Training Method and System for Green and Low-Carbon Evaluation Model Based on Multiple Supplier Types

Abstract

This invention discloses a training method and system for a green and low-carbon evaluation model based on multiple types of suppliers. The method includes: first, acquiring a set of green and low-carbon evaluation indicators and determining indicator sample units. Then, acquiring two types of sample propagation chains, one with supplier sample entities and the other with indicator sample units as source nodes, to characterize their topological structure. After extracting the state feature representations of each propagation chain, processing them through two branches of an initial evaluation network: the first branch obtains the feature representation of the indicator sample unit based on the features sourced from the indicators; the second branch combines the output features of the first branch with the features sourced from the suppliers to obtain the feature representation of the supplier sample entity. By training this network, a green and low-carbon evaluation model is finally obtained. This model can process the target propagation chain, outputting the feature representation of the target indicator unit and the feature representation of the target supplier, thereby accurately determining the green and low-carbon evaluation indicator unit most suitable for the target supplier.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and more specifically, to a training method and system for a green and low-carbon evaluation model based on multiple types of suppliers. Background Technology

[0002] In green supply chain management, accurate green and low-carbon performance evaluation of various types of suppliers is crucial. Traditional evaluation methods often rely on isolated indicator scores or simple weighted summaries, failing to fully consider the complex and multi-layered relationship topology between supplier entities and green and low-carbon indicators. Different types of suppliers differ significantly in business models, data foundations, and environmental performance dimensions, and existing methods lack the ability to deeply model and integrate this heterogeneous relationship information. This results in evaluation results that are often inaccurate, making it difficult to distinguish the true degree of fit of suppliers in specific green dimensions, and failing to provide reliable support for sustainable procurement decisions. Summary of the Invention

[0003] The purpose of this invention is to provide a training method and system for a green and low-carbon evaluation model based on multiple types of suppliers.

[0004] In a first aspect, embodiments of the present invention provide a method for training a green and low-carbon evaluation model based on multiple types of suppliers, including:

[0005] Obtain a set of green and low-carbon evaluation indicators, and determine indicator sample units based on the set of green and low-carbon evaluation indicators. The set of green and low-carbon evaluation indicators includes multiple evaluation indicators related to green and low-carbon.

[0006] At least one first sample propagation chain and at least one second sample propagation chain are obtained. Each first sample propagation chain is used to characterize the topology between the supplier sample entity and the indicator sample unit when the supplier sample entity is the source node. Each second sample propagation chain is used to characterize the topology between the indicator sample unit and the supplier sample entity when the indicator sample unit is the source node.

[0007] Extract the state feature representation of each first sample propagation chain and the state feature representation of each second sample propagation chain;

[0008] At least one state feature representation of the second sample propagation chain is used as the received feature data of the first network branch in the initial evaluation network to obtain the indicator sample unit feature representation. At least one state feature representation of the first sample propagation chain, either the indicator sample unit feature representation or the first correction feature representation, is used as the received feature data of the second network branch in the initial evaluation network to obtain the supplier sample entity feature representation. The first correction feature representation is used to characterize the parameter association representation when the parameter mapping of the second network branch is associated with the generated feature representation result of the first network branch.

[0009] The initial evaluation network is trained based on the feature representation of the indicator sample unit and the feature representation of the supplier sample entity to obtain a green and low-carbon evaluation model. The green and low-carbon evaluation model is used to process the state feature representation of at least one first target propagation chain and the state feature representation of at least one second target propagation chain to obtain the target indicator unit feature representation and the target supplier feature representation. The target indicator unit feature representation and the target supplier feature representation are used to determine the target-fit indicator unit.

[0010] In one possible implementation, the first network branch includes a first feature fusion module and a first propagation link module; the state feature representation of at least one second sample propagation link is used as the received feature data of the first network branch in the initial evaluation network to obtain the indicator sample unit feature representation, including:

[0011] The state feature representation of each second sample propagation chain is sent to the first feature fusion module to obtain the state association feature representation of each second sample propagation chain. The state association feature representation of each second sample propagation chain is used to characterize the state association of the propagation association state in the corresponding second sample propagation chain.

[0012] The state association feature representation of at least one second sample propagation chain is used as the received feature data of the first propagation link by the module to obtain the routing confidence coefficient of each second sample propagation chain.

[0013] Based on the first propagation link, the module performs confidence aggregation on the state association feature representation of at least one second sample propagation link and the routing confidence coefficient to obtain the feature representation of the index sample unit.

[0014] In one possible implementation, before feeding the state feature representation of each second sample propagation chain into the first feature fusion module to obtain the state association feature representation of each second sample propagation chain, the method further includes:

[0015] Obtain a second corrected feature representation, which is used to characterize the parameter association representation when the parameter mapping of the first network branch is associated with the generated feature representation result of the second network branch;

[0016] The state feature representation of each second sample propagation chain is fed into the first feature fusion module to obtain the state association feature representation of each second sample propagation chain, including:

[0017] The second correction feature representation is combined with the state feature representation of each second sample propagation chain to obtain the correction combination feature representation of each second sample propagation chain.

[0018] The corrected combined feature representation of each second sample propagation chain is fed into the first feature fusion module to obtain the state association feature representation of each second sample propagation chain.

[0019] In one possible implementation, based on the first propagation link, the module performs confidence aggregation on the state association feature representation of at least one second sample propagation link and the routing confidence coefficient to obtain the indicator sample unit feature representation, including:

[0020] Based on the first propagation link, the module performs confidence aggregation on the state association feature representation of at least one second sample propagation link and the routing confidence coefficient to obtain the aggregated feature representation;

[0021] The aggregated feature representation and the second corrected feature representation are combined to obtain the feature representation of the index sample unit.

[0022] In one possible implementation, the second network branch includes a second feature fusion module and a second propagation link module; using one of the indicator sample unit feature representations or the first correction feature representation, and at least one state feature representation of the first sample propagation link, as received feature data of the second network branch in the initial evaluation network to obtain supplier sample entity feature representations, including:

[0023] The first correction feature representation is combined with the state feature representation of each first sample propagation chain to obtain the combined feature representation of each first sample propagation chain.

[0024] From the combined feature representation of the first propagation subchain, extract the feature representation of the target supplier sample entity and the feature representation of each associated sample entity that is propagated with the target supplier sample entity, wherein the first propagation subchain is any sample propagation chain in the at least one first sample propagation chain, and the target supplier sample entity is any supplier sample entity in the first propagation subchain;

[0025] The feature representations of the target supplier sample entities and the feature representations of each associated sample entity are respectively subjected to exponential correlation calculation to obtain the sample routing confidence coefficient of each associated sample entity in the first propagation subchain;

[0026] Based on the sample routing confidence coefficient of each associated sample entity, confidence aggregation is performed on the feature representation of the corresponding associated sample entity to obtain the state association feature representation of the first propagation sub-chain. The state association feature representation of each first sample propagation chain is used to characterize the state association of object interactions in the corresponding first sample propagation chain.

[0027] The state association feature representation of each first sample propagation chain is used as the received feature data of the second propagation link module to obtain the routing confidence coefficient of each first sample propagation chain.

[0028] Based on the second propagation link module, the state association feature representation of at least one first sample propagation link is aggregated with the route confidence coefficient to obtain the first propagation link feature representation;

[0029] The first propagation link is processed by combining the feature representation with the first correction feature representation to obtain the supplier sample entity feature representation.

[0030] In one possible implementation, the second network branch includes at least one evaluation dimension extraction module, a pattern routing fusion module, and a balance aggregation module; it uses one of the indicator sample unit feature representations or the first correction feature representation, and at least one state feature representation of the first sample propagation chain, as the received feature data of the second network branch in the initial evaluation network to obtain the supplier sample entity feature representation, including:

[0031] From the state feature representation of the second propagation subchain, extract the feature representation of the target supplier sample entity and the feature representation of each associated sample entity that is propagated with the target supplier sample entity. The second propagation subchain is any sample propagation chain in the at least one first sample propagation chain, and the target supplier sample entity is any supplier sample entity in the second propagation subchain.

[0032] The feature representations of the target supplier sample entities are combined with at least one pattern parameter feature representation to obtain at least one first combined feature representation.

[0033] At least one of the first combined feature representations is subjected to exponential correlation calculation with the feature representation of each of the associated sample entities to obtain the sample routing confidence coefficient of each of the associated sample entities in the second propagation subchain.

[0034] Based on the sample routing confidence coefficient of each associated sample entity, confidence aggregation is performed on the feature representation of the corresponding associated sample entity to obtain at least one pattern association feature representation of the second propagation subchain. Each evaluation dimension extraction module outputs one pattern association feature representation. Each pattern parameter feature representation is used to characterize the pattern adaptation association degree of the corresponding evaluation dimension extraction module for the indicator sample unit.

[0035] At least one pattern association feature representation, at least one pattern parameter feature representation, and the index sample unit feature representation of each first sample propagation chain are transmitted to the pattern routing fusion module to obtain the expert feature representation of each first sample propagation chain.

[0036] The expert feature representation of at least one of the first sample propagation chains is transmitted to the equalization aggregation module to obtain the supplier sample entity feature representation.

[0037] In one possible implementation, at least one pattern association feature representation, at least one pattern parameter feature representation, and the index sample unit feature representation of each first sample propagation chain are transmitted to the pattern routing fusion module to obtain an expert feature representation of each first sample propagation chain, including:

[0038] For each of the first sample propagation chains, at least one of the pattern parameter feature representations is subjected to exponential correlation calculation with the feature representation of the index sample unit to obtain the confidence coefficient of each of the pattern correlation feature representations in each of the first sample propagation chains.

[0039] For each of the first sample propagation chains, based on the confidence coefficient of at least one of the pattern association feature representations, confidence aggregation is performed on the corresponding at least one of the pattern association feature representations to obtain the expert feature representation of each of the first sample propagation chains.

[0040] In one possible implementation, the state feature representation of each second sample propagation chain is fed into the first feature fusion module to obtain the state association feature representation of each second sample propagation chain, including:

[0041] From the state feature representation of the third propagation subchain, extract the feature representation of the target index sample unit and the feature representation of each associated sample entity that is propagated with the target index sample unit. The third propagation subchain is any sample propagation chain in the at least one second sample propagation chain, and the target index sample unit is any index sample unit in the third propagation subchain.

[0042] The feature representations of the target index sample units and the feature representations of each associated sample entity are respectively subjected to exponential correlation calculation to obtain the sample routing confidence coefficient of each associated sample entity in the third propagation subchain.

[0043] Based on the sample routing confidence coefficient of each associated sample entity, confidence aggregation is performed on the feature representation of the corresponding associated sample entity to obtain the state association feature representation of the third propagation subchain.

[0044] In one possible implementation, the initial evaluation network is trained based on the feature representations of the indicator sample units and the feature representations of the supplier sample entities to obtain a green and low-carbon evaluation model, including:

[0045] Calculate the deviation measure between the feature representation of the indicator sample unit and the feature representation of the supplier sample entity to obtain the deviation measure value;

[0046] The network connection weights of the initial evaluation network are adjusted based on the deviation metric to obtain a green and low-carbon evaluation model.

[0047] Secondly, embodiments of the present invention provide a training system for a green and low-carbon evaluation model based on multiple types of suppliers. The training system includes a processor and a memory, the memory and the processor being connected. The memory is used to store programs, instructions or code, and the processor is used to execute the programs, instructions or code in the memory to implement the green and low-carbon evaluation model training method based on multiple types of suppliers described in the second aspect.

[0048] Compared to existing technologies, the beneficial effects of this invention include: Using the green and low-carbon evaluation model training method and system disclosed in this invention based on multiple types of suppliers, a set of green and low-carbon evaluation indicators is obtained and indicator sample units are determined. Subsequently, two types of sample propagation chains are obtained, with supplier sample entities and indicator sample units as source nodes respectively, to characterize their topological structure. After extracting the state feature representations of each propagation chain, they are processed through two branches of the initial evaluation network: the first branch obtains the indicator sample unit feature representation based on the features sourced from the indicators; the second branch combines the output features of the first branch with the features sourced from the suppliers to obtain the supplier sample entity feature representation. By training this network, a green and low-carbon evaluation model is finally obtained. This model can process the target propagation chain, outputting the target indicator unit feature representation and the target supplier feature representation, thereby accurately determining the green and low-carbon evaluation indicator unit most suitable for the target supplier. Attached Figure Description

[0049] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly described below. It should be understood that the following drawings only show some embodiments of the present invention and should not be considered as limiting the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0050] Figure 1 A flowchart illustrating the steps of a green and low-carbon evaluation model training method based on multiple types of suppliers provided in an embodiment of the present invention;

[0051] Figure 2 A schematic block diagram of the structure of a computer device provided in an embodiment of the present invention. Detailed Implementation

[0052] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.

[0053] The specific embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0054] In order to solve the technical problems mentioned in the background art Figure 1 This is a flowchart illustrating the training method for a green and low-carbon evaluation model based on multiple types of suppliers provided in this embodiment. The following is a detailed description of the training method for the green and low-carbon evaluation model based on multiple types of suppliers.

[0055] Step S110: Obtain a set of green and low-carbon evaluation indicators, and determine indicator sample units based on the set of green and low-carbon evaluation indicators. The set of green and low-carbon evaluation indicators includes multiple evaluation indicators related to green and low-carbon.

[0056] The server first retrieves raw data constituting the green and low-carbon evaluation indicator set from a pre-set corporate social responsibility database, an environmental management system standard library (such as the ISO 14000 series), and industry-specific green supply chain management specifications. This indicator set is a structured collection containing multiple evaluation indicators closely related to green and low-carbon practices. For example, in an evaluation system for the automotive parts manufacturing industry, this indicator set might include dozens of specific indicators such as "carbon emission intensity per unit product," "proportion of renewable energy used in the production process," "level of control and substitution of hazardous substances," "compliance rate of wastewater and exhaust gas emissions," "product recyclability rate," "implementation of green logistics," and "environmental management system certification level."

[0057] After obtaining the original textual or numerical definitions of the indicators, the server needs to preprocess the indicator set to form indicator sample units that can be used for model calculation. This process includes: vectorizing and embedding textual indicators, for example, using a pre-trained language model to transform "environmental management system certification level" into a fixed-dimensional semantic vector; normalizing or standardizing numerical indicators to eliminate the influence of dimensions, for example, mapping the original value of "carbon emission intensity per unit product" to the [0, 1] interval; for indicators with hierarchical structures (such as multiple secondary indicators under a primary indicator), the server will perform feature fusion or aggregation to generate a single feature vector representing the composite indicator. Finally, each processed indicator, regardless of its original form, is represented as a high-dimensional feature vector, which is called an "indicator sample unit". Therefore, the indicator sample unit is the digital and vectorized representation of green and low-carbon evaluation indicators in the feature space. The server stores all these indicator sample units in memory or cache, forming a baseline feature library for subsequent processing.

[0058] Step S120: Obtain at least one first sample propagation chain and at least one second sample propagation chain. Each first sample propagation chain is used to characterize the topology between the supplier sample entity and the indicator sample unit when the supplier sample entity is the source node. Each second sample propagation chain is used to characterize the topology between the indicator sample unit and the supplier sample entity when the indicator sample unit is the source node.

[0059] The server then extracts data from the enterprise's Supplier Relationship Management (SRM) system, supply chain transaction record database, and environmental audit report to construct a topological network reflecting the relationship between suppliers and evaluation indicators. In this network, nodes are divided into two categories: supplier sample entities (representing specific supplier companies) and indicator sample units (i.e., the indicator feature vectors obtained in the previous step). Edges represent the relationship between the two, which can be direct (such as a supplier's specific indicator score in an audit project) or indirect (such as relationships derived from jointly participated supply chain projects, shared environmental technologies, or belonging to the same industrial park).

[0060] Based on this topology, the server performs traversal in two directions to generate two types of sample propagation chains. The first sample propagation chain starts with a specific supplier sample entity as the source node (starting point) and explores its association paths in the network. The server uses a restricted random walk or a meta-path-based traversal algorithm. For example, starting from supplier entity "Company A", the path might be "Company A" — (through the "supply contract" relationship) — "Product X" — (through the "environmental attributes of Product X" relationship) — "recyclable rate indicator unit". This path constitutes a first sample propagation chain, which depicts the topology from supplier A, through intermediate entities (such as specific products), and finally associated with a certain green and low-carbon indicator unit. The server generates multiple such propagation chains for each supplier sample entity in the training set to cover its different association patterns with different indicator units.

[0061] The second sample propagation chain involves reverse exploration starting from a specific indicator sample unit as the source node. For example, starting with the "carbon emission intensity indicator unit per unit product," the path might be: "carbon emission intensity indicator unit" — (through the "applied to assessment" relationship) — "industry energy efficiency benchmark report" — (through the "supplier mentioned in the report" relationship) — "Company B." This path constitutes a second sample propagation chain, depicting the topology from a certain green and low-carbon indicator, through intermediate information carriers or assessment activities, and ultimately linked to a specific supplier entity. Similarly, the server will generate multiple such propagation chains for important indicator sample units.

[0062] Each propagation chain (whether it is the first type or the second type) is serialized and stored by the server, which includes the feature representation of all nodes on the path (the supplier node has its own features such as enterprise size, industry classification, etc., the indicator node is its indicator sample unit feature, and the intermediate node also has corresponding features) as well as the type and weight information of the edges between nodes.

[0063] Step S130: Extract the state feature representation of each first sample propagation chain and the state feature representation of each second sample propagation chain;

[0064] For each generated sample propagation chain, the server needs to transform it from a discrete node-edge sequence into a continuous feature representation that can reflect the overall state information of the chain, i.e., a state feature representation.

[0065] The server employs a graph neural network (GNN)-based encoder or a dedicated sequence encoder (such as the Transformer encoder) to perform this extraction process. Specifically, for a propagation chain, the server treats it as a heterogeneous graph sequence. The encoder processes each node in the chain sequentially, along with its incoming and outgoing edges. For example, for a first-sample propagation chain "Supplier S—Product P—Indicator I", the encoder first reads the initial features of supplier node S (such as industry code, registration location feature vector, etc.) and the type features of the "Production" edge; then, this information is passed to product node P, where it is fused with P's initial features (such as product type, material composition vector, etc.) and the type features of the "Having" edge to update the hidden state of node P; finally, the updated state of node P, along with the features of the "Association" edge, is passed to indicator node I, where it is fused with I's indicator sample unit features to generate the final hidden state of indicator node I in the context of this chain. This final hidden state, or the vector obtained by pooling (such as average pooling) the hidden states of all nodes in the chain at the final layer, is defined by the server as the "state feature representation" of this first-sample propagation chain. It not only contains the indicator information at the end of the chain, but also encodes the path semantics from a specific supplier to that indicator.

[0066] Similarly, for the second sample propagation chain, such as "Indicator I - Report R - Supplier S", the encoder starts from indicator node I, passes through the intermediate report node R, and finally converges to supplier node S, generating a state feature representation that ends with the context state of supplier S from the perspective of a specific indicator. The server independently performs the above encoding operation on each first sample propagation chain and each second sample propagation chain, resulting in two sets: the set of state feature representations for the first sample propagation chain and the set of state feature representations for the second sample propagation chain.

[0067] Step S140: At least one state feature representation of the second sample propagation chain is used as the received feature data of the first network branch in the initial evaluation network to obtain the indicator sample unit feature representation; and one of the indicator sample unit feature representation or the first correction feature representation, and at least one state feature representation of the first sample propagation chain are used as the received feature data of the second network branch in the initial evaluation network to obtain the supplier sample entity feature representation. The first correction feature representation is used to characterize the parameter association representation when the parameter mapping of the second network branch is associated with the generated feature representation result of the first network branch.

[0068] The server constructs an initial evaluation network with a two-branch structure. The two branches of this network share some underlying parameters but have different structures to handle information from different perspectives.

[0069] The first network branch (indicator perspective branch): The goal of this branch is to aggregate multi-faceted information about a certain indicator sample unit derived from different supplier association paths, thereby generating a richer and more robust "indicator sample unit feature representation". The server collects the state feature representations of all second sample propagation chains obtained in the previous step, with the same target indicator sample unit (e.g., "carbon emission intensity indicator unit") as the source node, and uses them as the input of this branch.

[0070] In one specific implementation, the first network branch includes a first feature fusion module and a first propagation chain module. The server first inputs the state feature representation of each second sample propagation chain into the first feature fusion module. This module is typically a multilayer perceptron (MLP), which performs nonlinear transformation and refinement on the state representation of each chain to extract the "state association features" reflected by that chain, that is, focuses more on the core information of the "association pattern between indicators and specific supplier groups" expressed by that chain.

[0071] Next, the server inputs all refined state-association feature representations into the first propagation chain module. At the core of this module is an attention mechanism or routing network. It learns to calculate a "routing confidence coefficient" for the state-association features of each input second-sample propagation chain. This coefficient reflects the importance of the association pattern represented by the propagation chain in defining the overall features of the target metric sample unit during the current training phase. For example, a propagation chain that associates multiple leading suppliers through an "authoritative industry white paper" may have a higher confidence than a propagation chain that associates suppliers through a "single supplier self-declaration."

[0072] Finally, based on the calculated routing confidence coefficient, the server performs a weighted summation (i.e., confidence aggregation) of the state association features of all second sample propagation chains. This aggregated feature vector, which integrates information from different association paths and different confidence sources, is output by the server as the final, enhanced "indicator sample unit feature representation".

[0073] The second network branch (supplier perspective branch): The goal of this branch is to aggregate information from multiple paths originating from the target supplier and associated with different metrics, while also considering the global features of the metrics passed from the first branch, thereby generating a "supplier sample entity feature representation" for the supplier. The server collects the state feature representations of all first sample propagation chains with the same target supplier sample entity as the source node obtained in the previous step, and uses them as one of the main inputs to this branch.

[0074] Furthermore, this branch also needs to receive contextual information from the first network branch. This is achieved through a learnable "first corrected feature representation." This first corrected feature representation is a parameter vector that is optimized simultaneously with the network weights. Essentially, it is an association representation of how the parameter space of the second network branch is aligned and mapped with the generated result of the first network branch (i.e., the feature representation of the index sample unit). In the early stages of training or in some variant designs, the feature representation of the index sample unit generated by the first branch can also be directly used as part of the input.

[0075] In one specific implementation, the second network branch may include a second feature fusion module and a second propagation link module, whose processing flow is symmetrical to that of the first branch but with different inputs. The server first performs feature combination operations such as concatenation or addition between the first corrected feature representation and the state feature representation of each first sample propagation chain, injecting global indicator context information into each specific supplier-indicator path feature. Then, each combined feature undergoes processing similar to that of the first branch: refinement through the second feature fusion module, calculation of the confidence coefficient of each path through the second propagation link module, and finally weighted aggregation. The aggregation result is then combined with the first corrected feature representation to output the final "supplier sample entity feature representation." This process ensures that the supplier's feature representation is not only based on its own associated indicator paths but also implicitly considers the aggregated features of these indicators from a global perspective.

[0076] In another, more complex implementation, the second network branch can be designed to include multiple parallel "evaluation dimension extraction modules," a "pattern routing fusion module," and a "balanced aggregation module." Each evaluation dimension extraction module focuses on extracting features related to a certain evaluation pattern (such as "environmental compliance," "resource efficiency," "product greenness," etc.) from the supplier's propagation chain. It guides the direction of feature extraction by introducing a learnable "pattern parameter feature representation." Specifically, for a first sample propagation chain, the server extracts features from the target supplier node and associated intermediate nodes in the chain, combines the supplier features with a certain pattern parameter feature, calculates the correlation degree (sample routing confidence coefficient) with the features of other nodes in the chain, and then aggregates the information on the chain according to this coefficient to obtain the "pattern association feature representation" of the chain under that pattern. All first sample propagation chains of a supplier will pass through all evaluation dimension extraction modules, generating multiple sets of pattern association features.

[0077] Then, the server sends all the pattern association features, corresponding pattern parameter features, and indicator sample unit feature representations from the first branch of the first sample propagation chain into the pattern routing fusion module. This module calculates the fit confidence of each pattern for the current supplier (based on all its propagation chains) (calculated by the correlation between pattern parameter features and global indicator features), and then fuses the features of different patterns according to this confidence to generate a unified "expert feature representation" for each first sample propagation chain. This representation integrates multi-dimensional information.

[0078] Finally, the server feeds the expert feature representations of all first-sample propagation chains into the equilibrium aggregation module (e.g., using a self-attention mechanism), comprehensively considering the importance of different propagation chains, and performs the final aggregation to generate the "supplier sample entity feature representation" for that supplier. This design enables the model to explicitly model and integrate the differentiated performance of multiple types of suppliers across different green and low-carbon dimensions.

[0079] Step S150: The initial evaluation network is trained based on the feature representation of the indicator sample unit and the feature representation of the supplier sample entity to obtain a green and low-carbon evaluation model. The green and low-carbon evaluation model is used to process the state feature representation of at least one first target propagation chain and the state feature representation of at least one second target propagation chain to obtain the target indicator unit feature representation and the target supplier feature representation. The target indicator unit feature representation and the target supplier feature representation are used to determine the target-fit indicator unit.

[0080] In this embodiment of the invention, for example, the server drives the training of the entire initial evaluation network by defining a loss function. The core idea is that for a training sample (usually a known, well-performing "supplier-metric" pair), the feature representation of the supplier sample entity generated by that supplier and the feature representation of the metric sample unit generated by that metric should be close in the feature space; while for mismatched "supplier-metric" pairs, their feature representations should be far apart.

[0081] Therefore, the server calculates the similarity (e.g., cosine similarity) between the feature representations of all positive sample pairs (matching pairs) and the similarity between the feature representations of all negative sample pairs (unmatching pairs constructed through random sampling) within a batch. Then, it employs either a contrastive learning loss (e.g., InfoNCELoss) or a margin-based ranking loss (e.g., TripletLoss) to maximize the similarity of positive sample pairs and minimize the similarity of negative sample pairs, or to set the similarity of positive sample pairs above a predetermined margin value for negative sample pairs. This loss value is called the "deviation metric".

[0082] The server uses the backpropagation algorithm to update all trainable parameters in the initial evaluation network based on the calculated deviation metric (loss value), including the weights of each module in both branches, corrected feature representations, and mode parameter features. Through numerous training iterations, the network gradually learns how to extract effective information from the bidirectional propagation chain and generate embeddings that highly align the matching supplier and indicator feature representations.

[0083] After training, the server saves the final network parameters, resulting in the "Green and Low-Carbon Evaluation Model." This model possesses general processing capabilities: For a new target supplier and a set of target indicators, the server can first construct the supplier's first target propagation chain (associated with each indicator) and each indicator's second target propagation chain (associated with each supplier), extracting their state feature representations. Then, the state feature representation of the second target propagation chain is input into the model's first network branch to obtain an enhanced "target indicator unit feature representation." The state feature representation of the first target propagation chain and the model's internal first correction feature representation (or combined with the first branch output) are input into the model's second network branch to obtain the "target supplier feature representation." Finally, the similarity between the target supplier feature representation and each target indicator unit feature representation is calculated. The indicator unit with the highest similarity is determined by the server to be the "target-fit indicator unit" with the target supplier in the green and low-carbon dimension, and can be used for supplier capability assessment, green procurement decisions, or personalized improvement suggestion generation.

[0084] In summary, the server-executed method of this invention, by systematically constructing and utilizing bidirectional propagation chain topology information and designing a collaborative dual-branch deep network for feature learning and fusion, can overcome the shortcomings of isolated indicators and neglect of supplier type differences in traditional evaluation methods, and achieve deep, accurate and interpretable evaluation model training for the green and low-carbon performance of multiple types of suppliers.

[0085] After the server completes the extraction of state features from the second sample propagation chain, it begins to execute the processing flow of the first network branch. Its core objective is to fuse multi-angle information about the same target indicator sample unit gathered from different supplier association paths to generate an enhanced, global indicator feature representation.

[0086] Step S141: Send the state feature representation of each second sample propagation chain into the first feature fusion module to obtain the state association feature representation of each second sample propagation chain.

[0087] The server first focuses on a specific indicator sample unit, such as "carbon emission intensity per unit of product". Assume the server has previously constructed three second sample propagation chains with this indicator as the source node:

[0088] 1. Chain A: "Carbon Emission Intensity Indicator" - "Industry Annual Sustainable Development White Paper" - "Supplier A";

[0089] 2. Chain B: "Carbon Emission Intensity Index" — "Third-Party Environmental Audit Report R" — "Supplier B";

[0090] 3. Chain C: "Carbon Emission Intensity Indicator" - "Supply Chain Collaborative Emission Reduction Project Database" - "Supplier C".

[0091] The server retrieves the state feature representations of these three chains from memory, denoted as h_A, h_B, and h_C. Each h is a high-dimensional vector that encodes the complete path semantic information from the indicator node, through specific intermediate nodes (such as white papers and audit reports), to the final destination of a specific supplier.

[0092] Subsequently, the server inputs h_A, h_B, and h_C independently to the first feature fusion module. This module is typically a parameter-shared multilayer perceptron. This network performs a series of nonlinear transformations on the input vector of each chain. Its function is to refine and purify the features, extracting more core and focused "state-related features" from the relatively raw and mixed path state representations. Specifically, it learns to extract the essence of the specific "association pattern" represented by each chain. For example, for chain A, the module might strengthen the "industry benchmarking" pattern feature; for chain B, it might strengthen the "third-party authoritative verification" pattern feature; and for chain C, it might strengthen the "dynamic project participation" pattern feature. After processing by this module, the server obtains three new feature vectors h_A', h_B', and h_C', which are the "state-related feature representations" of each second sample propagation chain, respectively condensing the core state information representing the association between the indicator and different supplier groups through different channels.

[0093] Step S142: Use the state association feature representation of at least one second sample propagation chain as the received feature data of the first propagation link module to obtain the routing confidence coefficient of each second sample propagation chain.

[0094] The server inputs h_A', h_B', and h_C' obtained in the previous step into the first propagation link module. The core of this module is an attention mechanism. The server uses this module to calculate a routing weight vector. Specifically, the module first concatenates or stacks h_A', h_B', and h_C', then uses a feedforward network to calculate the unnormalized importance score of each feature vector. Next, the server normalizes these scores using the Softmax function, obtaining three values ​​between 0 and 1 that sum to 1: α_A, α_B, and α_C. These three values ​​are the "routing confidence coefficients" corresponding to each second sample propagation link.

[0095] The calculation of the coefficients is data-driven. For example, during training, the server might discover that supplier data linked through a "third-party environmental audit report" (chain B) is crucial for accurately characterizing the "carbon emission intensity" indicator due to its quality and stability; therefore, the model will learn to assign a higher value to α_B. Meanwhile, "industry white papers" (chain A) may have broad coverage but coarser data granularity, resulting in a moderate confidence level α_A. Data from the "supply chain project database" (chain C) may be highly real-time but relatively less authoritative, leading to a lower confidence level α_C. These coefficients dynamically reflect the relative reliability and importance of different information sources within the current task context.

[0096] Step S143: Based on the first propagation link, the module performs confidence aggregation on the state association feature representation and routing confidence coefficient of at least one second sample propagation link to obtain the feature representation of the index sample unit.

[0097] Finally, the server performs confidence aggregation. This is a crucial step in fusing multi-source information into a single, enhanced representation. The server uses a weighted summation method for aggregation. The specific calculation formula is: H_index = α_Ah_A' + α_Bh_B' + α_Ch_C'.

[0098] The server performs this vector operation, using the previously calculated route confidence coefficients α_A, α_B, and α_C as weights, multiplying them by the corresponding state association feature representations h_A', h_B', and h_C', respectively. Then, the three weighted vectors are summed to obtain the final aggregated feature vector H_index. This H_index is the "index sample unit feature representation" sought by the server.

[0099] Through this operation, the H_ index generated by the server is no longer an isolated, static index definition vector, but a rich representation that integrates multi-dimensional, dynamically related information. It includes not only the literal meaning of the "carbon emission intensity per unit product" index, but also incorporates multiple contextual information such as "how this index is cited in authoritative industry reports," "how this index is verified in rigorous audits," and "how this index is applied in actual supply chain projects." Furthermore, the routing confidence coefficient ensures that more reliable and relevant information occupies a higher weight in the final representation. This enhanced feature representation will be output for subsequent collaborative computation of the second network branch and target alignment during model training.

[0100] Before performing feature fusion of the first network branch, the server needs to perform a key operation: acquire and apply the second corrected feature representation to establish a parameter mapping association with the second network branch, ensuring the collaboration and alignment of the two branches in the feature learning process.

[0101] Step S140a: Obtain the second corrected feature representation.

[0102] When the server initializes and builds the initial evaluation network, it randomly generates and stores a trainable high-dimensional parameter vector, which is called the "second calibration feature representation," denoted as C_second. This vector does not come from any input data but serves as an internally optimizable parameter of the model, participating in training along with the network weights. Its physical meaning is to learn and represent the association rules that characterize how the parameter space of the first network branch should be mapped and aligned with the expected or already generated feature representations of the second network branch (the supplier's perspective branch). During training, C_second is continuously updated through backpropagation, gradually encoding the "translation" or "calibration" knowledge required for effective information exchange between the two branches.

[0103] Step S141a: Perform feature combination processing on the second corrected feature representation and the state feature representation of each second sample propagation chain to obtain the corrected combined feature representation.

[0104] The server focuses on the target indicator sample unit, such as "product recyclability". Assume there are three related second sample propagation chains, whose state characteristics are represented as h_D (related path involves industry standards), h_E (related path involves customer factory inspection reports), and h_F (related path involves waste material recycling transaction records).

[0105] Before sending h_D, h_E, and h_F into the first feature fusion module, the server first performs feature combination processing. Specifically, the server performs vector addition operations on the second corrected feature representation C_second and each state feature representation respectively. That is, the server calculates: h_D_combined = h_D + C_second; h_E_combined = h_E + C_second; h_F_combined = h_F + C_second.

[0106] This addition operation is not a simple concatenation, but a correction offset in the feature space. Its function is to inject calibration information C_second, representing "how to understand supplier characteristics," into the features of each associated path originating from the metric. This ensures that the feature representation of each path carries contextual cues from another perspective (the supplier's perspective) from the very beginning of subsequent processing, guiding the feature fusion module not only to focus on "how the metrics are associated," but also subconsciously considering "what characteristic patterns the supplier to which this association points might possess."

[0107] Step S141b: The corrected combined feature representation of each second sample propagation chain is sent to the first feature fusion module to obtain the state association feature representation of each second sample propagation chain.

[0108] Next, the server inputs the offset-corrected feature vectors h_D_combined, h_E_combined, and h_F_combined into the first feature fusion module. This module (e.g., a multilayer perceptron) now receives inputs that are combined features that fuse the original path information and cross-branch calibration signals.

[0109] The module performs a non-linear transformation on these inputs. Because the inputs include C_second, the module's internal computation process is adjusted by the calibration signal when extracting the core "state-related features" of the path. For example, for h_D_combined, when refining the "industry standard association" pattern, the module may simultaneously strengthen dimensions related to the "supplier compliance features" encoded by C_second. Finally, the server outputs new feature vectors h_D', h_E', and h_F' from this module. These are the more refined "state-related feature representations" after cross-branch correction.

[0110] These not only characterize the associations in the original propagation chain (e.g., "the metric is associated with the supplier through industry standards"), but also implicitly relate what this association pattern might mean for the other end (the supplier entity). This lays a crucial alignment foundation for subsequent steps, where the metric feature representation H_metric generated by the first network branch can be effectively compared and matched with the supplier feature representation H_supplier generated by the second network branch in a common space. The entire processing flow is executed automatically by the server, and C_second is dynamically adjusted to its optimal state during training by optimizing the loss function.

[0111] In this embodiment of the invention, based on the first propagation link, the module performs confidence aggregation on the state association feature representation of at least one second sample propagation link and the routing confidence coefficient to obtain the indicator sample unit feature representation, which can be implemented through the following example.

[0112] Based on the first propagation link, the module performs confidence aggregation on the state association feature representation of at least one second sample propagation link and the routing confidence coefficient to obtain the aggregated feature representation;

[0113] The aggregated feature representation and the second corrected feature representation are combined to obtain the feature representation of the index sample unit.

[0114] In this embodiment of the invention, for example, after calculating the route confidence coefficients, the server performs the final aggregation and correction steps. First, based on the calculated route confidence coefficients α_D, α_E, and α_F, the server performs a weighted summation operation on the corresponding state-related feature representations h_D', h_E', and h_F' to obtain a preliminary aggregated feature representation H_agg = α_D*h_D' + α_E*h_E' + α_F*h_F'. Then, the server performs a second feature combination process, adding H_agg to the previously obtained second correction feature representation C_second, resulting in H_index = H_agg + C_second. This H_index is the final indicator sample unit feature representation. Through this operation, the server integrates the cross-branch calibration information C_second into the aggregated indicator feature, ensuring that the representation not only incorporates multi-path information but also maintains alignment with the representation space of the other branch, laying the foundation for subsequent joint training and accurate matching.

[0115] After the server completes the processing of the indicator features in the first network branch, it synchronously executes the processing flow of the second network branch. Its core objective is to aggregate multiple path information from the target supplier to different indicators, and to integrate the calibration information from the first branch to generate an enhanced feature representation of the supplier.

[0116] Step S151: Perform feature combination processing on the first correction feature representation and the state feature representation of each first sample propagation chain to obtain the combined feature representation of each first sample propagation chain.

[0117] The server first retrieves the trainable internal parameter vector "first corrected feature representation," denoted as C_first, from the parameter storage. Its function is symmetric to C_second, representing how the parameter space of the second network branch should be mapped and aligned with the generated results of the first network branch. The server selects the target supplier sample entity, such as "supplier A." Assume three first sample propagation chains have been constructed with A as the source node:

[0118] 1. Chain X: "Supplier A" — "Product A" — "Unit Product Carbon Emission Intensity Index"

[0119] 2. Chain Y: "Supplier A" — "Process B" — "Production Wastewater Recycling Rate Index"

[0120] 3. Chain Z: "Supplier A" — "Management System Certification C" — "Hazardous Substance Control Indicators"

[0121] The server extracts the state feature representations s_X, s_Y, and s_Z from these three chains. Then, the server performs a vector addition operation with each s chain, C_first, to obtain the corrected combined feature representations: s_X_combined = s_X + C_first; s_Y_combined = s_Y + C_first; s_Z_combined = s_Z + C_first. This operation injects calibration information from the metric perspective branch into the initial features of each supplier-related path, introducing a global metric context for subsequent processing.

[0122] Step S152: Send the combined feature representation of each first sample propagation chain into the second feature fusion module to obtain the state association feature representation of each first sample propagation chain.

[0123] Next, the server inputs s_X_combined, s_Y_combined, and s_Z_combined into the second feature fusion module (a parameter-shared multilayer perceptron). This module performs nonlinear transformations and refinements on the combined features of each chain, extracting core features that are more focused on the "object interaction state association" reflected by the path. For example, for chain X, the module may extract the interaction pattern feature of "associating with carbon emission performance through specific products"; for chain Y, it may extract the interaction pattern feature of "associating with water resource management through specific production processes"; and for chain Z, it may extract the interaction pattern feature of "associating with chemical safety through management system certification". The server outputs the refined feature vectors s_X', s_Y', and s_Z', which are the "state association feature representations" of each first sample propagation chain.

[0124] Step S153: Use the state association feature representation of each first sample propagation chain as the received feature data of the second propagation link module to obtain the routing confidence coefficient of each first sample propagation chain.

[0125] The server simultaneously inputs s_X', s_Y', and s_Z' into the second propagation link module (an attention mechanism). This module calculates the importance weight of each path in defining the overall green characteristics of the current supplier "Supplier A". The server outputs three route confidence coefficients β_X, β_Y, and β_Z through the feedforward network and Softmax normalization of this module. These coefficients are dynamically learned; for example, if the training data indicates that "carbon emission intensity" and "hazardous substance control" are core aspects of the current industry evaluation, and "Supplier A" has high data quality on these two chains, the model may assign higher values ​​to β_X and β_Z.

[0126] Step S154: Based on the second propagation link module, perform confidence aggregation on the state association feature representation and routing confidence coefficient of at least one first sample propagation link to obtain the first propagation link feature representation.

[0127] The server performs weighted aggregation: H_route = β_X*s_X' + β_Y*s_Y' + β_Z*s_Z'. The resulting H_route is a "first propagation link represented by features" that integrates the path information associated with different indicators and considers the relative importance of each path.

[0128] Step S155: Combine the first propagation link with the first corrected feature representation to obtain the supplier sample entity feature representation.

[0129] Finally, the server performs a second feature combination, adding the route aggregation feature H_route to the first corrected feature representation C_first: H_supplier = H_route + C_first. This step integrates the global indicator-perspective calibration information into the supplier's aggregation feature. The output H_supplier is the "supplier sample entity feature representation," which includes detailed information about the supplier's association with various indicators through multiple specific paths. It also implicitly aligns the supplier's position with its associated indicators in the global feature space, laying the foundation for subsequent accurate similarity calculation with H_indicators.

[0130] In this embodiment of the invention, the combined feature representation of each of the first sample propagation chains is sent to the second feature fusion module to obtain the state association feature representation of each of the first sample propagation chains. This can be implemented through the following example.

[0131] Step S152a: Extract the feature representation of the target supplier sample entity and the feature representation of each associated sample entity that is propagated with the target supplier sample entity from the combined feature representation of the first propagation subchain.

[0132] The server focuses on a specific first-sample propagation chain, such as chain X: "Supplier A" — "Product A" — "Unit Product Carbon Emission Intensity Index". The combined feature representation s_X_combined of this chain is a serialized or graph-structured high-dimensional representation that integrates information from all nodes in the chain. The server first parses this representation and extracts the feature sub-vectors representing specific nodes.

[0133] Specifically, the server locates and separates from s_X_combined:

[0134] 1. Feature representation of the target supplier sample entity: That is, the feature vector of the source node "Supplier A" of the chain, denoted as v_supplier_A. This vector encodes the inherent attributes of Supplier A.

[0135] 2. Feature representation of the associated sample entities: That is, the feature vectors of the subsequent nodes on the chain. In this chain, it includes:

[0136] v_product_A: The feature vector representing the intermediate entity "Product A".

[0137] v_indicator_carbon_emission: The feature vector representing the "carbon emission intensity indicator per unit product" at the end of the chain (this is an instantiation of the feature representation of its indicator sample unit).

[0138] Step S152b: Respectively perform exponential association calculation processing on the feature representation of the target supplier sample entity and the feature representation of each associated sample entity to obtain the sample routing confidence coefficient of each associated sample entity in the first propagation sub-chain.

[0139] Next, the server calculates the importance weight of each associated entity on the chain for understanding "this specific path starting from Supplier A". The server performs an attention calculation process. For each associated sample entity (taking v_product_A as an example), the server calculates its association score with the target supplier entity v_supplier_A.

[0140] First, the server maps v_supplier_A to a "query" vector q through a learnable linear transformation matrix W_Q, and maps v_product_A to a "key" vector k_product through another linear transformation matrix W_K. Then, the server calculates the dot product (or scaled dot product) of q and k_product to obtain a scalar score score_product. This score preliminarily characterizes the direct association strength between "Product A" and "Supplier A" in the context of this path.

[0141] The server performs the same operation on another associated entity v_indicator_carbon_emission to obtain k_indicator and score_indicator.

[0142] The server then normalizes the two raw scores, `score_product` and `score_indicator`, using the Softmax function. The calculation is as follows: `γ_product = exp(score_product) / (exp(score_product) + exp(score_indicator))`, and `γ_indicator` is calculated similarly. This gives the server two values, `γ_product` and `γ_indicator`, between 0 and 1, that sum to 1. These are the "sample route confidence coefficients" for "Product A" and "carbon emission intensity index" in this chain X. A high `γ_product` means that in this path, the intermediate node "Product A" is a key hub for understanding the relationship between supplier A and carbon emissions; a high `γ_indicator` may mean that the information in this path is highly concentrated on the final indicator node itself.

[0143] Step S152c: Based on the sample routing confidence coefficient of each associated sample entity, perform confidence aggregation on the feature representation of the corresponding associated sample entity to obtain the state association feature representation of the first propagation subchain.

[0144] Finally, the server performs weighted aggregation of the information within the chain. The server multiplies the feature representation of each associated sample entity (usually using the "value" vectors v_product' and v_indicator' mapped through the W_V matrix) with its corresponding sample route confidence coefficient, and then sums them.

[0145] That is, the server calculates: s_X'=γ_product*v_product'+γ_indicator*v_indicator'.

[0146] This aggregated vector s_X' is the "state association feature representation" of chain X. It is no longer a simple list or average of all node features, but a fused representation that has been recalibrated based on the association importance between the node and the source supplier in this specific path. It more succinctly characterizes "which intermediate and final indicator information are the crucial interactive states in this specific path from supplier A to carbon emission indicators".

[0147] The server executes the complete process from S152a to S152c independently and in parallel for each first sample propagation chain (chain Y, chain Z, etc.), thereby generating a refined state association feature representation s_Y', s_Z', etc. for each chain, so that the subsequent second propagation chain can be aggregated across chains by the module.

[0148] In this embodiment of the invention, the method further includes:

[0149] Calculate the first squared deviation metric between the supplier sample entity feature representation and the second corrected feature representation;

[0150] The first correction feature representation and the second correction feature representation are optimized based on the first squared deviation metric.

[0151] In this embodiment of the invention, exemplarily, during model training, the server performs the following steps to optimize the synergy between the two branches. The server calculates the squared Euclidean distance between the supplier sample entity feature representation H_supplier generated in the current iteration and the second calibration feature representation C_second, obtaining a first squared bias metric L_align. This value measures the difference between the supplier features and the calibration signal from the index branch. Subsequently, during the backpropagation phase, the server uses L_align as part of the auxiliary loss term, calculates its gradient, and updates the internal parameters of the first calibration feature representation C_first and the second calibration feature representation C_second, respectively. Through iterative tuning, the server drives the evolution of the two calibration feature representations, enabling the supplier features generated by the second branch to better align with the parameter space of the first branch, thereby improving the consistency of information transmission between the two network branches and the overall collaborative learning effect of the model.

[0152] In this embodiment of the invention, the method further includes:

[0153] Calculate the second squared deviation metric between the feature representation of the index sample unit and the first corrected feature representation;

[0154] The second correction feature representation and the first correction feature representation are optimized based on the second squared deviation metric.

[0155] In this embodiment of the invention, exemplarily, during model training, the server performs the following symmetric optimization steps to enhance inter-branch collaboration. The server calculates the squared Euclidean distance between the feature representation H_index of the index sample unit generated in the current iteration and the first corrected feature representation C_first, obtaining a second squared bias metric L_align2. This value quantifies the difference between the index feature and the calibration signal from the supplier branch. Subsequently, during the backpropagation phase, the server uses L_align2 as another auxiliary loss term, calculates its gradient, and simultaneously updates the parameters of the second corrected feature representation C_second and the first corrected feature representation C_first. Through iterative optimization, the server promotes the co-evolution of the two corrected feature representations, ensuring that the index feature generated by the first branch can be effectively aligned with the parameter space of the second branch, thereby bidirectionally improving the consistency and matching accuracy of the model feature representation.

[0156] In this embodiment of the invention, the second network branch includes at least one evaluation dimension extraction module, a pattern routing fusion module, and a balanced aggregation module; the second network branch in the initial evaluation network receives feature data by taking one of the indicator sample unit feature representations or the first correction feature representations, and at least one of the state feature representations of the first sample propagation chain, as the supplier sample entity feature representations, to obtain the supplier sample entity feature representations. This can be implemented through the following example.

[0157] Step S151': Transmit the state feature representation and at least one mode parameter feature representation of each first sample propagation chain to at least one evaluation dimension extraction module to obtain at least one mode association feature representation of each first sample propagation chain.

[0158] The server initializes three independent "evaluation dimension extraction modules," each associated with a trainable pattern parameter feature representation. For example: Module 1 (compliance dimension): pattern parameter P_compliance; Module 2 (efficiency dimension): pattern parameter P_efficiency; Module 3 (product greenness dimension): pattern parameter P_greenness.

[0159] These mode parameters are vectors with random initial values, learned during training, and used to characterize the adaptation and correlation direction between the evaluation mode of interest to each module and the feature space of the indicator sample unit.

[0160] The server selects the target supplier "Supplier A" and its three first sample propagation chains, s_X, s_Y, and s_Z, as state features. For each chain, the server performs the following operations in parallel: Taking chain X ("Supplier A" - "Product A" - "Carbon Emission Intensity Index") as an example, the server inputs s_X and the three mode parameters P_compliance, P_efficiency, and P_greenness into the three evaluation dimension extraction modules respectively.

[0161] Each evaluation dimension extraction module internally executes a similar attention mechanism to step 6, but uses the pattern parameter as guidance. Specifically, in the compliance module, the server uses the pattern parameter P_compliance as a "query" and performs attention calculations with the associated entity features (product A, carbon emission indicators) extracted from chain X to obtain the weight distribution of the chain on the "compliance" dimension, and aggregates them to generate the "compliance pattern association feature representation" m_X_comp for chain X. Similarly, the efficiency module and the product greenness module output m_X_eff and m_X_grn, respectively. Therefore, after a propagation chain is processed by three modules, three pattern association feature representations of different dimensions are generated. The server performs the same operation on chains Y and Z, ultimately obtaining three sets (three for each chain) of pattern association features.

[0162] Step S152': Transmit at least one pattern association feature representation, at least one pattern parameter feature representation, and index sample unit feature representation of each first sample propagation chain to the pattern routing fusion module to obtain the expert feature representation of each first sample propagation chain.

[0163] The server obtains the "indicator sample unit feature representation" H_indicator (a global feature that integrates information from multiple indicators) generated by the first network branch and relevant to the current context. Subsequently, the server inputs the multi-dimensional pattern features of each chain, all pattern parameters, and the H_indicator into the pattern routing fusion module.

[0164] This module calculates the importance of each chain's different pattern dimensions within the context of the current global metric. Taking chain X as an example, the server calculates the similarity (e.g., dot product) between three pattern parameters, P_compliance, P_efficiency, and P_greenness, and the H_ metric, resulting in three unnormalized scores. Then, Softmax is applied to these three scores to obtain three pattern confidence coefficients, δ_comp, δ_eff, and δ_grn. These coefficients reflect the relative importance of chain X's information in the three dimensions of compliance, efficiency, and product greenness within the context of the global metric characteristics represented by the H_ metric.

[0165] Next, the server uses these coefficients to perform a weighted summation of the three pattern association features m_X_comp, m_X_eff, and m_X_grn of chain X: e_X = δ_comp * m_X_comp + δ_eff * m_X_eff + δ_grn * m_X_grn. The resulting e_X is the "expert feature representation" of chain X, which dynamically integrates information from different evaluation dimensions of the chain, with the integration weights adaptively determined by the global indicator features. The server performs the same operation on chains Y and Z, obtaining e_Y and e_Z.

[0166] Step S153': Transmit the expert feature representation of at least one first sample propagation chain to the equalization aggregation module to obtain the supplier sample entity feature representation.

[0167] Finally, the server inputs the expert feature representations e_X, e_Y, and e_Z from the three chains into the equalization aggregation module. This module is typically a self-attention layer. The server uses this module to calculate the mutual attention weights among e_X, e_Y, and e_Z, thereby evaluating the importance of each chain's expert features in comprehensively characterizing the overall green and low-carbon features of "Supplier A". Based on the calculated new weights, the server performs weighted aggregation on e_X, e_Y, and e_Z, ultimately outputting a comprehensive and robust "supplier sample entity feature representation" H_supplier. This representation not only integrates information from multiple associated paths of the supplier, but also undergoes multi-dimensional analysis within each path, and the aggregation between paths considers their interrelationships and overall contributions, thus achieving a deep characterization of the complex features of multiple types of suppliers.

[0168] In this embodiment of the invention, at least one state feature representation and at least one mode parameter feature representation of the first sample propagation chain are transmitted to the at least one evaluation dimension extraction module to obtain at least one mode association feature representation of each first sample propagation chain. This can be implemented through the following example.

[0169] From the state feature representation of the second propagation subchain, extract the feature representation of the target supplier sample entity and the feature representation of each associated sample entity that is propagated with the target supplier sample entity. The second propagation subchain is any sample propagation chain in the at least one first sample propagation chain, and the target supplier sample entity is any supplier sample entity in the second propagation subchain.

[0170] The feature representation of the target supplier sample entity is combined with at least one of the pattern parameter feature representations to obtain at least one first combined feature representation;

[0171] At least one of the first combined feature representations is subjected to exponential correlation calculation with the feature representation of each of the associated sample entities to obtain the sample routing confidence coefficient of each of the associated sample entities in the second propagation subchain.

[0172] Based on the sample routing confidence coefficient of each associated sample entity, confidence aggregation is performed on the feature representation of the corresponding associated sample entity to obtain at least one pattern association feature representation of the second propagation subchain.

[0173] In this embodiment of the invention, at least one pattern association feature representation, at least one pattern parameter feature representation, and the indicator sample unit feature representation of each first sample propagation chain are transmitted to the pattern routing fusion module to obtain an expert feature representation of each first sample propagation chain. This can be implemented through the following example.

[0174] For each of the first sample propagation chains, at least one of the pattern parameter feature representations is subjected to exponential correlation calculation with the feature representation of the index sample unit to obtain the confidence coefficient of each of the pattern correlation feature representations in each of the first sample propagation chains.

[0175] For each of the first sample propagation chains, based on the confidence coefficient of at least one of the pattern association feature representations, confidence aggregation is performed on the corresponding at least one of the pattern association feature representations to obtain the expert feature representation of each of the first sample propagation chains.

[0176] In this embodiment of the invention, the server executes the following specific operations of the pattern routing fusion module: First, the server retrieves the global "indicator sample unit feature representation" H_ index generated by the first network branch, as well as all pattern parameter feature representations (e.g., P_compliance, P_efficiency, P_greenness) from memory. For each first sample propagation chain (e.g., chain X) of the target supplier "Supplier A", the server performs the following processing: The server calculates the vector dot product of each pattern parameter feature representation and the H_ index to obtain three raw association scores. Subsequently, the server applies an exponential function (exp) to these three scores and performs normalization (Softmax) to obtain three "pattern confidence coefficients" (e.g., δ_X_comp, δ_X_eff, δ_X_grn) for chain X. These coefficients quantify the relative importance of the information of chain X in the three dimensions of compliance, efficiency, and product greenness within the context of the global indicator represented by the H_ index. Next, the server uses these coefficients as weights to perform a weighted summation of the three "pattern association feature representations" (m_X_comp, m_X_eff, m_X_grn) corresponding to chain X, i.e., calculating δ_X_comp*m_X_comp + δ_X_eff*m_X_eff + δ_X_grn*m_X_grn. The result of this calculation is the "expert feature representation" of chain X. The server repeats the above process for each propagation chain (chain Y, chain Z) of supplier A, thereby generating a dynamically integrated expert feature representation for each chain, which is then used by the subsequent balanced aggregation module.

[0177] In this embodiment of the invention, the state feature representation of each second sample propagation chain is sent to the first feature fusion module to obtain the state association feature representation of each second sample propagation chain. This can be implemented through the following example.

[0178] Step S141a: Extract the feature representation of the target index sample unit and the feature representation of each associated sample entity that is propagated with the target index sample unit from the state feature representation of the third propagation subchain.

[0179] The server focuses on a specific second-sample propagation chain, such as chain B: "Unit Product Carbon Emission Intensity Index" — "Third-Party Environmental Audit Report R" — "Supplier B". The state feature representation h_B of this chain is a serialized or graph-structured high-dimensional representation that integrates information from all nodes in the chain. The server first parses this representation and extracts the feature sub-vectors representing specific nodes.

[0180] Specifically, the server locates and separates from h_B:

[0181] 1. Feature representation of the target indicator sample unit: namely, the feature vector of the source node of the chain, "carbon emission intensity index per unit product", denoted as v_indicator_carbon. This vector encodes the semantic and numerical attributes of the indicator itself.

[0182] 2. Feature representation of associated sample entities: i.e., feature vectors of subsequent nodes in the chain. In this chain, it includes: v_report_R: the feature vector representing the intermediate entity "Third-Party Environmental Audit Report R"; v_supplier_Z: the feature vector representing the tail of the chain, "Supplier B".

[0183] Step S141b: Perform exponential correlation calculation on the feature representation of the target index sample unit and the feature representation of each associated sample entity to obtain the sample routing confidence coefficient of each associated sample entity in the third propagation subchain.

[0184] Next, the server calculates the importance weight of each associated entity in the chain for understanding "this specific path starting from the carbon emission indicator". The server performs an attention calculation process. For each associated sample entity (taking v_report_R as an example), the server calculates its association score with the target indicator entity v_indicator_carbon.

[0185] First, the server maps v_indicator_carbon to a "query" vector q_ind using a learnable linear transformation matrix W_Q', and v_report_R to a "key" vector k_report using another linear transformation matrix W_K'. Then, the server calculates the dot product of q_ind and k_report to obtain a scalar score score_report. This score initially characterizes the direct correlation strength between the "audit report R" and the "carbon emission intensity index" within this path context.

[0186] The server performs the same operation on another associated entity, v_supplier_Z, to obtain k_supplier and score_supplier.

[0187] Then, the server applies the Softmax function to normalize the two raw scores, score_report and score_supplier. The calculation process is: γ_report = exp(score_report) / (exp(score_report) + exp(score_supplier)), and γ_supplier is calculated similarly. Thus, the server obtains two values, γ_report and γ_supplier, between 0 and 1, that sum to 1. These are the "sample route confidence coefficients" for "Audit Report R" and "Supplier B" in this chain B. A high γ_report indicates that in this path, the intermediate carrier "Audit Report R" is a key information hub for understanding the association between this indicator and the supplier; a high γ_supplier may indicate that the information in this path is highly concentrated on the final supplier node itself.

[0188] Step S141c: Based on the sample routing confidence coefficient of each associated sample entity, perform confidence aggregation on the feature representation of the corresponding associated sample entity to obtain the state association feature representation of the third propagation subchain.

[0189] Finally, the server performs weighted aggregation of the information within the chain. The server multiplies the feature representation of each associated sample entity (usually using the "value" vectors v_report' and v_supplier' mapped through the W_V' matrix) with its corresponding sample route confidence coefficient, and then sums them.

[0190] That is, the server calculates: h_B'=γ_report*v_report'+γ_supplier*v_supplier'.

[0191] This aggregated vector h_B' is the "state association feature representation" of chain B. It is no longer a simple list or average of all node features, but a fused representation that has been recalibrated based on the association importance of nodes and source indicators in this specific path. It more succinctly characterizes "which intermediate information carriers and final supplier information are the crucial association states in this specific reverse path from carbon emission indicators to supplier B".

[0192] The server independently and in parallel executes the complete process from S141a to S141c for each second sample propagation chain (chain A, chain C, etc.), thereby generating a refined state association feature representation h_A', h_C', etc. for each chain, so that the subsequent first propagation chain can be aggregated across chains by the module to finally form the feature representation of the indicator sample unit.

[0193] In this embodiment of the invention, the initial evaluation network is trained based on the feature representation of the indicator sample unit and the feature representation of the supplier sample entity to obtain a green and low-carbon evaluation model, which can be implemented through the following example.

[0194] Calculate the deviation measure between the feature representation of the indicator sample unit and the feature representation of the supplier sample entity to obtain the deviation measure value;

[0195] The network connection weights of the initial evaluation network are adjusted based on the deviation metric to obtain a green and low-carbon evaluation model.

[0196] In this embodiment of the invention, exemplarily, during the model training phase, the server executes the following optimization process to obtain the final green and low-carbon evaluation model. The server first obtains the generated indicator sample unit feature representation H_indicator and supplier sample entity feature representation H_supplier from the current training batch. For each known positive sample pair in the batch (e.g., a known good match between "Supplier A" and "Carbon Emission Intensity Indicator"), the server calculates the similarity between H_Supplier_A and H_Indicator_Carbon, for example, using a cosine similarity function to obtain a high similarity score. Simultaneously, the server constructs several negative sample pairs for each positive sample pair through random sampling (e.g., pairing "Supplier A" with the unrelated "Wastewater Recycling Rate Indicator"), and calculates the similarity between the feature representations of these negative sample pairs, expecting to obtain a lower score.

[0197] Next, the server calculates a loss function based on the similarity scores of these positive and negative sample pairs, such as using contrastive learning loss (InfoNCELoss). This loss function encourages positive sample pairs to have the highest possible similarity scores and negative sample pairs to have the lowest possible scores; the result is the overall "deviation metric." This value quantifies the gap between the current model output and the ideal matching state.

[0198] Subsequently, the server initiates the backpropagation process. The server uses automatic differentiation to calculate the gradient of the deviation metric relative to all trainable parameters in the initial evaluation network (including the weights of each fusion module and routing module in the first and second network branches, as well as the first correction feature representation, the second correction feature representation, mode parameter features, etc.).

[0199] Finally, the server invokes an optimizer (such as the Adam optimizer) to iteratively update (tune) all connection weights and parameters in the network based on the calculated gradient direction and learning rate. This process is repeated over multiple training epochs until the model converges, i.e., the deviation metric stabilizes at a low level. At this point, the server saves the optimized network parameters, resulting in the trained "green and low-carbon evaluation model".

[0200] In this embodiment of the invention, obtaining at least one first sample propagation chain and at least one second sample propagation chain can be implemented through the following example.

[0201] Obtain multi-type association propagation structure data, which includes at least one first sample propagation chain and at least one second sample propagation chain. The multi-type association propagation structure data is used to characterize the propagation association structure between the supplier sample entity and the indicator sample unit.

[0202] In this embodiment of the invention, the server performs the following steps to obtain the sample propagation chain, as exemplarily. The server first extracts raw relational data from multiple connected enterprise data sources, including Supplier Relationship Management (SRM) systems, Enterprise Resource Planning (ERP) systems, Environmental Management Information Systems (EMIS), and third-party audit report databases. The server cleans, aligns, and merges this heterogeneous data to construct a unified heterogeneous information network. In this network, nodes are classified as "supplier sample entities" (e.g., "Supplier A," "Supplier B"), "product / process entities" (e.g., "Product A," "Process B"), "document / certification entities" (e.g., "Environmental Audit Report R," "Management System Certification C"), and "indicator sample units" (e.g., "Carbon Emission Intensity Indicator," "Wastewater Recycling Rate Indicator"). Edges between nodes are assigned types based on the actual relationships in the data, such as "producing," "having," "evaluated," "referenced," "compliant," etc., and may have weights.

[0203] Based on this constructed network, the server initiates two chain generation algorithms. For each supplier sample entity node, the server performs a restricted random walk or a meta-path-based traversal starting from that node. For example, starting from the "Supplier A" node, following the predefined meta-path "Supplier—Production—Product—Relationship—Indicator," a node sequence such as "Supplier A"—"Product A"—"Carbon Emission Intensity Indicator" is generated, forming a first sample propagation chain. The server generates multiple such chains for each supplier entity to cover its different relational aspects.

[0204] Meanwhile, for each indicator sample unit node, the server performs a reverse traversal. For example, starting from the "carbon emission intensity indicator" node, it follows the meta-path of "indicator - assessed - report - involved - supplier" to generate a node sequence such as "carbon emission intensity indicator" - "environmental audit report R" - "supplier B", which constitutes a second sample propagation chain.

[0205] The server serializes and stores all generated first and second sample propagation chains, along with node features and edge type information along their paths, into a structured dataset. This complete dataset is called "multi-type association propagation structure data." From a network topology perspective, it systematically depicts the complex, bidirectional propagation and association structure between supplier entities and green low-carbon indicator units through various intermediate nodes and relationship types, providing rich structured input for subsequent feature extraction and model training.

[0206] This invention provides a computer device 100, which includes a processor and a non-volatile memory storing computer instructions. When the computer instructions are executed by the processor, the computer device 100 executes the aforementioned training method for a green and low-carbon evaluation model based on multiple types of suppliers. Figure 2 As shown, Figure 2 This is a structural block diagram of a computer device 100 provided in an embodiment of the present invention. The computer device 100 includes a memory 111, a processor 112, and a communication unit 113. To enable data transmission or interaction, the memory 111, processor 112, and communication unit 113 are electrically connected to each other directly or indirectly. For example, these components can be electrically connected to each other through one or more communication buses or signal lines.

[0207] For illustrative purposes, the foregoing description has been made with reference to specific embodiments. However, the foregoing illustrative discussions are not intended to be exhaustive or to limit the present disclosure to the precise forms disclosed. Numerous modifications and variations are possible in accordance with the foregoing teachings. These embodiments were chosen and described in order to best illustrate the principles of the present disclosure and its practical application, thereby enabling those skilled in the art to best utilize the disclosure and to employ various embodiments with different modifications to suit a particular intended application.

Claims

1. A training method for a green and low-carbon evaluation model based on multiple types of suppliers, characterized in that, include: Obtain a set of green and low-carbon evaluation indicators, and determine indicator sample units based on the set of green and low-carbon evaluation indicators. The set of green and low-carbon evaluation indicators includes multiple evaluation indicators related to green and low-carbon. At least one first sample propagation chain and at least one second sample propagation chain are obtained. Each first sample propagation chain is used to characterize the topology between the supplier sample entity and the indicator sample unit when the supplier sample entity is the source node. Each second sample propagation chain is used to characterize the topology between the indicator sample unit and the supplier sample entity when the indicator sample unit is the source node. Extract the state feature representation of each first sample propagation chain and the state feature representation of each second sample propagation chain; At least one state feature representation of the second sample propagation chain is used as the received feature data of the first network branch in the initial evaluation network to obtain the indicator sample unit feature representation. At least one state feature representation of the first sample propagation chain, either the indicator sample unit feature representation or the first correction feature representation, is used as the received feature data of the second network branch in the initial evaluation network to obtain the supplier sample entity feature representation. The first correction feature representation is used to characterize the parameter association representation when the parameter mapping of the second network branch is associated with the generated feature representation result of the first network branch. The initial evaluation network is trained based on the feature representation of the indicator sample unit and the feature representation of the supplier sample entity to obtain a green and low-carbon evaluation model. The green and low-carbon evaluation model is used to process the state feature representation of at least one first target propagation chain and the state feature representation of at least one second target propagation chain to obtain the target indicator unit feature representation and the target supplier feature representation. The target indicator unit feature representation and the target supplier feature representation are used to determine the target-fit indicator unit.

2. The method according to claim 1, characterized in that, The first network branch includes a first feature fusion module and a first propagation link module; by using the state feature representation of at least one second sample propagation link as the received feature data of the first network branch in the initial evaluation network, the feature representation of the index sample unit is obtained, including: The state feature representation of each second sample propagation chain is sent to the first feature fusion module to obtain the state association feature representation of each second sample propagation chain. The state association feature representation of each second sample propagation chain is used to characterize the state association of the propagation association state in the corresponding second sample propagation chain. The state association feature representation of at least one second sample propagation chain is used as the received feature data of the first propagation link by the module to obtain the routing confidence coefficient of each second sample propagation chain. Based on the first propagation link, the module performs confidence aggregation on the state association feature representation of at least one second sample propagation link and the routing confidence coefficient to obtain the feature representation of the index sample unit.

3. The method according to claim 2, characterized in that, Before feeding the state feature representation of each second sample propagation chain into the first feature fusion module to obtain the state association feature representation of each second sample propagation chain, the method further includes: Obtain a second corrected feature representation, which is used to characterize the parameter association representation when the parameter mapping of the first network branch is associated with the generated feature representation result of the second network branch; The state feature representation of each second sample propagation chain is fed into the first feature fusion module to obtain the state association feature representation of each second sample propagation chain, including: The second correction feature representation is combined with the state feature representation of each second sample propagation chain to obtain the correction combination feature representation of each second sample propagation chain. The corrected combined feature representation of each second sample propagation chain is fed into the first feature fusion module to obtain the state association feature representation of each second sample propagation chain.

4. The method according to claim 3, characterized in that, Based on the first propagation link, the module performs confidence aggregation on the state association feature representation of at least one second sample propagation link and the routing confidence coefficient to obtain the feature representation of the index sample unit, including: Based on the first propagation link, the module performs confidence aggregation on the state association feature representation of at least one second sample propagation link and the routing confidence coefficient to obtain the aggregated feature representation; The aggregated feature representation and the second corrected feature representation are combined to obtain the feature representation of the index sample unit.

5. The method according to any one of claims 3 to 4, characterized in that, The second network branch includes a second feature fusion module and a second propagation link module; it uses one of the indicator sample unit feature representations or the first correction feature representation, and at least one of the state feature representations of the first sample propagation link, as the received feature data of the second network branch in the initial evaluation network to obtain the supplier sample entity feature representation, including: The first correction feature representation is combined with the state feature representation of each first sample propagation chain to obtain the combined feature representation of each first sample propagation chain. From the combined feature representation of the first propagation subchain, extract the feature representation of the target supplier sample entity and the feature representation of each associated sample entity that is propagated with the target supplier sample entity, wherein the first propagation subchain is any sample propagation chain in the at least one first sample propagation chain, and the target supplier sample entity is any supplier sample entity in the first propagation subchain; The feature representations of the target supplier sample entities and the feature representations of each associated sample entity are respectively subjected to exponential correlation calculation to obtain the sample routing confidence coefficient of each associated sample entity in the first propagation subchain; Based on the sample routing confidence coefficient of each associated sample entity, confidence aggregation is performed on the feature representation of the corresponding associated sample entity to obtain the state association feature representation of the first propagation sub-chain. The state association feature representation of each first sample propagation chain is used to characterize the state association of object interactions in the corresponding first sample propagation chain. The state association feature representation of each first sample propagation chain is used as the received feature data of the second propagation link module to obtain the routing confidence coefficient of each first sample propagation chain. Based on the second propagation link module, the state association feature representation of at least one first sample propagation link is aggregated with the route confidence coefficient to obtain the first propagation link feature representation; The first propagation link is processed by combining the feature representation with the first correction feature representation to obtain the supplier sample entity feature representation.

6. The method according to claim 2, characterized in that, The second network branch includes at least one evaluation dimension extraction module, a pattern routing fusion module, and a balance aggregation module; it uses one of the indicator sample unit feature representations or the first correction feature representation, and at least one of the state feature representations of the first sample propagation chain, as the received feature data of the second network branch in the initial evaluation network to obtain the supplier sample entity feature representation, including: From the state feature representation of the second propagation subchain, extract the feature representation of the target supplier sample entity and the feature representation of each associated sample entity that is propagated with the target supplier sample entity. The second propagation subchain is any sample propagation chain in the at least one first sample propagation chain, and the target supplier sample entity is any supplier sample entity in the second propagation subchain. The feature representations of the target supplier sample entities are combined with at least one pattern parameter feature representation to obtain at least one first combined feature representation. At least one of the first combined feature representations is subjected to exponential correlation calculation with the feature representation of each of the associated sample entities to obtain the sample routing confidence coefficient of each of the associated sample entities in the second propagation subchain. Based on the sample routing confidence coefficient of each associated sample entity, confidence aggregation is performed on the feature representation of the corresponding associated sample entity to obtain at least one pattern association feature representation of the second propagation subchain. Each evaluation dimension extraction module outputs one pattern association feature representation. Each pattern parameter feature representation is used to characterize the pattern adaptation association degree of the corresponding evaluation dimension extraction module for the indicator sample unit. At least one pattern association feature representation, at least one pattern parameter feature representation, and the index sample unit feature representation of each first sample propagation chain are transmitted to the pattern routing fusion module to obtain the expert feature representation of each first sample propagation chain. The expert feature representation of at least one of the first sample propagation chains is transmitted to the equalization aggregation module to obtain the supplier sample entity feature representation.

7. The method according to claim 6, characterized in that, At least one pattern association feature representation, at least one pattern parameter feature representation, and the index sample unit feature representation of each of the first sample propagation chains are transmitted to the pattern routing fusion module to obtain an expert feature representation for each of the first sample propagation chains, including: For each of the first sample propagation chains, at least one of the pattern parameter feature representations is subjected to exponential correlation calculation with the feature representation of the index sample unit to obtain the confidence coefficient of each of the pattern correlation feature representations in each of the first sample propagation chains. For each of the first sample propagation chains, based on the confidence coefficient of at least one of the pattern association feature representations, confidence aggregation is performed on the corresponding at least one of the pattern association feature representations to obtain the expert feature representation of each of the first sample propagation chains.

8. The method according to claim 2, characterized in that, The state feature representation of each second sample propagation chain is fed into the first feature fusion module to obtain the state association feature representation of each second sample propagation chain, including: From the state feature representation of the third propagation subchain, extract the feature representation of the target index sample unit and the feature representation of each associated sample entity that is propagated with the target index sample unit. The third propagation subchain is any sample propagation chain in the at least one second sample propagation chain, and the target index sample unit is any index sample unit in the third propagation subchain. The feature representations of the target index sample units and the feature representations of each associated sample entity are respectively subjected to exponential correlation calculation to obtain the sample routing confidence coefficient of each associated sample entity in the third propagation subchain. Based on the sample routing confidence coefficient of each associated sample entity, confidence aggregation is performed on the feature representation of the corresponding associated sample entity to obtain the state association feature representation of the third propagation subchain.

9. The method according to claim 1, characterized in that, The initial evaluation network is trained based on the feature representations of the indicator sample units and the feature representations of the supplier sample entities to obtain a green and low-carbon evaluation model, including: Calculate the deviation measure between the feature representation of the indicator sample unit and the feature representation of the supplier sample entity to obtain the deviation measure value; The network connection weights of the initial evaluation network are adjusted based on the deviation metric to obtain a green and low-carbon evaluation model.

10. A training system for a green and low-carbon evaluation model based on multiple types of suppliers, characterized in that, The green and low-carbon evaluation model training system based on multiple types of suppliers includes a processor and a memory, the memory and the processor being connected. The memory is used to store programs, instructions or code, and the processor is used to execute the programs, instructions or code in the memory to implement the green and low-carbon evaluation model training method based on multiple types of suppliers as described in any one of claims 1-9.

Images