Supplier evaluation method and device based on big data, equipment and medium

Abstract

The invention relates to a supplier evaluation method and device based on big data, equipment and a medium. The method comprises the steps that multi-source heterogeneous time series data of suppliers and associated entities are collected and preprocessed; identifying a time sequence causal relationship among the indexes by using a time sequence causal discovery algorithm, and constructing a partial directed acyclic graph; fusing the static relationship and the dynamic causal relationship to construct a dynamic causal knowledge graph; utilizing the pre-trained graph neural network to extract graph structure perception representation vectors of the supplier nodes to generate risk scores; analyzing the score cause based on the map causal structure to obtain a causal conduction path; performing mathematical modeling on the path relation to form a structural causal model; and finally, generating a supply chain elasticity analysis report through anti-fact deduction. By adopting the method, the'black box 'dilemma of a traditional assessment model can be broken through, the assessment process is traceable and understandable, quantitative deduction of the risk intervention effect is supported based on a real causal relationship, and decision support is provided for active risk management.

Description

Technical Field

[0001] This invention belongs to the field of data intelligent analysis, and in particular relates to a supplier evaluation method, apparatus, equipment and medium based on big data. Background Technology

[0002] With the widespread application of big data and artificial intelligence technologies, machine learning-based supplier evaluation methods have become mainstream. These methods predict risk by analyzing statistical correlations in massive amounts of data. The typical process involves first integrating multi-source data to construct feature indicators, then using algorithms such as logistic regression and random forest to train classification or regression models, and finally outputting the supplier's risk score or rating.

[0003] However, current methods suffer from two key drawbacks: First, the model's decision-making process is opaque and lacks interpretability. Complex algorithms act like a "black box," providing only results without offering convincing decision-making basis, making it difficult for business personnel to understand and trust the assessment conclusions. Second, the models rely on data correlation and lack causal reasoning capabilities. Existing methods can only reveal the statistical association between indicators and risks, failing to identify true causal relationships. Therefore, they cannot answer the root question of "why the risk is high," nor can they effectively extrapolate "how the risk will change after taking a certain measure," limiting their decision support value in proactive risk management. Summary of the Invention

[0004] Therefore, it is necessary to provide a big data-based supplier evaluation method, device, equipment, and medium that can overcome the "black box" dilemma of models, improve the interpretability of evaluation decisions, uncover the true causal relationship between indicators and risks, and support counterfactual inference to empower proactive risk management, in order to address the above-mentioned technical problems.

[0005] Firstly, this application provides a supplier evaluation method based on big data, including:

[0006] Collect multi-source heterogeneous time-series data related to suppliers and associated entities, and preprocess the multi-source heterogeneous time-series data to obtain the corresponding time-series event set and the time-series indicator sequence of each entity; the time-series event set includes entity, relationship, and timestamp.

[0007] The temporal causal discovery algorithm is used to identify the temporal causal relationship of temporal index sequences and construct a partially directed acyclic graph with temporal causal edges;

[0008] By combining the static relationships in a set of temporal events with the temporal causal relationships in some directed acyclic graphs, a dynamic causal knowledge graph is constructed.

[0009] The dynamic causal knowledge graph is input into a pre-trained graph neural network to extract the graph structure-aware representation vectors of the supplier nodes.

[0010] Based on the graph structure perception representation vector, generate a risk score for the supplier;

[0011] Based on the causal structure of the dynamic causal knowledge graph, the risk score is analyzed to obtain the causal transmission path;

[0012] Based on the causal transmission path, the causal relationship is mathematically modeled to obtain a structural causal model;

[0013] By inputting a set of time-series events into a structural causal model, a supply chain resilience analysis report containing counterfactual risk scores is generated through counterfactual inference.

[0014] In one embodiment, a dynamic causal knowledge graph is constructed by combining static relationships in a set of temporal events with temporal causal relationships in a partial directed acyclic graph, including:

[0015] Based on a time-series event set, a static supply chain relationship graph is constructed with suppliers, products, raw materials, and regions as nodes;

[0016] Extract the time-series index sequences corresponding to the entity nodes in the static supply chain relationship graph to obtain the set of time-series variables to be analyzed;

[0017] The temporal causal discovery algorithm based on conditional independence testing examines the conditional independence between variables in the set of time-series variables to be analyzed under different time delays, and obtains a set of conditional independence test results.

[0018] Based on the set of conditional independence test results, the causal direction and time lag between variables are inferred, resulting in a partially directed acyclic graph with temporal causal edges;

[0019] By fusing temporal causal edges in a portion of a directed acyclic graph with static relationship edges in a static supply chain relationship graph, a dynamic causal knowledge graph is generated; each temporal causal edge is accompanied by causal effect strength and confidence information.

[0020] Based on the above embodiments, and using the set of conditional independence test results, the causal direction and time lag between variables are inferred, resulting in a partially directed acyclic graph with temporal causal edges, including:

[0021] For the variable X in the time series variable set to be analyzed i and X j Given the maximum time delay τ max Within the range, for all possible time delays τ, construct a condition set Z, which contains all conditions except X. i and X j The values ​​of other variables at and before time t;

[0022] Based on the observed sample data of the time series variable set, calculate the variable X under the condition set Z. i (t-τ) and variable X j The conditional independence between (t) is obtained by setting the set of conditional independence test statistics;

[0023] Calculate the set of p-values ​​corresponding to the set of conditional independence test statistics. The p-value represents the probability of observing the current test statistic or a more extreme case if the null hypothesis is true.

[0024] The set of p-values ​​is compared with a preset significance level threshold α. When the p-value is less than the significance level threshold α, the variable X is generated. i To variable X j A preliminary set of causal links with a time delay of τ;

[0025] The initial set of causal links is integrated into a causal relationship adjacency matrix;

[0026] By removing the cyclic dependencies of the causal adjacency matrix using a delooping algorithm in graph theory, a partially directed acyclic graph with temporally causal edges is obtained.

[0027] In one embodiment, based on the causal structure of a dynamic causal knowledge graph, risk scores are analyzed to obtain causal transmission paths, including:

[0028] Based on the causal edge structure in the dynamic causal knowledge graph, the attention weights between nodes are calculated during the message passing process of the pre-trained graph neural network to obtain a set of attention weights that represent the importance of the edges.

[0029] Based on the attention weight set, initialize the learnable edge mask matrix corresponding to the causal edge set in the dynamic causal knowledge graph;

[0030] The learnable edge mask matrix and the high-confidence causal edge adjacency matrix extracted from the dynamic causal knowledge graph are input into a preset function construction rule to construct an optimized objective function for generating the explanatory subgraph.

[0031] The optimal edge mask matrix is ​​obtained by optimizing the objective function;

[0032] Based on the optimal edge mask matrix, causal transmission paths are extracted from the dynamic causal knowledge graph.

[0033] Based on the above embodiments, the expression for the optimized objective function is as follows:

[0034]

[0035] Where M is the learnable edge mask matrix, f(G,M) represents the prediction output of the graph neural network running on the subgraph G weighted by the mask matrix M, and y is the original risk score prediction value of the supplier. The loss function term that measures the difference between the subgraph prediction and the original prediction. For the L1 norm regularization term of the mask matrix M, and To balance the hyperparameter weights of different loss terms, For the causal alignment loss function term, It is an adjacency matrix extracted from a dynamic causal knowledge graph, representing high-confidence causal edges. Item used to punish The concealment of important causal edges.

[0036] In one embodiment, based on the causal transmission path, the causal relationship is mathematically modeled to obtain a structural causal model, including:

[0037] For node variables with in-degree causal edges in a causal transmission path Identify the set of causal parent nodes , Include variables , ;

[0038] Based on node variables With the set of causal parent nodes Establish structural equations and calculate node variables using the following formula. Theoretical values ​​of the variables:

[0039]

[0040] in It is a parameterizable function. These are the function parameters to be learned. This is the unobserved noise term;

[0041] Based on the theoretical values ​​of variables and historical time series data, a set of parameterized structural equations is obtained through time series regression or gradient descent optimization algorithms.

[0042] By integrating the parameterized set of structural equations, a structural causal model is obtained.

[0043] Secondly, this application also provides a supplier evaluation device based on big data, comprising:

[0044] The data acquisition and preprocessing module is used to collect multi-source heterogeneous time-series data related to suppliers and associated entities, and preprocess the multi-source heterogeneous time-series data to obtain the corresponding time-series event set and the time-series indicator sequence of each entity; the time-series event set includes entity, relationship, and timestamp.

[0045] The temporal causality discovery module is used to identify temporal causal relationships in temporal index sequences using temporal causality discovery algorithms and to construct a partially directed acyclic graph with temporal causal edges.

[0046] The dynamic causal knowledge graph construction module is used to combine the static relationships in the set of temporal events with the temporal causal relationships in some directed acyclic graphs to construct a dynamic causal knowledge graph.

[0047] The graph structure-aware representation generation module is used to input a dynamic causal knowledge graph into a pre-trained graph neural network and extract the graph structure-aware representation vectors of the supplier nodes.

[0048] The risk score generation module is used to generate risk scores for suppliers based on graph structure perception representation vectors.

[0049] The causal transmission path parsing module is used to parse risk scores and obtain causal transmission paths based on the causal structure of a dynamic causal knowledge graph.

[0050] The structural causal model building module is used to mathematically model causal relationships based on causal transmission paths to obtain structural causal models.

[0051] The report generation module is used to input a set of time-series events into a structural causal model and generate a supply chain resilience analysis report containing a counterfactual risk score through counterfactual inference.

[0052] Thirdly, this application also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of any of the methods in the first aspect of this application.

[0053] Fourthly, this application also provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the steps of any of the methods in the first aspect of this application.

[0054] This application discloses a supplier evaluation method, apparatus, equipment, and medium based on big data. It collects multi-source heterogeneous time-series data from suppliers and related entities, preprocesses it to obtain a set of time-series events and a sequence of time-series indicators, identifies time-series causal relationships using a time-series causal discovery algorithm to construct a partially directed acyclic graph, integrates static relationships and time-series causal relationships to generate a dynamic causal knowledge graph, inputs the graph into a pre-trained graph neural network to extract supplier node graph structure perception vectors to generate risk scores, obtains causal transmission paths based on graph causal structure analysis, obtains a structural causal model through mathematical modeling, and inputs the time-series event set to generate a supply chain resilience analysis report containing counterfactual risk scores through counterfactual deduction. Dynamic causal knowledge graphs clearly present the static relationships and temporal causal logic between entities, and the causal transmission path clearly defines the root cause of risk score generation, breaking the "black box" dilemma and making the assessment process traceable and understandable, significantly improving business personnel's trust in the assessment conclusions. The temporal causal discovery algorithm identifies the true causal relationship between indicators and risks, and the structural causal model and counterfactual inference technology can simulate risk changes after factors change, providing quantitative support for risk control measures, getting rid of dependence on statistical correlations, empowering proactive risk management, and comprehensively making up for the core defects of existing methods such as poor interpretability and lack of causal reasoning ability, providing a more scientific and practical solution for supply chain risk management. Attached Figure Description

[0055] To more clearly illustrate the technical solutions in the embodiments or related technologies of this application, the accompanying drawings used in the description of the embodiments or related technologies will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0056] Figure 1 This is a flowchart of a supplier evaluation method based on big data in one embodiment of the present invention;

[0057] Figure 2 This is a schematic diagram of a supplier evaluation device based on big data in one embodiment of the present invention. Detailed Implementation

[0058] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0059] In one embodiment, such as Figure 1As shown, a supplier evaluation method based on big data is provided. This embodiment illustrates the application of this method to a terminal for supplier evaluation. It is understood that this method can also be applied to a server, or to a system including both a terminal and a server, and is implemented through the interaction between the supplier evaluation terminal and the supplier evaluation server. In this embodiment, the method includes the following steps:

[0060] S101: Collect multi-source heterogeneous time-series data related to suppliers and associated entities, and preprocess the multi-source heterogeneous time-series data to obtain the corresponding time-series event set and the time-series indicator sequence of each entity; the time-series event set includes entities, relationships, and timestamps.

[0061] Among them, multi-source heterogeneous time-series data refers to differentiated data with time stamps collected from various sources and forms of suppliers and their related entities, which are integrated and processed to form structured time-series information that reflects the dynamic changes in the supply chain.

[0062] For example, the supplier evaluation terminal collects multi-source heterogeneous time-series data related to suppliers and associated entities. The multi-source heterogeneous time-series data is preprocessed through the following steps: cleaning the multi-source heterogeneous time-series data; supplementing missing values ​​of continuous indicators with linear interpolation; marking missing discrete events as unrelated events; retaining outliers after judgment based on business rules; deduplicating by entity-event-timestamp; fusing the multi-source heterogeneous time-series data; standardizing data of the same type using the Z-score standardization method, where the Z-score standardization method expression is:

[0063]

[0064] Where Z is the standardized value, which is The original value, The mean, Using the standard deviation, semantics are unified through ontology mapping to obtain the corresponding time-series event set and the time-series indicator sequence of each entity. The time-series event set is stored as a quadruple of entity1-relationship-entity2-timestamp; the time-series indicator sequence is organized according to "entity-indicator name-time series data", which accurately corresponds to the indicator change trajectory of each entity.

[0065] S102: Use the temporal causal discovery algorithm to identify the temporal causal relationship of the temporal index sequence and construct a partially directed acyclic graph with temporal causal edges.

[0066] Among them, the temporal causal discovery algorithm is a class of statistical and machine learning methods that automatically identify the direction, strength, and time lag of causal relationships between variables from time-stamped observation data. Its core lies in distinguishing between true causal drivers and simple temporal correlations. By testing conditional independence, analyzing Granger causality, or utilizing structural constraints, it infers the causal structure in which the "cause" event occurs before the "effect" event, and outputs a partially directed acyclic graph with temporal edges, providing an interpretable causal skeleton for constructing dynamic causal models.

[0067] For example, the supplier evaluation terminal uses a time-series causal discovery algorithm to transform causal identification into conditional independence testing within a specific time frame through a time window mechanism, avoiding time-series misjudgments. It standardizes the indicator sequence to eliminate dimensions and sets the time window and lag order according to the indicator change cycle. A complete undirected graph containing all indicators is constructed during initialization. In the edge deletion phase, a sliding window is used to retrieve data, and independence is tested and edges are deleted using the partial correlation coefficient formula. The partial correlation coefficient expression is:

[0068]

[0069] in, To control the partial correlation coefficient between X and Y at time Z, The Pearson correlation coefficient between X and Y. X is the multiple or multivariate correlation coefficient between X and the linear combination of all variables in the condition set Z. X and Y are two time-series index variables being tested for whether they have a direct causal relationship, and Z is the set of condition variables. In the edge direction determination stage, the direction is determined by combining the time series and the V structure, and loop edges are deleted to obtain a partially directed acyclic graph with time-series causal edges.

[0070] S103: Combine the static relationships in the set of time-series events with the time-series causal relationships in some directed acyclic graphs to construct a dynamic causal knowledge graph.

[0071] Among them, the dynamic causal knowledge graph is a graph structure that integrates static relationships between entities with temporal causal logic. It uses suppliers, products, regions, etc. as nodes and contains two types of edges: one is static relationship edges that reflect inherent connections, and the other is dynamic causal edges with time delay and intensity information inferred by temporal causal discovery algorithms. This graph not only depicts the static topology of the supply chain, but also reveals the mechanism of dynamic transmission of risks along temporal paths, providing an interpretable knowledge base that combines structure and causality for subsequent risk prediction and counterfactual inference of graph neural networks.

[0072] For example, the supplier evaluation terminal uses a static supply chain relationship graph to extract entity relationships. It identifies stable relationships between suppliers and related entities from multi-source data, forming standardized entity-relationship-entity triples. Stable relationships include types such as equity, cooperation, and supply and demand. The extraction technique can utilize mainstream methods based on deep learning or statistical learning. The supplier evaluation terminal introduces preliminary causal relationships, verifies the reliability of the relationships through conditional independence tests, infers the causal direction and time lag between variables based on temporal characteristics, and simultaneously obtains the strength and confidence information of causal effects through statistical analysis. The confidence level can be quantified using significance indicators. This temporal causal information is transformed into dynamic edges containing attributes of "cause-relationship-effect-time lag-strength-confidence level," which are then imported into a graph database along with the static relationship graph. After defining and standardizing the knowledge structure through ontology, redundant nodes are eliminated using an entity disambiguation algorithm, forming a knowledge graph that integrates static relationships and dynamic causality.

[0073] S104: Input the dynamic causal knowledge graph into the pre-trained graph neural network and extract the graph structure-aware representation vector of the supplier nodes.

[0074] Among them, the pre-trained graph neural network is a deep learning model pre-trained based on large-scale graph data, which can effectively learn the structural and semantic information of nodes and edges in the graph. In this method, it receives a dynamic causal knowledge graph as input, uses the learned patterns to perform deep representation of the supplier nodes and their causal relationships in the graph, and outputs a node vector that integrates topological structure and causal dependence, providing an interpretable and causally aware embedding basis for risk scoring.

[0075] For example, the supplier evaluation terminal inputs a dynamic causal knowledge graph into a pre-trained graph neural network to extract graph structure-aware representation vectors of supplier nodes. The graph neural network captures entity association characteristics through a neighbor node feature aggregation mechanism, improving the accuracy of feature representation; the Transform model is selected. During the pre-training phase, a general dataset from the supply chain domain and customized labeled data are used, with node attribute prediction and relationship type recognition as the core tasks, enabling the Transform model to learn the graph structure and semantic features in the supply chain scenario. During fine-tuning of the Transform model, discrete attributes of nodes are transformed into low-dimensional vectors using feature encoding techniques suitable for text or category information. Continuous attributes are standardized and then fused with the discrete attribute encoding results to form initial features. Neighbor features of the target node are aggregated through an attention mechanism or message passing mechanism, and the output results are optimized by combining a multi-view feature fusion strategy to obtain a graph structure-aware representation vector that integrates node attributes, graph topology relationships, and causal association information.

[0076] S105: Generate a risk score for the supplier based on the graph structure perception representation vector.

[0077] For example, the supplier evaluation terminal fuses the representation vector with recognized risk-related business features in the supply chain domain to achieve feature enhancement. These business features may include supplier qualification level, historical performance record, and cooperation stability. The fused feature set is standardized to eliminate dimensional differences and improve model training performance. An ensemble learning model with both interpretability and predictive performance is selected. Multi-model collaborative decision-making reduces the bias of a single model. Part of the base model is used to capture linear correlations between features to ensure interpretability, while another part is used to uncover complex nonlinear relationships between features to improve accuracy. The dataset is divided into training, validation, and test sets according to a conventional ratio. The training set is used to fit model parameters, the validation set is used for hyperparameter optimization and model tuning, and the test set is used to verify the model's generalization ability. After the model outputs the supplier risk association probability, it is converted into a risk score of 0-100 through probability mapping rules, and a threshold is set according to business needs to complete the risk level classification.

[0078] S106: Based on the causal structure of the dynamic causal knowledge graph, analyze the risk score and obtain the causal transmission path.

[0079] For example, starting with the risk score, the directed edges of the graph are traversed in reverse order. After identifying the primary influencing indicators, the process continues to trace back to the root cause indicators, forming a preliminary set of paths. Path selection uses a significance test, and the strength is calculated using the strength formula, which is expressed as:

[0080]

[0081] in, Let k be the total strength of the candidate causal propagation path. Let the strength of the i-th edge on the k-th candidate causal path be calculated; the strength is then used to retain high-threshold paths as candidates. The contribution formula is used to calculate the contribution:

[0082]

[0083] Where C represents the single-path contribution. Let m be the strength of the j-th candidate path and m be the number of candidates. The paths arranged in descending order of contribution are taken as the core transmission paths. The complete "root-middle-risk" link and the strength and contribution of each link are obtained, thus obtaining the causal transmission path.

[0084] S107: Based on the causal transmission path, the causal relationship is mathematically modeled to obtain a structural causal model.

[0085] Among them, the structural causal model is a quantitative mathematical model based on causal graphs. The structural causal model explicitly describes the causal mechanism and functional relationship between variables through a set of structural equations. In this method, the structural causal model is constructed by parametrically learning the causal transmission path identified in the dynamic causal knowledge graph. It accurately expresses how the "cause" variable affects the "effect" variable through specific time delays and functional relationships, thus providing a computable causal reasoning engine for subsequent counterfactual intervention inference.

[0086] For example, the core of obtaining a structural causal model is to quantify causal relationships through mathematical modeling. Key steps include parent node identification, structural equation construction, parameter optimization, and model integration. For node variables with preceding causal influences in the causal transmission path, their corresponding causal parent node set is accurately identified. This set includes all preceding variables that directly cause the target node and their corresponding time lag parameters. Based on the correlation characteristics between the target node variable and the causal parent node set, a parameterizable function is used to construct the structural equation between them, clarifying the quantitative correlation logic between variables. Unobserved noise terms are included in the equation to cover potential minor influencing factors. Combining historical time-series data, the parameters of the structural equation are learned and iteratively optimized using mainstream optimization algorithms such as time-series regression and gradient descent to obtain a structured equation with determined parameters. The parameterized structural equations corresponding to all node variables are systematically integrated to form a structural causal model that can accurately characterize causal relationships.

[0087] S108: Input the set of time-series events into the structural causal model, and generate a supply chain resilience analysis report containing counterfactual risk scores through counterfactual inference.

[0088] For example, the supplier assessment terminal inputs a set of time-series events into a structural causal model. Through counterfactual inference, it generates a supply chain resilience analysis report containing a counterfactual risk score. Based on the historical data characteristics of the time-series events, and focusing on the root indicators of the core causal transmission path, it designs counterfactual scenarios that meet actual business needs. Scenario types can cover typical supply chain intervention measures such as raw material price fluctuations, policy adjustments, and changes in fulfillment methods. Based on the core logic of causal intervention, variables other than the target root indicator are fixed in the structural causal model, and only the target indicator value is adjusted to the scenario-defined value. The risk score under this scenario, i.e., the counterfactual risk score, is calculated through the structural causal model. The counterfactual risk score is compared and analyzed with the actual risk score to quantify the differences and trends between the two scores, clarifying the impact of different intervention measures on risk. The resulting supply chain resilience analysis report must include a comparison of actual and counterfactual risk scores, risk change characteristics of each scenario, identification of key intervention nodes, and targeted management suggestions, providing quantitative decision support for supply chain risk management.

[0089] The technical solution provided in this application includes the following technical effects: This application discloses a supplier evaluation method, apparatus, equipment, and medium based on big data. By collecting multi-source heterogeneous time-series data of suppliers and related entities and preprocessing it to obtain a set of time-series events and a sequence of time-series indicators, a time-series causal discovery algorithm is used to identify time-series causal relationships and construct a partially directed acyclic graph. Static relationships and time-series causal relationships are fused to generate a dynamic causal knowledge graph. The graph is input into a pre-trained graph neural network to extract supplier node graph structure perception representation vectors to generate risk scores. Causal transmission paths are obtained based on graph causal structure analysis. A structural causal model is obtained through mathematical modeling. The time-series event set is input and counterfactual inference is used to generate a supply chain resilience analysis report containing counterfactual risk scores. Dynamic causal knowledge graphs clearly present the static relationships and temporal causal logic between entities, and the causal transmission path clearly defines the root cause of risk score generation, breaking the "black box" dilemma and making the assessment process traceable and understandable, significantly improving business personnel's trust in the assessment conclusions. The temporal causal discovery algorithm identifies the true causal relationship between indicators and risks, and the structural causal model and counterfactual inference technology can simulate risk changes after factors change, providing quantitative support for risk control measures, getting rid of dependence on statistical correlations, empowering proactive risk management, and comprehensively making up for the core defects of existing methods such as poor interpretability and lack of causal reasoning ability, providing a more scientific and practical solution for supply chain risk management.

[0090] In one embodiment of the present invention, a dynamic causal knowledge graph is constructed by combining static relationships in a set of temporal events with temporal causal relationships in a partial directed acyclic graph, including:

[0091] S201: Based on a set of time-series events, construct a static supply chain relationship graph with suppliers, products, raw materials, and regions as nodes.

[0092] For example, the supplier evaluation terminal extracts basic information on four types of entities—suppliers, products, raw materials, and regions—from a time-series event set. Entity identification technology is used to differentiate these entities, assigning each entity a unique identifier, such as a unified social credit code for suppliers and an administrative region code for regions. The core static relationships between entities are further identified, including the supply relationship between suppliers and products, the procurement relationship between suppliers and raw materials, and the compositional relationship between products and raw materials. These entities and relationships are then organized into a static supply chain relationship graph.

[0093] S202: Extract the time series index sequences corresponding to the entity nodes in the static supply chain relationship graph to obtain the set of time series variables to be analyzed.

[0094] For example, using entity nodes in a static supply chain relationship graph as indexes, the time-series indicator sequences corresponding to each node are matched and extracted, and aggregated to form a set of time-series variables to be analyzed. Core time-series indicators are matched for four types of entities: operational indicators such as fulfillment rate and capacity utilization rate for suppliers; performance indicators such as qualification rate and sales volume for products; supply indicators such as price fluctuations and inventory for raw materials; and environmental indicators such as logistics efficiency and policy support for regions. The association between indicators and nodes is established through unique entity identifiers, ensuring that all indicator data carries the timestamp of the bound entity. Multiple indicators of the same entity are aligned according to the time dimension to obtain the set of time-series variables to be analyzed.

[0095] S203: A time-series causal discovery algorithm that uses conditional independence testing to examine the conditional independence between variables in the time-series variable set to be analyzed under different time delays, and obtains a set of conditional independence test results.

[0096] For example, the supplier evaluation terminal employs a time-series causal discovery algorithm with integrated conditional independence testing functionality to examine the relationships between time-series variables under multiple time-lag parameters. The time-lag range is set based on the volatility characteristics of the indicators, with daily / weekly time lags for high-frequency indicators and monthly / quarterly time lags for low-frequency indicators. Partial correlation analysis and mutual information testing methods are used to test any pair of variables under various time lags by introducing a set of conditional variables. Core information such as "variable pair - time lag - conditional variable set - test result - significance index" is recorded and summarized to form a set of conditional independence test results.

[0097] S204: Based on the set of conditional independence test results, infer the causal direction and time lag between variables to obtain a partially directed acyclic graph with temporal causal edges.

[0098] For example,

[0099] S205: Merge the temporal causal edges in a portion of the directed acyclic graph with the static relationship edges in the static supply chain relationship graph to generate a dynamic causal knowledge graph; wherein each temporal causal edge is accompanied by causal effect strength and confidence information.

[0100] For example, the supplier evaluation terminal merges temporal causal edges from a portion of the directed acyclic graph with static relationship edges from the static supply chain relationship graph to generate a dynamic causal knowledge graph. Each temporal causal edge is appended with causal effect strength and confidence information. The positioning of static edges as representing stable entity associations and temporal causal edges as reflecting the dynamic causal effects of variables is clearly defined. The effect strength of temporal causal edges is quantified through statistical analysis, and the confidence level is characterized by the significance index of conditional independence tests. Both types of edges are imported into the same graph model, and through ontology definition standardization, redundancy and conflicts are eliminated via entity disambiguation, forming a knowledge graph that combines static association and dynamic causal characteristics.

[0101] The supplier evaluation method based on big data provided in this application achieves precise correlation between entities and time-series indicators by accurately constructing a static supply chain relationship graph with defined core nodes. Combined with conditional independence tests and causal inference under multi-dimensional time lags, it efficiently mines the real time-series causal relationships between variables and finally integrates them to generate a dynamic causal knowledge graph with key attributes. This not only ensures the structured presentation of supply chain entity relationships but also improves the accuracy and timeliness of causal relationship identification, providing more targeted and reliable structured knowledge support for subsequent supply chain risk analysis and decision support.

[0102] In one embodiment of the present invention, based on the set of conditional independence test results, the causal direction and time lag between variables are inferred to obtain a partially directed acyclic graph with temporal causal edges, including:

[0103] S301: For variable X in the set of time series variables to be analyzed i and X j Given the maximum time delay τ max Within the range, for all possible time delays τ, construct a condition set Z, which contains all conditions except X. i and X j The values ​​of other variables at and before time t.

[0104] Among these parameters, the maximum time delay is a pre-defined key parameter that defines the upper limit of the maximum time range for the algorithm to search for causal relationships. Its function is to constrain the causal variable X. i It can affect the outcome variable X j The longest delay time ensures that the discovered causal relationship has practical timeliness significance.

[0105] For example, the supplier evaluation terminal targets each pair of variables X in the set of time-series variables to be analyzed. i and X j Based on the business characteristics and change cycle of time-series variables, a reasonable maximum time delay τ is set. max Range. Within the maximum time delay τ max Iterate through all time delays τ within the range, for each group (X) i ,X j Construct the corresponding condition set Z, where the construction of condition set Z requires excluding variable X. i and X j Only historical observations of all other variables in the time series variable set at and before time t are included to ensure that the condition set can comprehensively cover all possible influences on X. i With X j Other variables that cause interference due to association relationships.

[0106] S302: Based on the observed sample data of the time-series variable set, calculate the variable X under the condition set Z.i (t-τ) and variable X j The conditional independence between (t) is obtained by setting the set of conditional independence test statistics.

[0107] For example, the supplier evaluation terminal, based on the observed sample data corresponding to the time series variable set, evaluates each group (X) i ,X j The variable X is calculated using a pre-defined conditional independence test method. i The value at time t-τ and the variable X j The conditional independence between the values ​​at time t. The test process uses the condition set Z as the control variable to mask its interference, and outputs the conditional independence test statistic for each test group. All statistics are summarized to form a set of conditional independence test statistics, which is directly related to the corresponding (X) values. i ,X j ,τ) combined information.

[0108] S303: Calculate the set of p-values ​​corresponding to the set of conditional independence test statistics. The p-value represents the probability of observing the current test statistic or a more extreme case if the null hypothesis is true.

[0109] The p-value is a probability indicator in hypothesis testing, representing the probability of observing the current sample data given that the null hypothesis is true. The smaller the p-value, the greater the inconsistency between the current observation and the null hypothesis, thus providing stronger statistical evidence to reject the null hypothesis.

[0110] For example, the supplier evaluation terminal calculates the p-value corresponding to each conditional independence test statistic based on the set of conditional independence test statistics and the distribution characteristics of the corresponding conditional independence test methods. All the calculated p-values ​​are summarized and integrated to form a structured p-value set, which is associated with each statistic in the set of conditional independence test statistics.

[0111] S304: Compare the set of p-values ​​with a preset significance level threshold α. When the p-value is less than the significance level threshold α, generate the variable X. i To variable X j A preliminary set of causal links with a time delay of τ.

[0112] The level threshold α is a key decision criterion pre-set in statistical hypothesis testing, usually set to 0.05 or 0.01, used to control the maximum acceptable probability of a Type I error, such as erroneous rejection of the true null hypothesis, in statistical testing.

[0113] For example, a significance level threshold α is preset to determine whether a causal relationship is significant. Each p-value in the set of p-values ​​is compared with the threshold α one by one. When a p-value is less than the threshold α, it indicates that the null hypothesis is not true, i.e., X < α. i (t-τ) and X j (t) shows a significant conditional dependency, based on which a conditional dependency is generated from variable X. i Pointing to variable X j Preliminary causal links with a time delay of τ are identified. All preliminary causal links that meet the criteria are aggregated to form a preliminary causal link set.

[0114] S305: Integrate the initial set of causal links into a causal adjacency matrix.

[0115] For example, a causal adjacency matrix is ​​constructed using all variables in the time-series variable set as the row and column indices of the matrix. The rule for the values ​​of the matrix elements is: if there exists a dependent variable X... i Pointing to variable X j A preliminary causal link with a time delay of τ corresponds to the matrix element (X). i ,X j If no corresponding initial causal link exists, the matrix element is assigned the value τ; if no corresponding initial causal link exists, the matrix element is assigned the value 0. The causal adjacency matrix can intuitively represent the variable relationships and time delay information of all initial causal links, realizing the structured integration of the initial causal link set.

[0116] S306: By removing the cyclic dependencies of the causal adjacency matrix using a delooping algorithm in graph theory, a partially directed acyclic graph with temporally causal edges is obtained.

[0117] For example, the supplier evaluation terminal uses mainstream delooping algorithms in graph theory to process the causal adjacency matrix. It identifies cyclic dependencies in the matrix through strongly connected component decomposition and breaks the cycles by deleting or adjusting edges with weaker causal strength in the cyclic links. After delooping, a graph structure is constructed based on the adjusted adjacency matrix. Each node in the graph corresponds to a temporal variable, and each directed edge corresponds to a verified temporal causal relationship and is labeled with a time delay, resulting in a partially directed acyclic graph with no cyclic dependencies and clear causal logic, containing temporal causal edges.

[0118] The supplier evaluation method based on big data provided in this application constructs a comprehensive condition set within a reasonable time lag range to mask interfering variables. It accurately identifies significant temporal causal links by combining conditional independence tests and p-value screening. After the causal relationship is structurally integrated through an adjacency matrix, a loop removal algorithm is used to eliminate circular dependencies. Finally, a directed acyclic graph of temporal causality with clear causal orientation and complete time lag information is obtained. This effectively eliminates false associations, ensures the rigor of causal logic, and significantly improves the accuracy and reliability of temporal causal relationship identification. It lays a solid graphical foundation for subsequent supply chain-related causal analysis and decision support.

[0119] In an exemplary embodiment, based on the causal structure of a dynamic causal knowledge graph, risk scores are analyzed to obtain causal transmission paths, including:

[0120] S401: Based on the causal edge structure in the dynamic causal knowledge graph, the attention weights between nodes are calculated during the message passing process of the pre-trained graph neural network to obtain a set of attention weights that represent the importance of the edges.

[0121] Among them, pre-trained graph neural networks are deep learning models that are pre-trained based on large-scale general or domain graph data. Their core advantage lies in their powerful basic feature extraction capabilities by learning the general structural patterns and semantic associations in the graph in advance.

[0122] For example, during the message passing phase, the graph neural network adaptively learns node interaction weights by combining the node attributes and causal relationship features connected by causal edges. These interaction weights represent the importance of the corresponding causal edge to the risk scoring causal analysis. The supplier evaluation terminal then aggregates the attention weights corresponding to all causal edges to obtain a structured set of attention weights representing the importance of each edge.

[0123] S402: Based on the attention weight set, initialize the learnable edge mask matrix corresponding to the causal edge set in the dynamic causal knowledge graph.

[0124] The learnable edge mask matrix is ​​a weight matrix with the same dimension as the edge set of the dynamic causal knowledge graph, whose element values ​​can be adjusted through an optimization algorithm. When parsing risk scores, the learnable edge mask matrix acts as an attention filter, automatically learning and quantifying the contribution of each edge in the graph to the final risk score prediction. By incorporating it together with a high-confidence causal edge adjacency matrix representing prior knowledge into the optimization objective function for training, the learnable edge mask matrix can be optimized, thereby accurately identifying and drawing the causal transmission path subgraphs that have a decisive impact on the risk score.

[0125] For example, the supplier evaluation terminal initializes a learnable edge mask matrix based on the attention weight set. The dimension of the learnable edge mask matrix matches the causal edge set in the dynamic causal knowledge graph. The initial values ​​of the matrix elements are obtained by mapping the attention weights of the corresponding causal edges. The higher the weight, the closer the initial mask value is to 1, and the lower the weight, the closer the initial mask value is to 0. This realizes the initial empowerment of the mask matrix by the importance of the edges, and the matrix has the learnable characteristic of subsequent iterative optimization.

[0126] S403: Input the learnable edge mask matrix and the high-confidence causal edge adjacency matrix extracted from the dynamic causal knowledge graph into the preset function construction rules to construct an optimized objective function for generating the explanatory subgraph.

[0127] For example, the supplier evaluation terminal selects causal edges with confidence levels meeting preset requirements from a dynamic causal knowledge graph, constructing a high-confidence causal edge adjacency matrix. The learnable edge mask matrix and the high-confidence causal edge adjacency matrix are then input into preset function construction rules to construct a multi-constraint collaborative optimization objective function. The construction rules aim to ensure subgraph prediction consistency, causal logic integrity, and mask matrix sparsity. By integrating constraints from different dimensions, a comprehensive optimization criterion is formed. The constructed objective function guides the accurate generation of subsequent interpretable subgraphs.

[0128] S404: The optimal edge mask matrix is ​​calculated by optimizing the objective function.

[0129] For example, the supplier evaluation terminal uses mainstream gradient optimization algorithms to iteratively optimize the constructed objective function. During the optimization process, the element values ​​of the learnable edge mask matrix are dynamically adjusted to balance the influence of various constraints and gradually reduce the overall loss value. When the loss value converges to a preset threshold or the number of iterations reaches a set upper limit, the optimization process stops, and the resulting edge mask matrix is ​​the optimal edge mask matrix, which can accurately characterize the criticality of each causal edge to the analysis of the causes of risk scoring.

[0130] S405: Extract causal transmission paths from the dynamic causal knowledge graph based on the optimal edge mask matrix.

[0131] For example, an element selection threshold is set for the optimal edge mask matrix, and causal edges in the dynamic causal knowledge graph corresponding to edges with element values ​​higher than the threshold are retained. Taking the supplier risk score as the endpoint, the key causal edges selected are traced backward to connect the complete link from the root indicator through intermediate indicators to the risk score, extracting the core causal transmission path and presenting the causal orientation and time sequence relationship between each node.

[0132] The deep learning-based wind and solar resource prediction method provided in this application accurately filters key causal edges through attention weighting and masking optimization, efficiently extracts core causal transmission paths with clear direction and time sequence, significantly improves the accuracy and interpretability of risk scoring cause analysis, and provides reliable causal basis for locating the root causes of supply chain risks and making control decisions.

[0133] Based on the above embodiments, the expression for the optimized objective function is as follows:

[0134]

[0135] Where M is the learnable edge mask matrix, f(G,M) represents the prediction output of the graph neural network running on the subgraph G weighted by the mask matrix M, and y is the original risk score prediction value of the supplier. The loss function term that measures the difference between the subgraph prediction and the original prediction. For the L1 norm regularization term of the mask matrix M, and To balance the hyperparameter weights of different loss terms, For the causal alignment loss function term, It is an adjacency matrix extracted from a dynamic causal knowledge graph, representing high-confidence causal edges. Item used to punish The concealment of important causal edges.

[0136] For example, the optimization objective is to minimize the comprehensive loss value to determine the optimal learnable edge mask matrix. The comprehensive loss includes multiple types of collaborative constraint terms. The first type of constraint term measures the difference between the prediction results on the mask-weighted subgraph and the original risk score prediction values, ensuring the effectiveness of the subgraph prediction. The second type of constraint term regularizes the mask matrix, improving its sparsity to filter key edges. The third type of constraint term is a causal alignment constraint, used to avoid masking high-confidence causal edges, ensuring the causal logic integrity of the subgraph. By adjusting the weight ratio of each type of constraint term through hyperparameters, and combining it with the adjacency matrix of high-confidence causal edges, causal alignment verification is achieved, forming a comprehensive optimization criterion that balances prediction performance and causal interpretability.

[0137] In an exemplary embodiment, based on the causal transmission path, the causal relationship is mathematically modeled to obtain a structural causal model, including:

[0138] S501: For node variables with in-degree causal edges in a causal propagation path. Identify the set of causal parent nodes , Include variables , .

[0139] For example, the supplier evaluation terminal focuses on causal transmission paths, targeting node variables with in-degree causal edges in the path. This involves identifying the set of causal parent nodes. By traversing the topology of the causal transmission path, all pairs of nodes are selected. The preceding variables that produce a direct causal effect constitute the set of causal parent nodes of that node variable. The set contains , Variables, among which , … These are the time delay parameters corresponding to the causal effects of each preceding variable on Yj(t).

[0140] S502: Based on node variables With the set of causal parent nodes Establish structural equations and calculate node variables using the following formula. Theoretical values ​​of the variables:

[0141]

[0142] in It is a parameterizable function. These are the function parameters to be learned. It is the unobserved noise term.

[0143] For example, the supplier evaluation terminal targets the node variables already identified in the causal transmission path. Given the causal parent node set Pa(Yj) and its corresponding set of parent nodes, establish a structural equation between them, and calculate the node variables using this equation. The theoretical values ​​of the variables. The constructed structural equations adopt a parameterizable functional form, as shown in this diagram. and The quantitative correlation between them; the equations include parameters to be learned to adapt to the actual data patterns, and also include unobserved noise terms to cover potential minor influencing factors that the model has not captured, so as to ensure the completeness and rationality of the calculation of the theoretical values ​​of variables.

[0144] S503: Based on the theoretical values ​​of variables and historical time-series data, a set of parameterized structural equations is obtained through time-series regression or gradient descent optimization algorithms.

[0145] For example, the supplier evaluation terminal uses theoretical values ​​of variables as a basis, combined with collected and preprocessed historical time-series data, to perform parameter learning for structural equations. Mainstream optimization algorithms such as time-series regression or gradient descent are selected, with the goal of minimizing the difference between theoretical variable values ​​and historical actual observations, to iteratively optimize the parameters to be learned in the structural equations. When the parameters converge to a stable state or reach a preset stopping condition, the optimization process stops, at which point a structured equation with determined parameters is obtained. The parameterized structural equations corresponding to all node variables are then summarized and organized to obtain a set of parameterized structural equations.

[0146] S504: Integrate the parameterized set of structural equations to obtain the structural causal model.

[0147] For example, the parameterized set of structural equations is systematically integrated to construct a structural causal model. During the integration process, the topological logic of the causal transmission path is strictly followed, and the upstream and downstream relationships of the node variables corresponding to each structural equation in the causal chain are systematically analyzed to ensure consistency in causal logic and temporal alignment among the equations. Based on the analyzed relationships, a unified model framework is built, and all parameterized structural equations are systematically incorporated into the framework to form a structural causal model that can accurately characterize the quantitative causal relationships between variables in the causal transmission path.

[0148] The deep learning-based wind and solar resource prediction method provided in this application accurately identifies causal parent nodes, constructs suitable structural equations, and optimizes parameters, ultimately integrating them into a structured structural causal model. This achieves accurate characterization of quantitative causal relationships between variables in the causal transmission path, improves the reliability and rigor of causal modeling, and provides solid quantitative model support for subsequent counterfactual deduction and risk intervention effect prediction.

[0149] It should be understood that although the steps in the flowcharts of the embodiments described above are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the embodiments described above may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.

[0150] Based on the same inventive concept, this application also provides a big data-based supplier evaluation apparatus for implementing the aforementioned big data-based supplier evaluation method. The solution provided by this apparatus is similar to the implementation described in the above method; therefore, the specific limitations in one or more embodiments of the big data-based supplier evaluation apparatus provided below can be found in the limitations of the big data-based supplier evaluation method described above, and will not be repeated here.

[0151] In one exemplary embodiment, such as Figure 2 As shown, a supplier evaluation device 600 based on big data is provided, including:

[0152] The data acquisition and preprocessing module 601 is used to acquire multi-source heterogeneous time-series data related to suppliers and associated entities, and preprocess the multi-source heterogeneous time-series data to obtain the corresponding time-series event set and the time-series indicator sequence of each entity; the time-series event set includes entities, relationships, and timestamps;

[0153] The temporal causality discovery module 602 is used to identify the temporal causal relationship of temporal index sequences using a temporal causality discovery algorithm and to construct a partially directed acyclic graph with temporal causal edges.

[0154] The dynamic causal knowledge graph construction module 603 is used to combine the static relationships in the set of temporal events with the temporal causal relationships in some directed acyclic graphs to construct a dynamic causal knowledge graph.

[0155] The graph structure-aware representation generation module 604 is used to input a dynamic causal knowledge graph into a pre-trained graph neural network and extract the graph structure-aware representation vectors of the supplier nodes therein;

[0156] The risk score generation module 605 is used to generate a risk score for the supplier based on the graph structure perception representation vector.

[0157] The causal transmission path parsing module 606 is used to parse risk scores and obtain causal transmission paths based on the causal structure of a dynamic causal knowledge graph.

[0158] The structural causal model building module 607 is used to mathematically model causal relationships based on causal transmission paths to obtain structural causal models.

[0159] The report generation module 608 is used to input a set of time-series events into a structural causal model and generate a supply chain resilience analysis report containing a counterfactual risk score through counterfactual inference.

[0160] In one embodiment of the present invention, the dynamic causal knowledge graph construction module 603 is further configured to:

[0161] Based on a time-series event set, a static supply chain relationship graph is constructed with suppliers, products, raw materials, and regions as nodes;

[0162] Extract the time-series index sequences corresponding to the entity nodes in the static supply chain relationship graph to obtain the set of time-series variables to be analyzed;

[0163] The temporal causal discovery algorithm based on conditional independence testing examines the conditional independence between variables in the set of time-series variables to be analyzed under different time delays, and obtains a set of conditional independence test results.

[0164] Based on the set of conditional independence test results, the causal direction and time lag between variables are inferred, resulting in a partially directed acyclic graph with temporal causal edges;

[0165] By fusing temporal causal edges in a portion of a directed acyclic graph with static relationship edges in a static supply chain relationship graph, a dynamic causal knowledge graph is generated; each temporal causal edge is accompanied by causal effect strength and confidence information.

[0166] Based on the above embodiments, the dynamic causal knowledge graph construction module 603 is further used for:

[0167] For each pair of variables X in the set of time series variables to be analyzed i and X j Given the maximum time delay τ max Within the range, for all possible time delays τ, construct a condition set Z, which contains all conditions except X. i and X j The values ​​of other variables at and before time t;

[0168] Based on the observed sample data of the time series variable set, calculate the variable X under the condition set Z. i (t-τ) and variable X j The conditional independence between (t) is obtained by setting the set of conditional independence test statistics;

[0169] Calculate the set of p-values ​​corresponding to the set of conditional independence test statistics. The p-value represents the probability of observing the current test statistic or a more extreme case if the null hypothesis is true.

[0170] The set of p-values ​​is compared with a preset significance level threshold α. When the p-value is less than the significance level threshold α, the variable X is generated. i To variable X j A preliminary set of causal links with a time delay of τ;

[0171] The initial set of causal links is integrated into a causal relationship adjacency matrix;

[0172] By removing the cyclic dependencies of the causal adjacency matrix using a delooping algorithm in graph theory, a partially directed acyclic graph with temporally causal edges is obtained.

[0173] In one embodiment of the present invention, the causal transmission path analysis module 606 is further configured to:

[0174] Based on the causal edge structure in the dynamic causal knowledge graph, the attention weights between nodes are calculated during the message passing process of the pre-trained graph neural network to obtain a set of attention weights that represent the importance of the edges.

[0175] Based on the attention weight set, initialize the learnable edge mask matrix corresponding to the causal edge set in the dynamic causal knowledge graph;

[0176] The learnable edge mask matrix and the high-confidence causal edge adjacency matrix extracted from the dynamic causal knowledge graph are input into a preset function construction rule to construct an optimized objective function for generating the explanatory subgraph.

[0177] The optimal edge mask matrix is ​​obtained by optimizing the objective function;

[0178] Based on the optimal edge mask matrix, causal transmission paths are extracted from the dynamic causal knowledge graph.

[0179] Based on the above embodiments, the expression for the optimized objective function is as follows:

[0180]

[0181] Where M is the learnable edge mask matrix, f(G,M) represents the prediction output of the graph neural network running on the subgraph G weighted by the mask matrix M, and y is the original risk score prediction value of the supplier. The loss function term that measures the difference between the subgraph prediction and the original prediction. For the L1 norm regularization term of the mask matrix M, and To balance the hyperparameter weights of different loss terms, For the causal alignment loss function term, It is an adjacency matrix extracted from a dynamic causal knowledge graph, representing high-confidence causal edges. Item used to punish The concealment of important causal edges.

[0182] In one embodiment of the present invention, the structural causal model construction module 607 is further configured to:

[0183] For node variables with in-degree causal edges in a causal transmission path Identify the set of causal parent nodes , Include variables , ;

[0184] Based on node variables With the set of causal parent nodes Establish structural equations and calculate node variables using the following formula. Theoretical values ​​of the variables:

[0185]

[0186] in It is a parameterizable function. These are the function parameters to be learned. This is the unobserved noise term;

[0187] Based on the theoretical values ​​of variables and historical time series data, a set of parameterized structural equations is obtained through time series regression or gradient descent optimization algorithms.

[0188] By integrating the parameterized set of structural equations, a structural causal model is obtained.

[0189] In one embodiment, this application provides a computer device including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps in the above-described method embodiments.

[0190] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the steps in the above method embodiments.

[0191] For the device embodiments, since they basically correspond to the method embodiments, the relevant parts can be referred to in the description of the method embodiments. The device embodiments described above are merely illustrative. The components described as separate parts may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this disclosure according to actual needs. Those skilled in the art can understand and implement this without creative effort.

[0192] The above-described embodiments are merely illustrative of several implementation methods of the embodiments of this application, and their descriptions are relatively specific and detailed. However, they should not be construed as limiting the scope of the patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the embodiments of this application, and these modifications and improvements all fall within the protection scope of the embodiments of this application.

Claims

1. A supplier evaluation method based on big data, characterized in that, The method includes: Collect multi-source heterogeneous time-series data related to suppliers and associated entities, and preprocess the multi-source heterogeneous time-series data to obtain the corresponding time-series event set and the time-series index sequence of each entity; the time-series event set includes entity, relationship, and timestamp. The temporal causal discovery algorithm is used to identify the temporal causal relationship of the temporal index sequence and construct a partially directed acyclic graph with temporal causal edges; By combining the static relationships in the set of temporal events with the temporal causal relationships in the partial directed acyclic graph, a dynamic causal knowledge graph is constructed. The dynamic causal knowledge graph is input into a pre-trained graph neural network to extract the graph structure-aware representation vectors of the supplier nodes. Based on the graph structure perception representation vector, generate a risk score for the supplier; Based on the causal structure of the dynamic causal knowledge graph, the risk score is analyzed to obtain the causal transmission path; Based on the aforementioned causal transmission path, the causal relationship is mathematically modeled to obtain a structural causal model; The set of time-series events is input into the structural causal model, and a supply chain resilience analysis report containing a counterfactual risk score is generated through counterfactual inference.

2. The method according to claim 1, characterized in that, The construction of a dynamic causal knowledge graph by combining the static relationships in the set of temporal events with the temporal causal relationships in the partial directed acyclic graph includes: Based on the aforementioned time-series event set, a static supply chain relationship graph is constructed with suppliers, products, raw materials, and regions as nodes; Extract the time-series index sequences corresponding to the entity nodes in the static supply chain relationship graph to obtain the set of time-series variables to be analyzed; The temporal causal discovery algorithm based on conditional independence testing examines the conditional independence between variables in the set of time-series variables to be analyzed under different time delays, and obtains a set of conditional independence test results. Based on the set of conditional independence test results, the causal direction and time lag between variables are inferred, resulting in a partially directed acyclic graph with temporal causal edges. The temporal causal edges in the partial directed acyclic graph are fused with the static relationship edges in the static supply chain relationship graph to generate the dynamic causal knowledge graph; wherein each temporal causal edge is accompanied by causal effect strength and confidence information.

3. The method according to claim 2, characterized in that, Based on the set of conditional independence test results, the causal direction and time lag between variables are inferred, resulting in a partially directed acyclic graph with temporal causal edges, including: For the variable X in the set of time series variables to be analyzed i and X j Given the maximum time delay τ max Within the range, for all possible time delays τ, construct a condition set Z, which contains all conditions except X. i and X j The values ​​of other variables at and before time t; Based on the observed sample data of the time-series variable set, calculate the variable X under the condition set Z. i (t-τ) and variable X j The conditional independence between (t) is obtained by setting the set of conditional independence test statistics; Calculate the set of p-values ​​corresponding to the set of conditional independence test statistics, where each p-value represents the probability of observing the current test statistic or a more extreme case if the null hypothesis is true. The set of p-values ​​is compared with a preset significance level threshold α. When the p-value is less than the significance level threshold α, the variable X is generated. i To variable X j A preliminary set of causal links with a time delay of τ; The initial set of causal links is integrated into a causal adjacency matrix; By removing the cyclic dependencies of the causal adjacency matrix using a delooping algorithm in graph theory, a partially directed acyclic graph with temporally causal edges is obtained.

4. The method according to claim 1, characterized in that, The causal structure based on the dynamic causal knowledge graph is used to analyze the risk score and obtain the causal transmission path, including: Based on the causal edge structure in the dynamic causal knowledge graph, the attention weights between nodes are calculated during the message passing process of the pre-trained graph neural network to obtain a set of attention weights representing the importance of the edges. Based on the attention weight set, initialize the learnable edge mask matrix corresponding to the causal edge set in the dynamic causal knowledge graph; The learnable edge mask matrix and the high-confidence causal edge adjacency matrix extracted from the dynamic causal knowledge graph are input into a preset function construction rule to construct an optimized objective function for generating the explanatory subgraph. The optimal edge mask matrix is ​​calculated using the aforementioned optimization objective function; Based on the optimal edge mask matrix, the causal transmission path is extracted from the dynamic causal knowledge graph.

5. The method according to claim 4, characterized in that, The expression for the optimization objective function is: Where M is the learnable edge mask matrix, f(G,M) represents the prediction output of the graph neural network running on the subgraph G weighted by the mask matrix M, and y is the original risk score prediction value of the supplier. The loss function term that measures the difference between the subgraph prediction and the original prediction. For the L1 norm regularization term of the mask matrix M, and To balance the hyperparameter weights of different loss terms, For the causal alignment loss function term, It is an adjacency matrix representing high-confidence causal edges extracted from the dynamic causal knowledge graph. Item used to punish The concealment of important causal edges.

6. The method according to claim 1, characterized in that, The step of mathematically modeling the causal relationship based on the causal transmission path to obtain a structural causal model includes: For node variables with in-degree causal edges in the causal transmission path Identify the set of causal parent nodes The Include variables , ; Based on the node variables With the set of causal parent nodes Establish structural equations and calculate the node variables using the following formula. Theoretical values ​​of the variables: in It is a parameterizable function. These are the function parameters to be learned. This is the unobserved noise term; Based on the theoretical values ​​of the variables and historical time series data, a set of parameterized structural equations is obtained through time series regression or gradient descent optimization algorithms. By integrating the parameterized set of structural equations, the structural causal model is obtained.

7. A supplier evaluation device based on big data, characterized in that, The device includes: The data acquisition and preprocessing module is used to collect multi-source heterogeneous time-series data related to suppliers and associated entities, and preprocess the multi-source heterogeneous time-series data to obtain the corresponding time-series event set and the time-series index sequence of each entity; the time-series event set includes entities, relationships, and timestamps; The temporal causal discovery module is used to identify the temporal causal relationship of the temporal index sequence using the temporal causal discovery algorithm, and to construct a partially directed acyclic graph with temporal causal edges; The dynamic causal knowledge graph construction module is used to combine the static relationships in the set of temporal events with the temporal causal relationships in the partial directed acyclic graph to construct a dynamic causal knowledge graph. The graph structure-aware representation generation module is used to input the dynamic causal knowledge graph into a pre-trained graph neural network and extract the graph structure-aware representation vectors of the supplier nodes therein; The risk score generation module is used to generate a risk score for the supplier based on the graph structure perception representation vector. The causal transmission path parsing module is used to parse the risk score based on the causal structure of the dynamic causal knowledge graph to obtain the causal transmission path; The structural causal model construction module is used to mathematically model the causal relationship based on the causal transmission path to obtain a structural causal model. The report generation module is used to input the time series event set into the structural causal model and generate a supply chain resilience analysis report containing a counterfactual risk score through counterfactual inference.

8. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 7.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 7.

Images