A double-path graph representation learning method for enterprise ESG rating task

Abstract

An enterprise ESG rating task-oriented double-path graph representation learning method is provided in the embodiment of the application, and belongs to the technical field of data processing, and specifically comprises the following steps: step 1, constructing an ESG heterogeneous graph atlas with a target enterprise as a center node, and extracting a target subgraph of multiple types of nodes and their corresponding relationships from the k-hop neighborhood thereof; step 2, dynamically selecting a structure enhancement mode for the target subgraph based on a learnable enhancement strategy, and generating multiple enhanced views with different structures but consistent semantics; step 3, performing semantic smoothing and structure denoising processing on each enhanced view to improve semantic consistency and structure purity; step 4, aggregating multiple denoised enhanced views to form a unified enterprise graph representation vector; and step 5, inputting the enterprise graph representation vector into a classification model to predict the enterprise ESG rating or risk category. Through the scheme of the application, the rating accuracy and adaptability are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of data processing, and in particular to a dual-path graph representation learning method for enterprise ESG rating tasks. BACKGROUND

[0002] At present, in the ESG evaluation system, the prediction and evaluation of enterprise ESG grades are realized by constructing the correlation network between enterprises (such as holding relationship, environmental events, topic reports, governance structure, etc.), which has become an important basis for enterprise rating and investment decision-making. Traditional ESG rating methods mostly rely on expert scoring, quantitative statistical indicators or supervised learning models based on structured data, but these methods face many challenges in practical application.

[0003] Therefore, there is an urgent need for a dual-path graph representation learning method for enterprise ESG rating tasks that can simultaneously handle graph structure perturbation, semantic noise and multi-view learning ability to improve the accuracy and adaptability of enterprise rating tasks in real complex scenarios. SUMMARY

[0004] Therefore, the embodiments of the present application provide a dual-path graph representation learning method for enterprise ESG rating tasks, which at least partially solves the problem of poor precision and adaptability in the prior art.

[0005] The embodiments of the present application provide a dual-path graph representation learning method for enterprise ESG rating tasks, comprising:

[0006] Step 1, constructing an ESG heterogeneous graph with a target enterprise as a center node, and extracting a target subgraph of multiple types of nodes and their corresponding relationships from the k-hop neighborhood thereof, wherein the multiple types of nodes include enterprises, senior managers, events, topics and supervision;

[0007] Step 2, dynamically selecting a structure enhancement method for the target subgraph based on a learnable enhancement strategy, generating multiple enhanced views with different structures but consistent semantics;

[0008] Step 3, performing semantic smoothing and structure denoising processing on each enhanced view to improve semantic consistency and structure purity;

[0009] Step 4, aggregating multiple denoised enhanced views to form a unified enterprise graph representation vector;

[0010] Step 5, inputting the enterprise graph representation vector into a classification model to predict the enterprise ESG grade or risk category.

[0011] According to a specific implementation mode of the embodiments of the present application, the step 1 specifically comprises:

[0012] Step 1.1. Constructing an ESG heterogeneous graph from enterprise multi-source data

[0013] ;

[0014] wherein, denotes a set of all nodes in the graph, denotes a set of edges, denotes a node type mapping function, for all possible node types, denotes an edge type mapping function, for all relationship types, denotes a feature matrix of all nodes, wherein is an attribute vector of node ;

[0015] Step 1.2. Extracting an initial subgraph of the k-hop neighborhood of the target enterprise node from the ESG heterogeneous graph, and filtering the initial subgraph through a node type filtering function and an edge confidence screening function to obtain a target subgraph, wherein the node type filtering function is

[0016] ;

[0017] wherein, denotes a set of node types meeting the ESG evaluation requirements, denotes a type label of node v, and if the label falls within , the node is retained;

[0018] the edge confidence screening function is

[0019] ;

[0020] wherein, denotes a confidence score of edge , is an adjustable threshold value.

[0021] According to a specific implementation manner of an embodiment of the present application, the step 2 specifically comprises:

[0022] Step 2.1. Dividing the target subgraph into a plurality of substructures, and defining a set of candidate enhancement strategies for each substructure;

[0023] Step 2.2. Dynamically selecting an enhancement strategy from the set of candidate enhancement strategies through a learnable strategy selector to generate a plurality of enhanced views.

[0024] According to a specific implementation manner of an embodiment of the present application, the step 2.2 specifically comprises:

[0025] Step 2.2.1, for each target subgraph The subgraph embedding is obtained by encoding using a graph neural network:

[0026] ;

[0027] in, This is a graph-level aggregation operation. Represents the set of real numbers. For embedded dimensions, Represents a d-dimensional real vector;

[0028] Step 2.2.2, embed the subgraph Input the policy selector MLP to generate the unnormalized score of the augmentation policy:

[0029] ;

[0030] in: For the number of candidate enhancement strategies, , These are learnable parameters;

[0031] Step 2.2.3: Use the Gumbel-Softmax approximation sampling strategy to select the enhancement strategy and generate the enhanced view.

[0032] According to a specific implementation of an embodiment of the present invention, step 3 specifically includes:

[0033] Step 3.1: Use the semantic smoothing module to perform fuzzy fusion on semantically similar relationship types in each enhanced view;

[0034] Step 3.2: Remove low-confidence edges using the structural denoising module, and retain high-risk propagation edges;

[0035] Step 3.3: Combine the semantic and structural denoising results to generate the purified subgraph representation.

[0036] According to a specific implementation of an embodiment of the present invention, step 3.1 specifically includes:

[0037] Step 3.1.1, assume that there are common elements in the enhanced view. There are several relation types, and for each relation type Initialize a learnable embedding vector:

[0038] ;

[0039] Step 3.1.2, for each pair of relation types Calculate their semantic similarity :

[0040] ;

[0041] Step 3.1.3, for each edge the relationship type carried , the corresponding similarity values of r and all K candidate relationships are taken out from the semantic similarity, and are spliced into a similarity vector , and a suitable fuzzy relationship type is sampled from the set of all semantically similar candidate relationships The sampling method is a differentiable Gumbel-Softmax sampling:

[0042] ;

[0043] wherein, is the relationship type and the similarity vector between the K candidate relationship types, temperature parameter, the final weight used for fuzzy fusion;

[0044] Step 3.1.4, the original relationship type embedding is represented as a weighted average of its multiple semantic neighbor relationship embeddings:

[0045] ;

[0046] wherein, is the set of relationships semantically similar to the relationship ;

[0047] Step 3.1.5, finally, the relationship embedding of each edge in the original enhanced view after fuzzy fusion is replaced to construct a semantically consistent subgraph structure:

[0048] ;

[0049] wherein, denotes the node set in the subgraph, denotes the edge set.

[0050] According to a specific implementation manner of an embodiment of the present application, the step 3.2 specifically comprises:

[0051] Step 3.2.1, based on the local node features and relationship embeddings of the edges in the enhanced view, the local confidence score of the edge is predicted through the edge scorer;

[0052] Step 3.2.2, based on the preset high-risk node set in the enhanced view, the path conduction factor of each edge in the risk propagation path is calculated;

[0053] Step 3.2.3, fusing the local confidence score and the path conduction factor to obtain the final confidence of the edge, and screening the edge set after structure refining based on a preset threshold to obtain the enhanced view after structure denoising.

[0054] According to a specific implementation manner of an embodiment of the present application, the step 3.3 specifically includes:

[0055] Step 3.3.1, fusing the fuzzy fused relationship embedding and the edge set after structure refining to construct a fusion subgraph with purified structure and consistent semantics;

[0056] Step 3.3.2, inputting the fusion subgraph into a graph neural network encoder, and obtaining a graph-level embedding representation of the fusion subgraph as the enhanced view after denoising through a graph-level aggregation function.

[0057] According to a specific implementation manner of an embodiment of the present application, the step 4 specifically includes:

[0058] Step 4.1, calculating the risk density of each enhanced view;

[0059] Step 4.2, setting a view aggregation weight according to the risk density;

[0060] Step 4.3, weightedly aggregating the graph representations of multiple enhanced views to obtain a final enterprise graph representation vector.

[0061] According to a specific implementation manner of an embodiment of the present application, the step 5 specifically includes:

[0062] Step 5.1, using a classification model to predict the ESG level of the enterprise graph representation vector;

[0063] Step 5.2, constructing a weighted loss function combined with structure, semantics and risk for model training, wherein the weighted loss function is

[0064] ;

[0065] wherein, is a loss weighting factor, is a standard cross-entropy loss, is a risk weighted loss, is the average value of the entropy of the edge relationship distribution of the enterprise graph representation vector, is a weight parameter for regulating the contribution of each item;

[0066] Step 5.3, outputting the ESG prediction label of the target enterprise by using the trained classification model.

[0067] The double-path graph representation learning scheme for the enterprise ESG rating task in the embodiment of the application comprises the following steps: step 1, constructing an ESG heterogeneous graph atlas with a target enterprise as a center node, and extracting a target subgraph of multiple types of nodes and their corresponding relationships from the k-hop neighborhood of the target enterprise, wherein the multiple types of nodes include enterprises, senior managers, events, topics and supervision; step 2, dynamically selecting a structure enhancement mode for the target subgraph based on a learnable enhancement strategy, and generating multiple enhanced views with different structures but consistent semantics; step 3, performing semantic smoothing and structure denoising processing on each enhanced view to improve semantic consistency and structure purity; step 4, aggregating multiple denoised enhanced views to form a unified enterprise graph representation vector; and step 5, inputting the enterprise graph representation vector into a classification model to predict the enterprise ESG rating or risk category.

[0068] The beneficial effects of the embodiment of the application are as follows: through the scheme of the application, the subgraph division and learnable enhancement strategy generation module are introduced, multiple enhanced views with diverse structures and consistent semantics can be constructed for different enterprises. Through node covering, edge disturbance, subgraph exchange and other operations, the structural changes that may occur in the enterprise graph atlas in the real world are simulated under the premise of preserving semantic consistency, thereby improving the generalization ability and robustness of the model; for the redundant relationships, inconsistent expressions and low-confidence relationship edges widely existing in the enterprise graph atlas, the application proposes a denoising strategy combining semantic ambiguity and structure refinement. The semantic module performs fuzzy fusion on synonymous relationships, and the structure module identifies and removes noise edges in the graph through a learnable edge weight mechanism, effectively improving the quality of the graph representation; the application can flexibly handle heterogeneous nodes (such as enterprises, senior managers, events and media) and multiple types of relationships in the ESG scenario, adapt to the ESG multi-dimensional index system, and has strong universality and expandability. BRIEF DESCRIPTION OF DRAWINGS

[0069] In order to more clearly illustrate the technical solutions of the embodiments of the application, the following will briefly introduce the drawings needed in the embodiments. Obviously, the drawings in the following description are only some embodiments of the application, and other drawings can be obtained by those skilled in the art without creative labor.

[0070] Figure 1 A flowchart of a double-path graph representation learning method for an enterprise ESG rating task provided by the embodiment of the application is shown. DETAILED DESCRIPTION

[0071] The embodiments of the application will be described in detail below with reference to the drawings.

[0072] Following, the embodiments of the present application will be described by specific examples, and other advantages and effects of the present application can be easily understood by those skilled in the art from the disclosure. Obviously, the described embodiments are only a part of the embodiments of the present application, and not all the embodiments. The present application can also be implemented or applied by other different specific embodiments, and various modifications or changes can be made based on different views and applications without departing from the spirit of the present application. It should be noted that the following embodiments and features in the embodiments can be combined with each other without conflict. Based on the embodiments of the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0073] It should be noted that various aspects of the embodiments described below are within the scope of the appended claims. It should be apparent that the aspects described herein can be embodied in a wide variety of forms and that any specific structure and / or function described herein is merely illustrative. Based on the teachings herein one skilled in the art should appreciate that an aspect described herein can be implemented independently of any other aspects and that an aspect can be implemented both as any number of software, firmware, and / or hardware structures.

[0074] It should also be noted that the figures provided in the following embodiments are only schematically illustrate the basic concept of the present application, and only show the components related to the present application in the figures, not drawn according to the number, shape and size of the components in actual implementation, the type, number and proportion of each component in actual implementation can be a random change, and the component layout type can also be more complex.

[0075] In addition, in the following description, specific details are provided in order to facilitate a thorough understanding of examples. However, one skilled in the art will understand that the aspects described can be practiced without these specific details.

[0076] First, ESG related data usually exists in the form of multi-source heterogeneous, strong timeliness and complex structure, such as news reports, policy documents, enterprise announcements and other unstructured information, which are often extracted as various forms of graph structure data. How to effectively represent these graph structures and extract robust and generalizable enterprise features from them is a key problem in current ESG graph representation learning.

[0077] Secondly, there are inevitably noise information in the actual graph, such as exaggerated topics, relationship redundancy or inconsistent expressions (such as "reported" and "complaints exist") and the like. These information interfere with the understanding of the real state of the model for the enterprise, and reduce the accuracy and interpretability of the prediction.

[0078] Finally, the traditional graph learning method has strong dependence on the graph structure, and is easily affected by the structure disturbance, resulting in poor robustness of the model in the scene of incomplete or frequent evolution data, and the model cannot adapt to the dynamics and diversity of the ESG graph.

[0079] The embodiment of the application provides a double-path graph representation learning method for an enterprise ESG rating task, which can be applied to the ESG rating process in the enterprise management scene.

[0080] Referring to Figure 1 , a flowchart of a double-path graph representation learning method for an enterprise ESG rating task provided by the embodiment of the application is shown. As Figure 1 shown, the method mainly includes the following steps:

[0081] Step 1, constructing an ESG heterogeneous graph with a target enterprise as a center node, and extracting a target subgraph of multiple types of nodes and their corresponding relationships from the k-hop neighborhood thereof, wherein the multiple types of nodes include enterprises, senior managers, events, topics and supervision;

[0082] In specific implementation, the specific process of constructing an ESG heterogeneous graph with a target enterprise as a center node and extracting a target subgraph of multiple types of nodes and their corresponding relationships from the k-hop neighborhood thereof can be as follows:

[0083] Step 1.1, obtaining the heterogeneous graph of enterprise ESG from the enterprise multi-source data through entity recognition (NER), relationship extraction (RE), rule matching and the like, and representing the graph as:

[0084] ;

[0085] Wherein:

[0086] represents the set of all nodes in the graph;

[0087] represents the set of edges;

[0088] represents a node type mapping function, for all possible node types (such as enterprises, topics, policies, senior managers, etc.);

[0089] represents an edge type mapping function, For all relationship types (such as holding, investment, association, punishment, etc.);

[0090] is the feature matrix of all nodes, where is the attribute vector of node (such as industry classification, carbon emissions, media sentiment score, etc.).

[0091] In addition, enterprise multi-source data can include: structured data such as annual reports, public announcements, unstructured topic data such as media reports, social platforms, and field data such as policy regulations, regulatory information, and industry chain databases.

[0092] Step 1.2, taking the target enterprise as the center node, extracting its closed subgraph from the graph. The extraction process is as follows:

[0093] Let the target enterprise node be , and the hop parameter be . Then the target subgraph is denoted as:

[0094] ;

[0095] Where:

[0096] , represents the set of nodes with a shortest path length of no more than between ;

[0097] , represents the edge set between these nodes;

[0098] represents the shortest path length between node and node in graph .

[0099] To improve the quality of the subgraph and semantic relevance, type filtering functions and weight threshold controls can be added in the node and edge screening process, as follows:

[0100] (1) Define the node type filtering function:

[0101] ;

[0102] Where, represents the set of node types that meet the ESG evaluation requirements, represents the "type label" of node v, such as "enterprise", "topic", "high-level official", etc. If its label falls within , the node is retained.

[0103] (2) Define the edge confidence screening function:

[0104] ;

[0105] wherein, represents the confidence score of edge . is an adjustable threshold.

[0106] (3) The finally constructed subgraph:

[0107] ;

[0108] wherein,

[0109] ;

[0110] ;

[0111] The subgraph retains the core structural neighborhood of the target enterprise under the ESG semantics, and is used for subsequent enhancement, denoising and rating task modeling.

[0112] Step 2, dynamically select the structural enhancement method for the target subgraph based on the learnable enhancement strategy, and generate multiple enhanced views with different structures but consistent semantics;

[0113] In specific implementation, the specific process of generating multiple enhanced views with different structures but consistent semantics is as follows:

[0114] Step 2.1, divide the extracted target enterprise subgraph into multiple structural modules (sub-subgraphs), and design an enhancement candidate space for it. The specific operation is as follows:

[0115] Step 2.1.1, divide the subgraph into several substructures by using a community detection algorithm (such as Louvain algorithm or label propagation). Through the substructure modules obtained by division, provide semantic boundaries for the subsequent enhancement strategy:

[0116] ;

[0117] wherein, represents the node set of the th subgraph community (sub-subgraph).

[0118] Step 2.1.2, define a set of candidate enhancement strategies for each subgraph, including:

[0119] ;

[0120] Each of the above strategies can act on the subgraph structure, simulate possible data disturbance and graph structure change, and enhance the robustness of the model.

[0121] Step 2.2. To make the structural reinforcement process adaptive and differentiable, this step constructs a set of reinforcement strategy selectors for each subgraph, dynamically selects the optimal reinforcement strategy based on its structural semantic features, and generates a structural perturbation view for subsequent representation learning. The specific process is as follows:

[0122] Step 2.2.1. For each target enterprise subgraph , first use a graph neural network (such as GIN) to encode to obtain a graph-level embedding representation:

[0123] ;

[0124] where is a graph-level aggregation operation (such as mean-pooling), is the embedding dimension.

[0125] Step 2.2.2. Input the subgraph embedding into the strategy selector MLP to generate the non-normalized score of the reinforcement strategy:

[0126] ;

[0127] where:

[0128] is the number of candidate reinforcement strategies;

[0129] , are learnable parameters.

[0130] Step 2.2.3. To realize the differentiable training of the strategy selection process, introduce the Gumbel-Softmax approximate one-hot sampling strategy:

[0131] ;

[0132] where:

[0133] is the approximate one-hot vector;

[0134] is the temperature parameter, when the approximate hard sampling, when the approximate uniform distribution;

[0135] is a random variable sampled from a standard Gumbel distribution.

[0136] Step 2.2.4. Each reinforcement strategy can act on the original subgraph , generate an enhanced view . Finally, the strategy corresponding to the maximum probability is selected to generate the enhanced view:

[0137] ;

[0138] The above process can be repeated to generate multiple structural perturbation versions of the enhanced view for each target subgraph for subsequent semantic and structural denoising modules:

[0139] .

[0140] Step 3, perform semantic smoothing and structural denoising processing on each enhanced view to improve semantic consistency and structural purity;

[0141] In specific implementation, the specific process of performing semantic smoothing and structural denoising processing on each enhanced view to improve semantic consistency and structural purity is as follows:

[0142] Step 3.1, based on the semantic smoothing module of the differentiable relationship fusion mechanism, the edges of similar semantics are merged, and the specific process is as follows:

[0143] Step 3.1.1, suppose there are kinds of edge types (relationship types) in the graph, initialize a learnable embedding vector for each relationship type :

[0144] ;

[0145] wherein is the relationship embedding dimension.

[0146] Step 3.1.2, for each pair of relationship types , calculate its semantic similarity :

[0147] ;

[0148] The similarity is used to identify relationships that are similar in semantics (such as "reported" and "complaint").

[0149] Step 3.1.3, for each edge carrying the relationship type , first take out the corresponding similarity values of r and all K candidate relationships from the semantic similarity calculated in step 3.1.2 to form a similarity vector , and then sample a suitable fuzzy relationship type from its set of all semantically similar candidate relationships. The sampling method is differentiable Gumbel-Softmax sampling:

[0150] ;

[0151] in:

[0152] For relation type and Similarity vectors between candidate relation types;

[0153] For temperature parameters;

[0154] These are the weights ultimately used for fuzzy fusion.

[0155] Step 3.1.4: Represent the original relation type embedding as a weighted average of its multiple semantically nearest relation embeddings:

[0156] ;

[0157] in, For relationship A set of semantically similar relationships.

[0158] Step 3.1.5: Finally, embed the fuzzy fusion relationship into each edge of the original graph. This substitution allows for the construction of semantically consistent and less noisy subgraph structures:

[0159] ;

[0160] in, Describe the set of nodes in the subgraph. Describe the set of edges.

[0161] Step 3.2 involves a structural denoising mechanism that integrates local features with upstream and downstream risk propagation logic to further remove low-quality or invalid edges from the ESG graph, while retaining important structures with potential risk transmission relationships. The specific process is as follows:

[0162] Step 3.2.1, Local Edge Confidence Calculation: First, based on the local information of the edge, predict whether it is a structural noise edge. The specific method is as follows:

[0163] (1) For each edge in the graph Construct its representation:

[0164] ;

[0165] in:

[0166] They are nodes Embedding;

[0167] the relation embedding output by the aforementioned semantic smoothing module;

[0168] denotes vector concatenation.

[0169] (2) Input representations into the edge scorer to predict the local confidence score of the edge:

[0170] ;

[0171] where is a sigmoid function, and are learnable parameters.

[0172] Step 3.2.2, modeling of upstream and downstream risk path factors: to supplement the limitations of local feature judgment, introduce a path-level risk transmission factor to evaluate the importance of the edge in the ESG risk propagation chain.

[0173] (1) Let there be a set of high-risk nodes (such as ESG violation enterprises or reported entities) in the graph, which constitute the risk source set .

[0174] (2) For any edge , calculate its path transmission factor from the risk source:

[0175] ;

[0176] where:

[0177] denotes the set of all paths from any to the edge ;

[0178] denotes the length of the path ;

[0179] is a risk attenuation factor, the longer the path, the smaller the propagation impact.

[0180] The larger the value, the more likely it is that the edge is located in the risk transmission chain and should be given priority to retain.

[0181] Step 3.2.3, fuse local score and path transmission factor to calculate the final edge confidence:

[0182] ;

[0183] where is the weight coefficient, which regulates the relative importance of the two kinds of information.

[0184] Based on the fusion confidence, set the edge screening threshold , construct the edge set after structure refinement:

[0185] ;

[0186] Finally, the subgraph after structure denoising is obtained:

[0187] ;

[0188] Step 3.3, fuse the processing results of the semantic smoothing and structure refinement parts to generate subgraph embedding for enterprise ESG task. The specific steps are as follows:

[0189] Step 3.3.1, based on the edge relationship embedding set after semantic denoising and the edge set after structure denoising , the graph after structure purification and semantic consistency is obtained after fusion:

[0190] ;

[0191] Each edge is accompanied by the embedding after semantic smoothing , and the confidence after upstream and downstream risk adjustment has been used for structure refinement.

[0192] Step 3.3.2, input the fused subgraph into a graph neural network model (such as GCN, GIN, or heterogeneous GNN) to obtain a graph-level embedding:

[0193] ;

[0194] Where:

[0195] represents a graph encoder with relationship embedding;

[0196] represents a graph-level aggregation function, such as average pooling, attention pooling, etc.

[0197] is the final structure-semantic joint representation of the enterprise subgraph.

[0198] This representation will be aggregated with other enhanced view results in subsequent steps for the final ESG level prediction task.

[0199] Step 4, aggregate multiple denoised enhanced views to form a unified enterprise graph representation vector;

[0200] In practice, the specific process of aggregating multiple denoised and enhanced views to form a unified enterprise graph representation vector is as follows:

[0201] Step 4.1, considering that corporate risks in the ESG graph mainly spread through paths such as "holding chains," "supply chains," and "topic chains," introduces a perception mechanism for ESG risk propagation and integrates a graph representation of multiple enhanced views. Build target enterprise Unified representation vector In step 3.2.2, the structural denoising module has calculated the impact factor of each edge in the risk propagation path. Based on this, the proportion of high-risk edges in each view is calculated:

[0202] ;

[0203] in:

[0204] Risk threshold;

[0205] Represents a view Risk density;

[0206] For indicator functions, satisfying , .

[0207] Step 4.2: Set the view aggregation weights based on risk density; here, an exponential weighting method is used.

[0208] ;

[0209] in:

[0210] This is a preset positive real number used to adjust the intensity of the impact of risk density on view weights.

[0211] Step 4.3, the final aggregation is represented as:

[0212] ;

[0213] According to a specific implementation of an embodiment of the present invention, step 5 specifically includes:

[0214] Step 5.1: Use a lightweight supervised classification model to predict labels on the graph representation. By default, a fully connected feedforward neural network (MLP) is used to build the classifier.

[0215] ;

[0216] in:

[0217] is the weight matrix of the two-layer classifier;

[0218] is the bias term;

[0219] is the number of classes, i.e., the total number of ESG rating labels;

[0220] is the predicted class probability distribution.

[0221] According to the business scenario, the classification model can be replaced by a more complex structure (such as Transformer, graph attention network, etc.), but it remains decoupled from the graph embedding module.

[0222] Step 5, input the enterprise graph representation vector into the classification model to predict the enterprise ESG rating or risk category.

[0223] In specific implementation, the specific process of inputting the enterprise graph representation vector into the classification model to predict the enterprise ESG rating or risk category specifically includes:

[0224] Step 5.1, use a lightweight supervised classification model to perform label prediction on the graph representation, and by default use a fully connected feedforward neural network (MLP) to construct the classifier:

[0225] ;

[0226] wherein:

[0227] is the weight matrix of the two-layer classifier;

[0228] is the bias term;

[0229] is the number of classes, i.e., the total number of ESG rating labels;

[0230] is the predicted class probability distribution.

[0231] According to the business scenario, the classification model can be replaced by a more complex structure (such as Transformer, graph attention network, etc.), but it remains decoupled from the graph embedding module.

[0232] Step 5.2, on the basis of the traditional cross-entropy loss, fuse the three types of intermediate parameters of structural confidence, risk density and semantic ambiguity obtained in the previous steps to construct a structure-semantic-risk joint weighted loss function, and improve the robustness of the model to uncertain information and the sensitivity to risk-oriented structure. Specifically as follows:

[0233] Step 5.2.1, for the final graph after aggregation , the standard cross-entropy loss is used as the basic supervision signal:

[0234] ;

[0235] wherein:

[0236] is the one-hot encoding of the true label;

[0237] is the softmax output probability;

[0238] is the number of classes, i.e. the number of ESG rating categories.

[0239] Step 5.2.2, the edge confidence output in the structure denoising module is used to calculate the overall confidence of the graph structure as a loss weighting factor:

[0240] ;

[0241] The higher the confidence, the cleaner the sample structure, and the higher the weight should be given.

[0242] Step 5.2.3, with the multi-view risk density distribution calculated in the aggregation stage , set the risk-oriented weight , and build a risk-weighted loss based on individual supervision of each view:

[0243] ;

[0244] wherein is the modulation factor, the higher the proportion of high-risk views, the greater the weight.

[0245] Step 5.2.4, from the semantic smoothing module, calculate the entropy value of each edge relationship distribution , representing its semantic uncertainty, and take the mean value of the whole graph:

[0246] ;

[0247] wherein is the relationship type distribution after blurring of the edge . The higher the blurring degree, the more cautious the training should be.

[0248] Step 5.2.5, combine the above three modulation factors to build the final supervision objective function of the present application as follows:

[0249] ;

[0250] ​wherein:

[0251] weight parameters (can be set or jointly optimized) for regulating the contributions of each sub-item;

[0252] The three items respectively regulate structural reliability, risk focus, and semantic certainty.

[0253] Step 5.3, the final output is the target enterprise ESG prediction label:

[0254] .

[0255] The dual-path graph representation learning method for enterprise ESG rating tasks provided by the embodiment can construct multiple enhanced views with diverse structures and consistent semantics for different enterprises by introducing a subgraph division and a learnable enhancement strategy generation module. By means of node covering, edge disturbance, subgraph exchange and other operations, structural changes that may occur in the enterprise graph in the real world are simulated under the premise of preserving semantic consistency, thereby improving the generalization ability and robustness of the model. For the redundant relationships, inconsistent expressions and low-confidence relationship edges widely existing in the enterprise graph, the invention proposes a denoising strategy combining semantic ambiguity and structural refinement. The semantic module performs fuzzy fusion on synonymous relationships, and the structural module identifies and removes noise edges in the graph through a learnable edge weight mechanism, effectively improving the quality of graph representation. The invention can flexibly handle heterogeneous nodes (such as enterprises, executives, events, media) and multi-type relationships in the ESG scenario, adapt to the multi-dimensional index system of ESG, and has strong universality and expandability.

[0256] It should be understood that parts of the present application can be realized by hardware, software, firmware or a combination thereof.

[0257] The above is merely a specific implementation of the present application, but the protection scope of the present application is not limited thereto. Any changes or replacements within the technical scope disclosed by the present application can be easily thought of by those skilled in the art, and should be covered within the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

Claims

1. A dual-path graph representation learning method for enterprise ESG rating tasks, characterized in that, include: Step 1: Construct an ESG heterogeneous graph with the target enterprise as the central node, and extract the target subgraph with multiple types of nodes and their corresponding relationships from its k-hop neighborhood. The multiple types of nodes include enterprises, executives, events, topics and regulations. Step 2: Dynamically select structural enhancement methods for the target subgraph based on learnable enhancement strategies to generate multiple enhanced views with different structures but consistent semantics; Step 2 specifically includes: Step 2.1: Divide the target subgraph into multiple substructures and define a set of candidate enhancement strategies for each substructure; Step 2.2: Dynamically select an enhancement strategy from the candidate enhancement strategy set using a learnable strategy selector to generate multiple enhancement views; Step 2.2 specifically includes: Step 2.2.1, for each target subgraph The subgraph embedding is obtained by encoding using a graph neural network: ； in, This is a graph-level aggregation operation. Represents the set of real numbers. For embedded dimensions, Represents a d-dimensional real vector; Step 2.2.2, embed the subgraph Input the policy selector MLP to generate the unnormalized score of the augmentation policy: ； in: For the number of candidate enhancement strategies, , These are learnable parameters; Step 2.2.3: Use the Gumbel-Softmax approximation sampling strategy to select the enhancement strategy and generate the enhanced view; Step 3: Perform semantic smoothing and structural denoising on each enhanced view to improve semantic consistency and structural purity; Step 3 specifically includes: Step 3.1: Use the semantic smoothing module to perform fuzzy fusion on semantically similar relationship types in each enhanced view; Step 3.2: Remove low-confidence edges using the structural denoising module, and retain high-risk propagation edges; Step 3.3: Combine the semantic and structural denoising results to generate the purified subgraph representation; Step 3.1 specifically includes: Step 3.1.1, assume that there are common elements in the enhanced view. There are several relation types, and for each relation type Initialize a learnable embedding vector: ； Step 3.1.2, for each pair of relation types Calculate their semantic similarity : ； Step 3.1.3, for each edge The type of relationship carried Extract the corresponding similarity values ​​of r with all K candidate relations from the semantic similarity and concatenate them into a similarity vector. Then, sample appropriate fuzzy relation types from the set of all semantically similar candidate relations. The sampling method is differentiable Gumbel-Softmax sampling: ； in, For relation type and Similarity vectors between candidate relation types For temperature parameters, These are the weights ultimately used for fuzzy fusion; Step 3.1.4: Represent the original relation type embedding as a weighted average of its multiple semantically nearest relation embeddings: ； in, For relationship A set of semantically similar relations; Step 3.1.5: Finally, embed the relationships of each edge in the original enhanced view using fuzzy fusion. By substitution, a semantically consistent subgraph structure is constructed: ； in, Describe the set of nodes in the subgraph. Describe the edge set; Step 3.2 specifically includes: Step 3.2.1: Based on the local node features and relationship embeddings of edges in the enhanced view, predict the local confidence score of the edges using the edge scorer; Step 3.2.2: Based on the preset set of high-risk nodes in the enhanced view, calculate the path transmission factor of each edge in the risk propagation path; Step 3.2.3: Combine the local confidence score and the path transmission factor to obtain the final confidence of the edge, and filter out the set of edges after structural refinement based on the preset threshold to obtain the enhanced view after structural denoising; Step 4: Aggregate multiple denoised and enhanced views to form a unified enterprise graph representation vector; Step 5: Input the enterprise graph representation vector into the classification model to predict the enterprise's ESG rating or risk category.

2. The method according to claim 1, characterized in that, Step 1 specifically includes: Step 1.1: Construct an ESG heterogeneous map from enterprise multi-source data. ； in, This represents the set of all nodes in the graph. Describe the set of edges. Represents the node type mapping function, For all possible node types, Represents the edge type mapping function. For all relation types, Let represent the feature matrix of all nodes, where For nodes Attribute vectors; Step 1.2, with the target enterprise node Centered on a target subgraph, an initial subgraph of k-hop neighborhoods is extracted from the ESG heterogeneous graph. This initial subgraph is then filtered using a node type filtering function and an edge confidence filtering function to obtain the target subgraph. The node type filtering function is as follows: ； in, This represents the set of node types that meet the requirements of ESG assessment. The type label of node v is indicated if its label falls within the range of 100. If the node is inside, then that node will be retained; The edge confidence filtering function is: ； in, Representing an edge Confidence score, The threshold is adjustable.

3. The method according to claim 2, characterized in that, Step 3.3 specifically includes: Step 3.3.1: Merge the fuzzy fused relationship embedding with the structurally refined edge set to construct a fused subgraph with purified structure and consistent semantics; Step 3.3.2: Input the fused subgraph into the graph neural network encoder, and obtain the graph-level embedding representation of the fused subgraph as the enhanced view after denoising through the graph-level aggregation function.

4. The method according to claim 3, characterized in that, Step 4 specifically includes: Step 4.1, calculate the risk density for each enhanced view; Step 4.2: Set the view aggregation weights based on risk density; Step 4.3: Weighted aggregation of graph representations from multiple enhanced views yields the final enterprise graph representation vector.

5. The method according to claim 4, characterized in that, Step 5 specifically includes: Step 5.1: Use a classification model to predict the ESG level of the enterprise graph representation vector; Step 5.2: Construct a structure-semantic-risk joint weighted loss function for model training, wherein the weighted loss function is... ； in, As the loss weighting factor, For standard cross-entropy loss, Risk-weighted loss, Let the mean entropy value be the distribution of the relationships along each edge of the enterprise graph vector. Weighting parameters for adjusting the contribution of each component; Step 5.3: Use the trained classification model to output the ESG prediction label of the target enterprise.

Images