Intelligent decision-making method and system based on neural network

Abstract

The invention discloses an intelligent decision-making method and system based on a neural network, and relates to data processing. The method comprises the steps of obtaining place data and corresponding attribute data of a to-be-decided project; establishing a fuzzy relation matrix between the places and the attributes; according to the fuzzy relation matrix, generating a network model reflecting trust relation strength among users through trust propagation operation of a graph neural network; according to the network model, a community structure containing the community overlapping degree is identified through a community discovery algorithm; calculating the influence weight of a decision maker according to the trust relationship strength between the users and the community overlapping degree; and forming a group consensus through a robust optimization method according to the influence weight and the community structure, and generating a decision result of the project to be decided. The traditional project decision often neglects the trust relationship among the users, and the community structure including the overlapping degree is identified by combining the egret optimization algorithm and the improved K-means clustering, and the influence weight is calculated based on the trust intensity, so that the scientificity and fairness of the decision are improved.

Description

Technical Field

[0001] This application relates to data processing, and particularly to an intelligent decision-making method and system based on a neural network. Background Art

[0002] With the in-depth promotion of the construction of digital government, government public decision-making projects are increasingly showing the characteristics of diversification and complexity. Government departments need to make scientific, reasonable and widely recognized decisions in multiple fields such as urban planning, infrastructure construction, and public resource allocation. Traditional government decision-making often relies on expert evaluation, public surveys and administrative procedures. In the face of a complex and changing social environment, these methods are difficult to fully capture and process the complex relationship network among stakeholders, resulting in a gap between the decision-making effect and public expectations.

[0003] In recent years, the application of artificial intelligence and big data technologies in the field of decision-making assistance has become increasingly widespread. Data-driven decision-making models can process massive amounts of information, discover potential laws, and provide an objective basis for government public projects. However, most existing intelligent decision-making systems focus on the analysis and processing of data itself, ignoring the important influence of the interpersonal network structure in the decision-making process, especially the key role of the trust relationship among users in decision-making recognition and implementation effect.

[0004] However, traditional group decision-making methods usually assume that decision-makers are independent of each other, or only consider simple weight allocation, ignoring the trust relationship network structure among decision-makers. In fact, in projects such as government procurement, urban renewal, and public facility location, there are complex trust relationship networks among various stakeholders (such as government departments, experts and scholars, community representatives, and enterprise units), and these relationships have an important impact on opinion formation and dissemination.

[0005] Existing community discovery algorithms have limitations in dealing with government decision-making networks. Stakeholders involved in government decision-making often belong to multiple communities at the same time (such as being both professional and technical personnel and community resident representatives). Traditional community discovery algorithms are difficult to identify this overlap, resulting in a one-sided understanding of the community structure and being unable to accurately grasp the dynamic process of opinion dissemination and consensus formation.

[0006] When existing group decision-making support systems integrate multiple opinions, they mostly adopt simple weighted average or voting mechanisms, lacking the ability to robustly handle the uncertainty of the decision-making environment. In government public project decision-making, in the face of the diverse needs of different regions and different groups, simple decision-making mechanisms are difficult to form a group consensus that takes into account fairness and representativeness. Summary of the Invention

[0007] In view of the fact that traditional project decision-making often ignores the trust relationship among users, this application provides an intelligent decision-making method and system based on neural networks. By means of the trust propagation mechanism of graph neural networks, a trust relationship network among users is constructed. The heron optimization algorithm and improved K-means clustering are combined to identify the community structure with overlap degree, and the influence weight is calculated based on the trust strength and community attributes. Finally, a more representative group consensus is formed through a robust optimization method, providing accurate decision-making support for the project to be decided.

[0008] One aspect of this application provides an intelligent decision-making method based on neural networks, including: obtaining the location data and corresponding attribute data of the project to be decided; establishing a fuzzy relationship matrix between the location and the attributes according to the location data and the attribute data; generating a network model reflecting the trust relationship strength among users through the trust propagation operation of graph neural networks according to the fuzzy relationship matrix; identifying the community structure with community overlap degree through a community discovery algorithm according to the network model; calculating the influence weight of the decision maker according to the trust relationship strength among users and the community overlap degree; forming a group consensus through a robust optimization method according to the influence weight and the community structure, and generating the decision result of the project to be decided.

[0009] Among them, the project to be decided has clear geographical location attributes and time attributes, and requires multiple stakeholders to participate in the decision-making process. For example, the site selection of newly built public facilities (such as hospitals, schools, cultural centers, etc.); the evaluation of the renovation plan of the old urban area in the urban renewal project; the layout optimization of transportation infrastructure (such as bus stops, subway stations); the location planning of affordable housing projects; the comprehensive evaluation of multiple alternative plans in government procurement; the balanced decision-making between ecological environment protection and industrial development, etc. The project to be decided contains location data elements, can form an alternative plan set, has time elements related to dates, can be comprehensively evaluated from income-type and cost-type indicators, involves multiple decision-making subjects, has a complex trust relationship network, and the decision result has different impacts on different community groups.

[0010] Further, establishing a fuzzy relationship matrix between the location and the attributes includes: collecting an alternative plan set composed of place names according to the preset place name elements; collecting an attribute set composed of dates according to the preset time elements; constructing an initial relationship matrix according to each location element in the alternative plan set and each time element in the attribute set; using fuzzy numbers to represent the indicators of each element in the initial relationship matrix, and the indicators include income-type indicators and cost-type indicators; where the income-type indicator represents an indicator that the larger the original value is, the more beneficial it is to the decision-making, and the cost-type indicator represents an indicator that the smaller the original value is, the more beneficial it is to the decision-making; through normalization processing, the indicators are converted into membership degree values; and a fuzzy relationship matrix is constructed according to the membership degree values.

[0011] Further, through normalization processing, converting the indicators into membership degree values includes: for income-type indicators: ; For cost - type indicators: ; Among them, represents the evaluation value of the k - th location under the j - th attribute after normalization, represents the relationship between the j - th attribute and the k - th type of location, j represents the j - th day of the week, and k represents the k - th type of location in the alternative solutions; represents the minimum value of the j - th attribute among all alternative locations; represents the maximum value of the j - th attribute among all alternative locations.

[0012] Further, a network model reflecting the strength of trust relationships among users is generated through the trust propagation operation of the graph neural network, including: comprehensively scoring alternative locations according to the fuzzy relationship matrix, and selecting locations with scores greater than the threshold to form a sample set; constructing an initial interaction network G containing user nodes and location nodes based on the sample set; performing feature transformation on the nodes and edges in the initial interaction network G to obtain network features in the same representation space; obtaining the propagated trust features between nodes according to the network features through the trust propagation operation of the graph neural network; and obtaining a network model reflecting the strength of trust relationships among users through information aggregation based on the propagated trust features between nodes.

[0013] Further, obtaining network features in the same representation space includes: through linear transformation, projecting all nodes in the initial interaction network G into the same - dimensional space to obtain the transformed nodes , and the projection formula is: , where is the original attribute vector of node v, is the transformed node representation, is the learnable parameter matrix; through parameterization, performing feature transformation on different types of edges in the initial interaction network G to obtain the transformed edges : , where is the original attribute vector of the i - th type of edge, is the parameter matrix specific to this type of edge, is the transformed edge representation; initializing the missing node and edge attributes in the initial interaction network G as random vectors, and using the random vectors as the learnable parameters of the graph neural network; taking the transformed nodes , and the transformed edges , as network features in the same representation space.

[0014] Further, obtaining the propagated trust features between nodes includes: setting a trust chain in the graph neural network, and presetting a hyperparameter K as the maximum propagation length of the trust chain to prevent information propagation distortion caused by overly long trust chains; among them, the trust chain includes the trustee node and the delegator node; according to the nodes And edge , calculate the information received by the recipient node v on the trust chain of length k , , where is the representation of the starting node of the trust chain, to are the representations of each edge on the trust chain, represents the Hadamard product operation, where , represents the complex plane, modulus ; for each recipient node , aggregate the information received by the recipient node on different types of trust chains to generate the propagation trust feature between nodes : , where is the original representation of node v, is the sum of all information received by node v on a specific type of trust chain, is the learnable parameter matrix of this type of trust chain.

[0015] Furthermore, obtain a network model reflecting the strength of the trust relationship between users, including: calculating the weights of different types of trust chains ; according to the weights , calculate the embedding representations Z of the recipient node and the delegator node respectively, : , ; according to the embedding representations Z and , through the aggregation representation, obtain the final node representation , , where W is the learnable connection weight matrix, and "||" represents the connection operation; according to the final node representation , construct a network model reflecting the strength of the trust relationship between users.

[0016] Furthermore, calculate the weights of different types of trust chains , using the following formula: , ; where is the total number of nodes, is the learnable query vector, is the attention network parameter, is the propagation trust feature of different types, and b is the bias term.

[0017] Furthermore, a community structure including community overlap is identified through a community discovery algorithm, which includes: partitioning the network model using the Shrike Optimization Algorithm (SBOA) to obtain a community partitioning scheme; clustering the community partitioning scheme using the Manhattan distance-based K-means clustering algorithm to obtain community centers; calculating the community overlap of each decision maker by identifying the overlapping parts where the decision maker belongs to multiple communities; and generating a community structure including community overlap based on the community centers and community overlap.

[0018] Furthermore, obtaining the community partitioning scheme includes: randomly partitioning N decision makers according to a preset community quantity Q value to generate a community partitioning matrix C of size N×N, applying a rounding function to each element in C to ensure that each decision maker is assigned a fixed community number; calculating the initial group consensus level based on the community partitioning matrix C and the network model matrix T reflecting the strength of trust relationships among users:; recording the initial community partitioning matrix C and its corresponding GCL value as the current optimal solution and. Apply the rounding function to process so that , ensuring that each decision maker is assigned to a fixed community number; calculating the initial group consensus level according to the community partitioning matrix C and the network model matrix T reflecting the strength of trust relationships among users : ; recording the initial community partitioning matrix C and its corresponding GCL value as the current optimal solution and .

[0019] Iteratively update the community partitioning matrix according to the following steps:

[0020] (a) Hunting for prey: Generate a candidate solution using the formula, where t is the current iteration number, T is the maximum iteration number, and and are two randomly selected existing partitioning schemes, and is a random vector between 0 and 1; where t is the current iteration number, T is the maximum iteration number, and are two randomly selected existing partitioning schemes, is a random vector between 0 and 1;

[0021] (b) Consuming the prey: Based on the result of the previous step, generate a new candidate solution using the formula, where b is a control parameter and r is a random number; where b is a control parameter, r is a random number;

[0022] (c) Attacking the prey: Based on the result of the previous step, generate the final candidate solution using the formula, where β is a scaling factor and RL is a perturbation vector generated based on Levy flight; where β is a scaling factor, RL is a perturbation vector generated based on Levy flight;

[0023] Apply the rounding function to the candidate solution generated in each step to ensure the validity of the community number and calculate its corresponding GCL value; if the value corresponding to the candidate solution is greater than the current optimal value, then update, corresponding to value is greater than the current optimal value , then update , ; When the preset 100 iterations are completed or the GCL improvement in 10 consecutive iterations does not exceed the preset threshold ε (ε = 0.001), the hunting phase ends, and is used as the optimal solution in the hunting phase; Based on the optimal solution in the hunting phase, the formula is applied for further optimization, where ζ is a random integer of 1 or 2, is a random vector conforming to the normal distribution, is a newly generated partitioning scheme randomly; The rounding function is applied to the candidate solution generated in the escape phase and its GCL value is calculated. If it is better than the current optimal solution, is updated; Repeat the optimization in the escape phase 100 times or until the GCL value improvement in 10 consecutive times does not exceed the threshold ε; The final is used as the community partitioning scheme for subsequent optimization of the community center using the K-means clustering algorithm based on the Manhattan distance.

[0024] Furthermore, obtaining the community center includes: calculating the opinion distance between decision-makers using the improved Manhattan distance , , where, and represent the opinion vectors of decision-makers i and j respectively; Calculate the opinion distance matrix D between any two decision-makers in the community partitioning scheme; According to the trust relationship T between decision-makers, calculate the comprehensive distance , , where, represents the degree of trust of the j-th expert in the i-th expert, and M represents the number of decision-makers in the community; According to the comprehensive distance , calculate the average comprehensive distance from each decision-maker to other decision-makers in the community, and select the decision-maker with the smallest average comprehensive distance as the center point of the corresponding community; According to the center point, reassign each non-center decision-maker to the community with the closest comprehensive distance; Repeat the determination of the center point and the assignment of decision-makers until the position of the community center point is stable or the preset number of iterations is reached, and output the community center obtained by clustering.

[0025] Furthermore, calculating the community overlap degree of each decision-maker includes: calculating the comprehensive distance between each decision-maker i and each community center point j according to the determined community center point; Calculate the membership coefficient according to , where Q is the number of communities; Set the community overlap determination threshold θ. When the membership degree of the decision-maker to a certain community is met, it is considered that the decision-maker partially belongs to this community; Calculate the community overlap degree of each decision-maker, where, It is an indicator function that takes the value of 1 when the condition is met and 0 otherwise.

[0026] Preferably, generating the decision result of the item to be decided includes: constructing a trust weight influence function based on the degree of mutual trust among experts. Let the trust influence degree of the i-th expert be: Where, represents the degree of trust of the -th expert in the -th expert, and represents the overall trust level obtained by the

[0027] -th expert in the group. Considering the degree of structural overlap of experts in the expert community network, construct an overlapping influence weight function. The community structure influence degree of the -th expert is expressed as: Where, represents the degree of overlap between the expert community where the i-th expert is located and other communities, is a regulation parameter used to balance the influence degree of the overlap degree on the weight. This value reflects the decision-making reference value brought by the activity degree of experts in multiple functional communities.

[0028] Combining the trust influence degree and the structural influence degree, calculate the preliminary weight of the i-th expert: .

[0029] Normalize the preliminary weights of all experts to obtain the set of expert weights finally used in the project decision-making system: Where, represents the normalized weight of the i-th expert, which is used in subsequent modules such as user behavior evaluation, credit score calculation, or intelligent scheduling recommendation, etc., to improve the accuracy and stability of system judgment.

[0030] Promote the realization of group consensus by means of trust relationships, and on the premise of fully considering robustness and fairness, optimize the consensus reaching process through an optimization algorithm to generate the decision information of the placement location of the item to be decided, including: introducing the influence of the trust relationship and overlapping community structure among experts on the consensus process, constructing a weight adjustment factor, and realizing dynamic weight adjustment among multiple agents by combining individual opinions and trust values of experts. By integrating the information bridge effect, improve the stability and accuracy of consensus formation in a complex environment, Where, use to represent the opinion of decision maker i, to represent the opinion of decision maker i after adjustment and modification, to represent the final collective opinion, to represent the degree of trust of expert i in the overall group, and T represents the overall trust level of all current experts. Denote the consensus enhancement factor when the expert is in the overlapping community structure, and its definition is as follows: , where is the number of sub - communities to which expert i belongs. This factor reflects the "information bridge" effect brought by the degree of participation of experts in multiple sub - communities, which helps to improve their group integration degree.

[0031] The group consensus level is obtained by the average of the individual consensus degrees of all experts, and the calculation formula is as follows: Calculate the unit opinion adjustment cost of the decision - maker to optimize the fairness utility of the group. For a certain decision - maker The unit opinion adjustment cost can be calculated by the formula, using to represent the opinion adjustment cost paid to the decision - maker, and the decision - maker The calculation of the fairness utility is as follows:

[0032] ,

[0033] ,

[0034] ,

[0035] where represents the initial opinion of the decision - maker , represents the opinion of the decision - maker after adjustment, the envy preference coefficient , and the sympathy / pride preference coefficient .

[0036] Control the formula cost under uncertain costs to maximize the group fairness utility level. The constructed optimization equation is as follows:

[0037] .

[0038] In the above optimization equation, the interference for the cost can be replaced by three uncertainty sets and the optimization equations for different robust groups are given.

[0039] The box - type uncertainty set is defined as follows: . Use the box - type uncertainty set to simulate the unit opinion adjustment cost paid by the adjuster to the decision - maker. Rewrite the optimization adjustment equation under the given fairness effect as follows:

[0040]

[0041] The ellipsoidal uncertainty set is defined as follows:

[0042] 。

[0043] Use an ellipsoidal uncertainty set to simulate the unit opinion adjustment cost that the adjuster pays to the decision maker. The optimization adjustment equation under the given fairness effect is rewritten as follows:

[0044]

[0045] Polyhedral uncertainty set is defined as follows:

[0046] 。

[0047] Use a box uncertainty set to simulate the unit opinion adjustment cost that the adjuster pays to the decision maker. The optimization adjustment equation under the given fairness effect is rewritten as follows:

[0048] 。

[0049] Another aspect of the present application also provides an intelligent decision-making system based on a neural network, including a collection module for obtaining location data and corresponding attribute data of the item to be decided; a fuzzy relationship module for constructing a fuzzy relationship matrix according to the location data and the attribute data; a network model module for generating a network model reflecting the strength of the trust relationship between users through graph neural network trust propagation operations according to the fuzzy relationship matrix; a community recognition module for partitioning the network model by using the goshawk algorithm and an improved K-means clustering algorithm, and identifying a community structure including community overlap; and a decision-making module for calculating the influence weight of the decision maker according to the strength of the trust relationship between users and the community overlap, and forming a group consensus through a robust optimization algorithm according to the influence weight, and generating a decision result of the item to be decided.

[0050] Compared with the prior art, the advantages of the present application are as follows:

[0051] When dealing with user preferences and regional planning, traditional decision-making methods for item placement to be decided generally adopt static statistical analysis and simple voting mechanisms, but there are defects such as the neglect of the trust relationship between users, the single description of the community structure, and the inaccurate quantification of the influence of the decision maker, resulting in difficulty in forming a truly representative consensus in group decision-making. The present application constructs a dynamic trust relationship network through a graph neural network, and adopts a parameterized trust chain propagation mechanism to capture complex user mutual trust patterns; identifies a community structure with overlap through an improved goshawk optimization algorithm and Manhattan distance K-means clustering, which more accurately reflects the multi-community belonging characteristics of the decision maker; calculates the influence weight based on the trust strength and community overlap to achieve accurate quantification of the decision-making influence; and finally forms a more representative group consensus through a robust optimization method, greatly improving the accuracy, adaptability and user satisfaction of the decision-making for the item to be decided. Brief Description of the Drawings

[0052] This application will be further described by way of exemplary embodiments, which will be described in detail with reference to the accompanying drawings. These embodiments are not restrictive. In these embodiments, the same numbers represent the same structures, where:

[0053] Figure 1 is a flowchart of an intelligent decision-making method based on a neural network provided in Embodiment 1 of this application;

[0054] Figure 2 is a flowchart of generating a fuzzy relation set;

[0055] Figure 3 is a flowchart of identifying key communities;

[0056] Figure 4 is a flowchart of learnable and composable trust propagation based on a graph neural network in the decision-making method;

[0057] Figure 5 is a flowchart of an overlapping sub-community detection method based on the combination of three-way clustering and K-means clustering in the decision-making method;

[0058] Figure 6 is a flowchart of calculating weights under the influence of trust and overlapping communities in the decision-making method;

[0059] Figure 7 is a flowchart of generating decision information in the decision-making method;

[0060] Figure 8 is a module diagram of an intelligent decision-making system based on a neural network provided in Embodiment 2 of this application. Detailed Embodiments

[0061] The methods and systems provided in the embodiments of this application will be described in detail below with reference to the accompanying drawings.

[0062] Embodiment 1

[0063] As Figure 1 shown, the method includes the following steps: obtaining location data and corresponding attribute data of the item to be decided; establishing a fuzzy relation matrix between the location and the attributes according to the location data and the attribute data; generating a network model reflecting the strength of the trust relationship between users through trust propagation operations of a graph neural network according to the fuzzy relation matrix; identifying a community structure including community overlap degree through a community discovery algorithm according to the network model; calculating the influence weight of the decision maker according to the strength of the trust relationship between users and the community overlap degree; forming a group consensus through a robust optimization method according to the influence weight and the community structure, and generating a decision result of the item to be decided.

[0064] In an embodiment of the present application, the decision-making item recommendation system can be expressed as , where is a positive integer; represents the attribute set, represents the first day of the week, represents the second day of the week, represents the third day of the week, represents the fourth day of the week, represents the fifth day of the week, represents the sixth day of the week, represents the seventh day of the week; represents the location set matching the attribute set, represents H101, represents H102, represents H103, represents H104, represents H105; represents the relationship set of brands and attributes given by different experts, where .

[0065] In an embodiment of the present application, the relationship set generated according to the attribute set and the item set can be represented by Tables 1 to 20:

[0066] Table 1 is the relationship set of the item set and the attribute set given by Expert 1

[0067]

[0068] Table 2 is the relationship set of the item set and the attribute set given by Expert 2

[0069]

[0070] Table 3 is the relationship set of the item set and the attribute set given by Expert 3

[0071]

[0072] Table 4 is the relationship set of the item set and the attribute set given by Expert 4

[0073]

[0074] Table 5 is the relationship set of the item set and the attribute set given by Expert 5

[0075]

[0076] Table 6 is the relationship set of the item set and the attribute set given by Expert 6

[0077]

[0078] Table 7 is the relationship set of the item set and the attribute set given by Expert 7

[0079]

[0080] Table 8 is the relationship set of the item set and the attribute set given by Expert 8

[0081]

[0082] Table 9 is the relationship set of the item set and the attribute set given by Expert 9

[0083]

[0084] Table 10 is the relationship set of the item set and the attribute set given by Expert 10

[0085]

[0086] Table 11 is the relationship set of the item set and the attribute set given by Expert 11

[0087]

[0088] Table 12 is the relationship set of the item set and the attribute set given by Expert 12

[0089]

[0090] Table 13 is the relationship set of the item set and the attribute set given by Expert 13

[0091]

[0092] Table 14 is the relationship set of the item set and the attribute set given by Expert 14

[0093]

[0094] Table 15 is the relationship set of the item set and the attribute set given by Expert 15

[0095]

[0096] Table 16 is the relationship set of the item set and the attribute set given by Expert 16

[0097]

[0098] Table 17 is the relationship set of the item set and the attribute set given by Expert 17

[0099]

[0100] Table 18 is the relationship set between the item set and the attribute set given by Expert 18

[0101]

[0102] Table 19 is the relationship set between the item set and the attribute set given by Expert 19

[0103]

[0104] Table 20 is the relationship set between the item set and the attribute set given by Expert 20

[0105]

[0106] In an embodiment of the present application, as Figure 2 shown, the fuzzy relationship set is calculated through the following steps: Convert the fuzzy numbers of the objective attribute indicators into the first membership degree and the first non-membership degree according to the linear normalization formula. The first membership degree represents profitability, and the second membership degree represents cost type. Wherein, the linear normalization formula is expressed as:

[0107]

[0108] Wherein, represents the evaluation value of the k-th location under the j-th attribute after the decision matrix is normalized, represents the relationship between the th attribute and the k-th type of location, represents the th day of the week, represents the k-th type of location in the alternative solutions; the profit type means that the larger the original variable, the better, and the cost type means that the smaller the original variable, the better. Convert the indicators into membership degree values, and construct a fuzzy relationship matrix according to the membership degree values.

[0109] In summary, the to-be-decided project management system constructed based on a series of steps such as the fuzzy relationship set can effectively handle the complex relationships between the to-be-decided projects and attributes, and solve the uncertainty problem in demand assessment. By converting the evaluation values obtained from the public data set into quantifiable values through scientific methods such as linear normalization, the relevant information of the to-be-decided projects in real life is presented in a clear and intuitive manner, and it can optimize the resource allocation of the to-be-decided projects in the city for urban planners.

[0110] Figure 3A flowchart for identifying key communities includes the following steps: Obtain the location set and attribute set of the item to be decided. Input the set of samples to be selected into the item management system to be decided for processing, and after processing, obtain a fuzzy relationship set that matches the set of samples to be selected. Perform learnable and composable trust propagation operations based on a graph neural network to generate an accurate user trust relationship model. Identify the key communities during the use of the item to be decided by combining the overlapping sub-community detection method of three-way clustering and K-means clustering.

[0111] In an embodiment of the present application, as Figure 4 shown, the user trust relationship model is generated through the following steps: The initial trust matrix among 20 decision-makers is shown in Table 21:

[0112] Table 21 Initial unfilled trust matrix

[0113]

[0114] Perform a linear transformation on all nodes to project them onto a unified dimension. Transform the nodes and edges through the following formula: ; ; where is the attribute vector of node , is the transformed representation of node , is the learnable parameter matrix, is the attribute vector of the i-th edge type, is the learnable parameter matrix specific to the i-th edge type, is the transformed representation of the i-th edge type. The processing of the attributes in the nodes and edges is performed using a method similar to RotatE, and their attributes are combined in the complex plane. The information received by node on the chain is calculated by the following formula: ; where , represents the complex plane, the modulus , and is the Hadamard (element-wise) product.

[0115] For node and the set chain of the th type, TrustGNN aggregates information through the following formula: , where is a mapping function indicating which type the trust chain is, is the th type of trust chain, and is specific to learnable parameter matrix. is the representation of the node of the j-th chain type . For the aggregated information , TrustGNN aggregates them with different weights. TrustGNN first transforms the chain type-specific representation through a non-linear transformation ( ), and then measures the score of the chain type-specific representation as the similarity between the transformed representation and the chain type-level attention vector . TrustGNN averages the scores of all chain type-specific node representations, and the calculation formula is as follows: , where is a learnable parameter matrix, is a bias vector, is the attention vector at the chain type level. All chain types share the parameters and . After that, the scores are normalized through .

[0116] TrustGNN aggregates these chain type-specific representations to obtain an embedding, and the calculation formula is as follows, , where is the node representation regarding the delegator role on the chain type. TrustGNN represents nodes from both the trustee and delegator aspects. Only aggregates information regarding the trustee role. Therefore, TrustGNN needs to calculate the embedding of the delegator role through the formula , where is the node representation regarding the delegator role on the chain type. To fully retain the information of the node as the trustee and the trustee, TrustGNN connects and through the formula to obtain the final embedding. Among them, is the concatenation operation,

[0117] Table 22 The completed trust matrix

[0118]

[0119] In an embodiment of the present application, as Figure 5 shown, the key community model is identified through the following steps: Use the goshawk optimization algorithm to partition the initial community:​

[0120] Initial community division: Randomly divide the decision-makers according to the determined value of k. At this time, take , , where both i and j are equal to the number of decision-makers. At this time, generate an N*N random vector. Round each number in the initialized result vector to ensure that each decision-maker falls into the specified community division.

[0121] Group consensus level: Calculate the average consensus level of each division result in the above random vector, and give the average consensus level of different initial communities. .

[0122] Update in the hunting stage: Use the three operations with the same time interval shown by the heron in the hunting stage to update the community division in turn. For each division plan in the initialized community division result C, simulate the three behaviors of the heron in the hunting stage in turn. Update the community division matrix, and round each result to ensure that each decision-maker falls into the required community division. Calculate the new average consensus level using the updated result. .

[0123] Compare the average consensus level of different communities after the update with the original average consensus level. If the updated community division has a better average group consensus level, replace the original result with the updated result. Repeat this step for 100 times for the division matrix.

[0124] Update in the escape stage: Continue to update the community division by imitating the escape strategy of the heron when it observes that there are predators around posing a threat to it. Further optimize the result, update the above matrix and round each result. After the decision-makers fall into the required community division result, record it in the result matrix. Calculate the new average consensus level using the updated result. .

[0125] Compare the average consensus level of different communities after the update with the original average consensus level. If the updated community division has a better average group consensus level, replace the original result with the updated result.

[0126] Repeat this step for 100 times for the division matrix. The initial clustering center division result is shown in Table 23 below:

[0127] Table 23 Initial clustering center

[0128]

[0129] Using Manhattan distance to improve the community center confirmation of K-means clustering: In this paper, the improved Manhattan distance is used to calculate the distance between two adjacent opinions. Assume that the opinions of the two decision makers are , The distance between the two is calculated as follows, , for the opinions between any two decision makers within the divided community results, the distance D is calculated, and M represents the number of decision makers in the subcommunity.

[0130]

[0131] The above distance matrix and the trust relationship T between decision makers are used to give the comprehensive distance DT between decision maker i and other people. The calculation formula of DT is as follows: ,in, Indicates Experts on The trust level of each expert is calculated by the distance calculation method above to calculate the distance between a decision maker and other decision makers, and the decision maker with the smallest distance is selected as the current cluster center. Then iterate until the cluster center no longer changes. The final cluster center is determined as shown in the following table:

[0132] Table 24 Final cluster centers

[0133]

[0134] The upper and lower bounds of the Heron Eagle optimization algorithm are selected and the overlapping parts are determined: the division parameters UB and LB of the core domain and the boundary domain in the three clusters are initialized, and the upper and lower bounds when the current group consensus is the highest are determined through the measurement of group consensus.

[0135] Then imitate the hunting stage of the egret eagle to update the parameters. Use the above formulas to update the upper and lower limit parameters in turn and save the parameters for the better division results. Imitate the escape process of the egret eagle to update the parameters again. If the formula is used to update the parameters and a better effect is obtained, save the parameters. Use the determined parameters to divide the community and give the final community division results. The final community division results are shown in Table 25 below:

[0136] Table 25 Clustering results display

[0137]

[0138] The weight of each decision maker is determined based on the analysis of the impact of trust factors and overlapping community factors on decision-making. The realization of group consensus is promoted through trust relationships, robustness and fairness are fully considered, and the consensus-reaching process is optimized by applying optimization algorithms to generate decision information for the items to be decided.

[0139] In one embodiment of the present application, as Figure 6 shown, the weights under the influence of trust and overlapping communities are calculated through the following steps: Based on the degree of mutual trust among experts, a trust weight influence function is constructed. Let the trust influence degree of the i-th expert be: where represents the degree of trust of the -th expert in the -th expert, and represents the total number of experts. This value reflects the overall trust level obtained by the -th expert in the group.

[0140] Considering the degree of structural overlap of experts in the expert community network, an overlapping influence weight function is constructed. The community structure influence degree of the -th expert is expressed as: where represents the degree of overlap between the expert community where the i-th expert is located and other communities, and is a regulation parameter used to balance the influence degree of the overlap degree on the weight. This value reflects the decision-making reference value brought by the activity degree of experts in multiple functional communities.

[0141] Combining the trust influence degree and the structural influence degree, the preliminary weight of the i-th expert is calculated: .

[0142] Normalize the preliminary weights of all experts to obtain the set of expert weights finally used in the project decision-making system: where represents the normalized weight of the i-th expert, which is used in subsequent modules such as user behavior evaluation, credit score calculation, or intelligent scheduling recommendation, etc., to improve the accuracy and stability of system judgment.

[0143] According to the complete trust degree matrix calculated in the above steps and the results of the community detection of the decision-makers, the weights of different decision-makers are as shown in Table 26 below:

[0144] Table 26 Calculation Results of Expert Weights

[0145]

[0146] In one embodiment of the present application, as Figure 7 shown, the group consensus under the influence of trust relationships is achieved through the following steps, and the optimized consensus process under the fair consideration of robust groups is carried out: Introduce the influence of the trust relationship and overlapping community structure among experts on the consensus process, construct a weight adjustment factor, and realize the dynamic weight adjustment among multiple agents by combining the individual opinions of experts and trust values. By integrating the information bridge effect, the stability and accuracy of consensus formation in a complex environment are improved.

[0147]

[0148] Among them, use to represent the opinion of decision maker i, to represent the opinion of decision maker i after adjustment and modification, to represent the final collective opinion, to represent the level of trust of expert i in the overall group, and T represents the overall trust level of all current experts. to represent the consensus enhancement factor when an expert is in an overlapping community structure, and its definition is as follows: , where is the number of sub-communities to which expert i belongs. This factor reflects the "information bridge" effect brought by the degree of participation of experts in multiple sub-communities, which helps to improve their group integration degree.

[0149] The group consensus level is obtained by taking the average of the consensus degrees of all expert individuals, and the calculation formula is as follows: Calculate the unit opinion adjustment cost of the decision maker to optimize the fairness utility of the group. For a certain decision maker Unit opinion adjustment cost can be calculated by the formula, using to represent the opinion adjustment cost paid to the decision maker, and the decision maker The calculation of fairness utility is as follows:

[0150] ,

[0151] ,

[0152] ,

[0153] Among them, represents the initial opinion of decision maker , represents the opinion of decision maker after adjustment, the envy preference coefficient , the sympathy / pride preference coefficient .

[0154] Control the formula cost under uncertain costs to maximize the group fairness utility level. The constructed optimization equation is as follows:

[0155] .

[0156] In the above optimization equation, the interference with the cost can be replaced by three uncertain sets and the optimization equations for different robust groups are given.

[0157] Box uncertain set is defined as follows:

[0158] .

[0159] Use the box uncertainty set to simulate the unit opinion adjustment cost paid by the adjuster to the decision maker. The optimization adjustment equation under the given fairness effect is rewritten as follows:

[0160]

[0161] Ellipsoidal uncertainty set is defined as follows: .

[0162] Use the ellipsoidal uncertainty set to simulate the unit opinion adjustment cost paid by the adjuster to the decision maker. The optimization adjustment equation under the given fairness effect is rewritten as follows:

[0163]

[0164] Polyhedral uncertainty set is defined as follows: .

[0165] Use the box uncertainty set to simulate the unit opinion adjustment cost paid by the adjuster to the decision maker. The optimization adjustment equation under the given fairness effect is rewritten as follows:

[0166] .

[0167] In the first-stage CRP, optimize and adjust within each sub-community to reach a consensus among the sub-groups and select the adjustment plan with the minimum opinion adjustment payment cost as the final result. For the collective opinion result after optimizing and adjusting the first community to reach a consensus, see Table 27:

[0168] Table 27 Display of Collective Opinion Results of Community 1

[0169]

[0170] For the collective opinion result after optimizing and adjusting the second community to reach a consensus, see Table 28:

[0171] Table 28 Display of Collective Opinion Results of Community 2

[0172]

[0173] For the collective opinion result after optimizing and adjusting the third community to reach a formula, see Table 29:

[0174] Table 29 Display of Collective Opinion Results of Community 3

[0175]

[0176] In the second stage, directly take the average of the collective opinions of the three groups as the final overall opinion.

[0177] The final score results after comprehensive calculation in the second stage are shown in Table 30:

[0178] Table 30 Display of the Final Collective Opinion Results

[0179]

[0180] Example 2

[0181] To implement the above embodiments, as Figure 8 shown, to implement the above embodiments, the embodiments of the present application propose an intelligent decision-making system based on a neural network, including: an acquisition module that obtains location data and corresponding attribute data of the item to be decided; a fuzzy relationship module that constructs a fuzzy relationship matrix according to the location data and attribute data; a network model module that generates a network model reflecting the strength of the trust relationship among users through the trust propagation operation of the graph neural network according to the fuzzy relationship matrix; a community recognition module that divides the network model by using the goshawk algorithm and an improved K-means clustering algorithm, and identifies the community structure including the community overlap degree; a decision-making module that calculates the influence weight of the decision maker according to the strength of the trust relationship among users and the community overlap degree, and forms a group consensus through a robust optimization algorithm according to the influence weight, and generates a decision result of the item to be decided. According to the analysis of the influence of trust factors and overlapping community factors on decision-making, determine the weights of each decision maker. Promote the realization of group consensus through the trust relationship, fully consider robustness and fairness, apply an optimization algorithm to optimize the consensus reaching process, and generate decision information on the placement location of the item to be decided.

[0182] To implement the above embodiments, the embodiments of the present application propose a computer device, including a memory, a processor, and a computer program stored on the memory and executable on the processor. When the processor executes the computer program, the method in Embodiment 1 is implemented.

[0183] In addition, each functional unit in each embodiment of the present application can be integrated in a processing module, or each unit can exist physically alone, or two or more units can be integrated in a module. The above integrated module can be implemented in the form of hardware or in the form of a software functional module. When the integrated module is implemented in the form of a software functional module and sold or used as an independent product, it can also be stored in a computer-readable storage medium.

[0184] The above-mentioned storage medium may be a read-only memory, a magnetic disk, an optical disc, etc. Although the embodiments of the present application have been shown and described above, it can be understood that the above embodiments are exemplary and should not be construed as limiting the present application. Those of ordinary skill in the art can make changes, modifications, substitutions, and variations to the above embodiments within the scope of the present application.

Claims

1. An intelligent decision-making method based on a neural network, characterized in that: include: Obtain location data and corresponding attribute data of the items to be decided; According to the location data and attribute data, a fuzzy relationship matrix between locations and attributes is established; According to the fuzzy relationship matrix, a network model that reflects the strength of trust relationships between users is generated through the trust propagation operation of the graph neural network. According to the network model, the community structure including community overlap is identified through the community discovery algorithm; Calculate the influence weight of the decision maker based on the strength of trust relationships between users and the degree of community overlap; According to the influence weight and community structure, a group consensus is formed through a robust optimization method to generate the decision results of the items to be decided.

2. The intelligent decision-making method based on a neural network according to claim 1, characterized in that: Establish a fuzzy relationship matrix between places and attributes, including: Collecting a set of alternative solutions consisting of place names according to preset place name elements; Collect attribute sets consisting of dates according to preset time elements; Construct an initial relationship matrix based on each location element in the alternative set and each time element in the attribute set; Fuzzy numbers are used to represent the index of each element in the initial relationship matrix, and the index includes a benefit index and a cost index; Through normalization, the indicators are converted into membership values; Construct the fuzzy relationship matrix based on the membership values.

3. The intelligent decision-making method based on neural network according to claim 2, characterized in that: Through normalization, the indicators are converted into membership values, including: For income indicators: ; For cost indicators: ; in, represents the evaluation value of the kth location under the jth attribute after normalization, represents the relationship between the jth attribute and the kth type of place, j represents the jth day of the week, and k represents the kth type of place in the alternatives; represents the minimum value of the jth attribute among all the alternative locations; Represents the maximum value of the jth attribute among all alternative locations.

4. The intelligent decision-making method based on a neural network according to claim 2, characterized in that: The trust propagation operation of the graph neural network generates a network model that reflects the strength of the trust relationship between users, including: Comprehensively score the candidate locations according to the fuzzy relationship matrix, and select the locations with scores greater than the threshold to form a sample set; Construct an initial interaction network G including user nodes and location nodes based on the sample set; Perform feature transformation on the nodes and edges in the initial interaction network G to obtain network features in the same representation space; According to the network characteristics, the trust propagation operation of the graph neural network is used to obtain the propagation trust characteristics between nodes; According to the propagation trust characteristics between nodes, a network model reflecting the strength of trust relationships between users is obtained through information aggregation.

5. The intelligent decision-making method based on neural network according to claim 4, characterized in that: The network features under the same representation space are obtained, including: Through linear transformation, all nodes in the initial interaction network G are projected into the same dimensional space to obtain the transformed nodes ; Through parameterization, the features of different types of edges in the initial interaction network G are transformed to obtain the transformed edges. ; Initialize the missing node and edge attributes in the initial interaction network G into random vectors, and use the random vectors as learnable parameters of the graph neural network; The transformed nodes , the transformed edge , as network features under the same representation space.

6. The intelligent decision-making method based on a neural network according to claim 4, characterized in that: Obtain the propagation trust characteristics between nodes, including: A trust chain is set in the graph neural network, and a hyperparameter K is preset as the maximum propagation length of the trust chain; wherein the trust chain includes a trustee node and a delegator node; According to the node and edge , calculate the information received by the trustee node v on the trust chain of length k ; For each trusted node , aggregated trustee nodes Information received on different types of trust chains generates propagation trust features between nodes .

7. The intelligent decision-making method based on neural network according to claim 4, characterized in that: The network model reflecting the strength of trust relationship between users is obtained, including: Calculating the weights of different types of trust chains ; According to weight , respectively calculate the embedding representation Z of the trustee node and the delegator node, ; According to the embedding representation Z and , through aggregation representation, we get the final node representation ; According to the final node representation ,construct a network model that reflects the strength of trust relationships between users.

8. The intelligent decision-making method based on neural network according to claim 7, characterized in that: Calculating the weights of different types of trust chains , using the following formula: ； ； in, is the total number of nodes, is the learnable query vector, are the attention network parameters, are different types of propagation trust features, and b is the bias term.

9. The intelligent decision-making method based on a neural network according to claim 4, characterized in that: The community structure including community overlap is identified through community discovery algorithms, including: The network model is divided using the Heron Eagle Optimization Algorithm (SBOA) to obtain a community division scheme. The K-means clustering algorithm based on Manhattan distance is used to cluster the community division scheme and obtain the community center; The community overlap of each decision maker is calculated by identifying the overlapping parts of the decision maker belonging to multiple communities at the same time; According to the community center and community overlap, a community structure including community overlap is generated.

10. The intelligent decision-making method based on neural network according to claim 9, characterized in that: Get the community division plan, including: Generate a random community partition matrix C for N decision makers according to the preset community number Q value; According to the community partition matrix C and the network model that reflects the strength of trust relationships between users, the group consensus level GCL is calculated; According to the Heron Eagle Optimization Algorithm (SBOA), the community partition matrix C is processed by differential mutation, optimal solution guidance based on Brownian motion, and Levy flight disturbance strategy to generate candidate community partition schemes. Calculate the group consensus level GCL of the candidate community division schemes respectively, and select the community division scheme with the highest group consensus level GCL as the current optimal solution; Use the escape strategy of the Heron Eagle Optimization Algorithm SBOA to optimize the current optimal solution; When the preset number of iterations is reached or the improvement of multiple consecutive iterations is lower than the preset threshold, the iteration is terminated and the final community division scheme is obtained.

11. The intelligent decision-making method based on a neural network according to claim 10, characterized in that: Access to community centres, including: Calculate the opinion distance between decision makers using the improved Manhattan distance , ,in, and Represent the opinion vectors of decision makers i and j respectively; Calculate the opinion distance matrix D between any two decision makers in the community division scheme; According to the trust relationship T between decision makers, the comprehensive distance is calculated , ,in, represents the trust level of the jth expert in the ith expert, and M represents the number of decision makers in the community; According to the comprehensive distance , calculate the average comprehensive distance from each decision maker to other decision makers in the community, and select the decision maker with the smallest average comprehensive distance as the center point of the corresponding community; Based on the central point, each non-central decision maker is reassigned to the community with the closest comprehensive distance; Repeat the center point determination and decision maker allocation until the location of the community center point is stable or the preset number of iterations is reached, and output the community center obtained by clustering.

12. The intelligent decision-making method based on neural network according to claim 10, characterized in that: Calculate the community overlap for each decision maker, including: According to the determined community center point, calculate the comprehensive distance between each decision maker i and each community center point j ; according to Calculate the membership coefficient , where U is the number of communities; Set the community overlap judgment threshold θ, when the decision maker's membership of a community When the decision maker is considered to be part of the community; Calculate the community overlap for each decision maker ,in, It is an indicator function, which takes the value 1 when the condition is met, otherwise it takes the value 0.

13. An intelligent decision-making system based on a neural network, characterized in that: include: The acquisition module obtains the location data and corresponding attribute data of the items to be decided; Fuzzy relationship module, which constructs fuzzy relationship matrix based on location data and attribute data; The network model module generates a network model that reflects the strength of trust relationships between users through the graph neural network trust propagation operation based on the fuzzy relationship matrix; The community identification module uses the Heron Eagle algorithm and the improved K-means clustering algorithm to partition the network model and identify the community structure including community overlap; The decision-making module calculates the influence weight of the decision maker according to the strength of the trust relationship between users and the degree of community overlap. Based on the influence weight, a group consensus is formed through a robust optimization algorithm to generate the decision results of the items to be decided.

Images