuation, and node out-degree; and mapping them through a multilayer perceptron to form individual semantic expression features; performing edge graph convolution aggregation on the user node and aggregating neighbor information to obtain the social semantic expression features of the user node as a trustee role; and using a gated feature fusion mechanism to determine the weights of the social semantic expression features and the individual semantic expression features respectively, and performing weighted fusion to form the trustee role features of the user node.

4. The social network semantic alignment trust evaluation system based on graph convolutional networks according to claim 1, characterized in that, Obtaining the dual-role trust features includes: performing an in-edge graph convolution operation on the calibrated rating semantic features of the user node to generate a trusted party role feature; and concatenating the trusted party role feature with the trustor role feature to obtain the dual-role trust features.

5. The social network semantic alignment trust evaluation system based on graph convolutional networks according to claim 1, characterized in that, The dual-role trust features are subjected to convolutional enhancement processing based on global semantic ranking to generate the final trust features of user nodes. This includes: linearly mapping the dual-role trust features to obtain the semantic ranking score of each node; globally ranking all nodes according to the semantic ranking score to obtain a ranked trust feature vector sequence and a corresponding global ranking index; arranging the feature vectors of each node in order based on the global ranking index to construct a global node representation matrix; applying a one-dimensional convolution kernel to the ranking dimension of the global node representation matrix to perform a sliding window operation to extract and aggregate node features with similar trust patterns; using the inverse mapping of the global ranking index, restoring the features output by the convolution operation to the original node order to obtain trust semantic enhancement features; and performing a residual connection between the trust semantic enhancement features of the same user node and the dual-role trust features to generate the final trust features.

Images