Supply chain financial risk control model training method and system based on multi-modal data

Abstract

The application provides a supply chain financial risk control model training method and system based on multi-modal data, relates to the technical field of supply chain financial risk control, and comprises the following steps: constructing a risk control model and performing supervised training based on a risk control database; acquiring multi-modal data, extracting features through an encoder to generate a risk control vector; performing multi-strategy retrieval and semantic conflict detection in the risk control database, and outputting a labeling result; constructing a risk credibility evaluation model, inputting the labeling result into the model to output a credibility index, and generating a credibility score through weighted fusion and confidence calibration; classifying the labeling result according to the credibility score, generating high-quality labeling samples through artificial review and correction, performing incremental training on the risk control model, and obtaining a trained risk control model, so that unified representation of multi-modal data and efficient reuse of historical samples can be realized, a full-link, closed-loop and self-evolving risk control model training system can be constructed, and the risk identification credibility and cross-scene adaptation efficiency can be significantly improved.

Description

Technical Field

[0001] This invention relates to the field of supply chain finance risk control technology, and in particular to a method and system for training supply chain finance risk control models based on multimodal data. Background Technology

[0002] Supply chain finance is a key business model for serving the financing needs of small and medium-sized enterprises and stabilizing industrial and supply chains. Its risk control capabilities directly determine the quality of credit assets and the scale of business.

[0003] However, existing training methods for supply chain finance risk control models still suffer from the following technical shortcomings in practical applications: First, multi-source heterogeneous data cannot be uniformly represented. Transaction logs, logistics timelines, invoice and contract images, corporate financial statements, public opinion texts, and credit knowledge graphs are processed independently, resulting in severe modal fragmentation and making it difficult to uncover cross-modal risk correlation features, thus forming data silos. Second, a large number of labeled corporate default, overdue, and fraud samples have not been compiled into an efficient retrieval vector library. Similar companies and similar transaction scenarios require repeated manual labeling, leading to high labeling costs, long cycles, and poor consistency. Third, the reliability of risk predictions cannot be quantitatively verified. The model only outputs a single confidence level, which is prone to high-confidence erroneous judgments. It lacks multi-dimensional evaluation mechanisms such as uncertainty, label consistency, and out-of-distribution detection, resulting in a high risk misjudgment rate. Finally, the parameters are fixed after the model is deployed, making it impossible to achieve incremental learning and continuous optimization through manual review and risk backtracking. After long-term operation, the risk identification capability declines significantly.

[0004] Therefore, it is necessary to provide a training method and system for supply chain finance risk control models based on multimodal data to solve the above-mentioned technical problems. Summary of the Invention

[0005] To address the aforementioned technical problems, this invention provides a training method and system for supply chain finance risk control models based on multimodal data. This system is used to solve the problems in existing supply chain finance risk control, such as the difficulty in uniformly representing multi-source heterogeneous data, low efficiency in reusing historical labeled samples, inability to quantitatively assess risk credibility, and the lack of closed-loop self-evolution capabilities in the model.

[0006] The present invention provides a method for training a supply chain finance risk control model based on multimodal data, the method comprising: A supply chain finance risk control model is constructed, and the model is trained under supervision based on a pre-set supply chain finance risk control database. Acquire multimodal data of supply chain finance, and extract features from the multimodal encoder in the supply chain finance risk control model to generate a multimodal risk control representation vector; Based on the multimodal risk control representation vector, multi-strategy retrieval and multimodal semantic conflict detection are performed in the supply chain finance risk control database, and risk control annotation results are output. A multidimensional risk credibility assessment model is constructed. The risk control labeling results are input into the multidimensional risk credibility assessment model, and multidimensional credibility indicators are output. The multidimensional credibility indicators are weighted, fused, and calibrated to generate a risk credibility score. The risk control labeling results are classified according to the risk credibility score, and high-quality labeled training samples are generated through manual review and correction. The supply chain finance risk control model is incrementally trained based on the high-quality labeled training samples to obtain the trained supply chain finance risk control model.

[0007] Preferably, the construction of the supply chain finance risk control model, based on a pre-set supply chain finance risk control database, includes supervised training of the supply chain finance risk control model, specifically including: Obtain historical multimodal data samples and risk labels corresponding to the historical multimodal data samples from the supply chain finance risk control database; The historical multimodal data samples are input into the supply chain finance risk control model, and the features of the historical multimodal data samples are extracted by the multimodal encoder to generate a historical multimodal risk control representation vector. The historical multimodal risk control representation vector is mapped to an initial predicted risk label using the risk classifier in the supply chain finance risk control model. Based on the cross-entropy loss function, the loss value between the initial predicted risk label and the risk label corresponding to the historical multimodal data sample is calculated, and the model parameters of the multimodal encoder and the risk classifier are iteratively updated according to the loss value until the preset convergence condition is met, so as to obtain the pre-trained supply chain finance risk control model.

[0008] Preferably, the step of acquiring supply chain finance multimodal data, and extracting features from the supply chain finance multimodal data using the multimodal encoder in the supply chain finance risk control model to generate a multimodal risk control representation vector, specifically includes: Obtain single-modal data of supply chain finance from the multimodal data of supply chain finance; Based on the data type of the single-modal supply chain finance data, the corresponding single-modal encoder is called from the multimodal encoder, and the single-modal encoder is used to extract features from the single-modal supply chain finance data to generate a single-modal supply chain finance modal feature map. Based on the cross-modal contrastive learning mechanism, positive and negative sample pairs of the feature maps of the supply chain finance modality in different modalities are constructed. With the constraint of aligning the semantic features of the positive sample pairs and dispersing the semantic features of the negative sample pairs, the single modality encoder is iteratively optimized so that the feature maps of the supply chain finance modality are mapped to a shared semantic space and fused to generate a basic unified vector. A preset enterprise meta-information vector is obtained, and an attention mechanism is used to calculate the dimensional modulation weight of the enterprise meta-information vector on the basic unified vector. The enterprise meta-information vector is then embedded into the basic unified vector according to the dimensional modulation weight to generate the multimodal risk control representation vector.

[0009] Preferably, based on the multimodal risk control representation vector, a multi-strategy retrieval is performed in the supply chain finance risk control database to generate preliminary risk labels and preliminary default probabilities, specifically including: Based on the multimodal risk control representation vector, the enterprise meta-information vector is matched layer by layer in the supply chain finance risk control database to select candidate samples that meet the preset layer matching conditions and generate a candidate sample set. In the candidate sample set, a similarity algorithm is used to perform precise similarity retrieval on the multimodal risk control representation vector to generate a first similarity value, and risk labels corresponding to candidate samples that meet the preset first similarity conditions are selected to generate a first risk label set; In the candidate sample set, based on the transaction pattern and payment period structure, a weighted similarity algorithm is used to perform cross-scenario generalization retrieval on the multimodal risk control representation vector to generate a second similarity value, and to select the risk labels corresponding to the candidate samples that meet the preset second similarity conditions to generate a second risk label set; Integrate the first risk tag set and the second risk tag set, and remove duplicates according to the preset risk tag categories to generate a candidate risk tag set. Normalize the similarity values ​​in the candidate risk tag set to generate the voting weight of the risk tag. Based on the risk label category, the voting weights of the risk labels in the same category are accumulated and aggregated to generate a risk label weighted probability. The risk label with the highest risk label weighted probability is taken as the initial risk label, and the risk label weighted probability is taken as the initial default probability.

[0010] Preferably, performing multimodal semantic conflict detection on the preliminary risk labels specifically includes: Obtain the supply chain finance modal feature map, input the supply chain finance modal feature map into a preset modal risk classifier, and output the modal prediction probability vector; Based on the modality prediction probability vector, the semantic conflict degree C of the preliminary risk label is calculated, and the corresponding calculation formula is as follows: In the formula, M represents the total number of modalities detected in multimodal semantic conflict detection; K represents the total number of categories of the initial risk label; This represents the probability value of the initial risk label of the k-th class in the modality prediction probability vector of the m-th modality; This represents the average probability that all M modalities are predicted to be the initial risk label of the kth class; If the semantic conflict degree is less than the preset conflict threshold, then the preliminary risk label is determined to have passed the multimodal semantic conflict detection, and the preliminary risk label is retained. If the semantic conflict degree is greater than or equal to the preset conflict threshold, it is determined that the preliminary risk label has failed the multimodal semantic conflict detection. The parameters of the multi-strategy retrieval process are adjusted to iteratively execute the multi-strategy retrieval and multimodal semantic conflict detection until the number of iterations reaches the preset maximum number of iterations. The preliminary risk label with the lowest semantic conflict degree is retained. The preliminary risk label and the preliminary default probability are used as the risk control labeling results.

[0011] Preferably, the construction of the multidimensional risk credibility assessment model involves inputting the risk control labeling results into the multidimensional risk credibility assessment model, outputting multidimensional credibility indicators, and performing weighted fusion and confidence calibration on the multidimensional credibility indicators to generate a risk credibility score. Specifically, this includes: The multidimensional risk credibility assessment model is constructed, wherein the multidimensional risk credibility assessment model includes a model uncertainty assessment unit, a label consistency assessment unit, an out-of-distribution assessment unit, a cross-modal consistency assessment unit, and an anti-interference robustness assessment unit; The risk control labeling results and intermediate representation data generated during the risk control labeling process are input into the multidimensional risk credibility assessment model. The assessment unit calculates and outputs the model uncertainty index, label consistency index, out-of-distribution index, cross-modal consistency index, and anti-interference robustness index, respectively. The model uncertainty index, the label consistency index, the out-of-distribution index, the cross-modal consistency index, and the anti-interference robustness index are weighted and fused to generate an initial credibility score; Based on a preset calibration sample set, with the goal of minimizing the error between the initial credibility score and the experienced default frequency in the calibration sample set, the initial credibility score is probabilistically mapped and calibrated, mapping the initial credibility score to the numerical value of the experienced default frequency, thereby generating the risk credibility score.

[0012] Preferably, the step of classifying the risk control labeling results based on the risk credibility score and generating high-quality labeled training samples through manual review and correction specifically includes: The risk credibility score is compared with a preset first credibility threshold and a second credibility threshold, wherein the first credibility threshold is greater than the second credibility threshold; If the risk credibility score is greater than or equal to the first credibility threshold, the risk control labeling result is marked as a high-credibility automatic confirmation sample. If the risk credibility score is greater than or equal to the second credibility threshold and less than the first credibility threshold, then the risk control labeling result is marked as a sample to be manually reviewed, the risk credibility scores are sorted in ascending order, and a manual review queue is generated. If the risk credibility score is less than the second credibility threshold, the risk control labeling result is marked as a high-risk priority processing sample and placed at the top of the manual review queue. Based on the highly reliable automatically confirmed samples, the samples awaiting manual review after verification, and the high-risk priority processing samples, the high-quality labeled training samples are generated and written into the supply chain finance risk control database.

[0013] Preferably, the incremental training of the supply chain finance risk control model based on the high-quality labeled training samples to obtain the trained supply chain finance risk control model specifically includes: Monitor the cumulative number of the high-quality labeled training samples in the supply chain finance risk control database, and trigger incremental training when the cumulative number reaches a preset incremental trigger threshold; Based on the high-quality labeled training samples and the historical labeled training samples in the supply chain finance risk control database, an incremental training sample set is constructed. Based on the incremental training sample set, the underlying network parameters used to extract basic semantic features in the multimodal encoder are frozen, the incremental training sample set is input into the supply chain finance risk control model, and the predicted risk label is output through the risk classifier in the supply chain finance risk control model. The parameters of the high-level network in the multimodal encoder and the parameters of the risk classifier are updated in multiple steps using the cross-entropy loss between the predicted risk label and the risk label corresponding to the high-quality labeled training sample. Using the difference between the high-level network parameters and the risk classifier parameters after the multi-step inner loop adaptive update and the parameters before incremental training as the meta gradient, the outer loop optimization update is performed on the high-level network parameters and the risk classifier parameters to generate the trained supply chain finance risk control model.

[0014] Preferably, the outer-loop optimization update is performed on the parameters of the high-level network and the parameters of the risk classifier, and the corresponding calculation formula is as follows: In the formula, This represents the set of parameters for the high-level network and the risk classifier after the outer ring optimization and update; This represents the set of parameters for the high-level network and the risk classifier before incremental training. This represents the preset outer loop learning rate; This represents the number of meta-learning tasks divided from the incremental training sample set; This represents the set of high-level network parameters and risk classifier parameters after multi-step inner loop adaptive update in the i-th meta-learning task; Let represent the meta-gradient of the i-th meta-learning task; Based on the updated high-level network parameters after outer ring optimization and the parameter set of the risk classifier The parameters of the high-level network and the parameters of the risk classifier in the supply chain finance risk control model are updated to generate the trained supply chain finance risk control model.

[0015] A supply chain finance risk control model training system based on multimodal data, the system comprising: The model pre-training module is used to build a supply chain finance risk control model and to conduct supervised training of the supply chain finance risk control model based on a pre-set supply chain finance risk control database. The multimodal representation module is used to acquire multimodal data of supply chain finance, and to extract features from the multimodal data of supply chain finance through the multimodal encoder in the supply chain finance risk control model to generate multimodal risk control representation vectors. The retrieval and conflict detection module is used to perform multi-strategy retrieval and multi-modal semantic conflict detection in the supply chain finance risk control database based on the multimodal risk control representation vector, and output the risk control labeling results. The credibility assessment module is used to construct a multi-dimensional risk credibility assessment model. The risk control labeling results are input into the multi-dimensional risk credibility assessment model, and multi-dimensional credibility indicators are output. The multi-dimensional credibility indicators are weighted, fused, and calibrated to generate a risk credibility score. The incremental optimization module is used to classify the risk control labeling results according to the risk credibility score, generate high-quality labeled training samples through manual review and correction, and perform incremental training on the supply chain finance risk control model based on the high-quality labeled training samples to obtain the trained supply chain finance risk control model.

[0016] Compared with related technologies, the supply chain finance risk control model training method and system based on multimodal data provided by this invention have the following beneficial effects: This invention constructs a supply chain finance risk control model and conducts supervised training on the model based on a pre-set supply chain finance risk control database. It acquires multimodal supply chain finance data, extracts features from this data using a multimodal encoder within the risk control model, and generates multimodal risk control representation vectors. Based on these vectors, it performs multi-strategy retrieval and multimodal semantic conflict detection within the supply chain finance risk control database, outputting risk control annotation results. Finally, it constructs a multidimensional risk credibility assessment model and inputs the risk control annotation results into this model. The system outputs multi-dimensional credibility indicators, performs weighted fusion and confidence calibration on these indicators, and generates a risk credibility score. Based on the risk credibility score, the risk control labeling results are categorized, and high-quality labeled training samples are generated through manual review and correction. Incremental training of the supply chain finance risk control model is then performed based on these high-quality labeled training samples, resulting in a trained supply chain finance risk control model. This enables unified representation of multi-modal data and efficient reuse of historical samples, constructing a full-link, closed-loop, and self-evolving risk control model training system, significantly improving the high credibility of risk identification and cross-scenario adaptability.

[0017] This invention constructs a supply chain finance risk control model and conducts supervised training based on a pre-set supply chain finance risk control database. It utilizes a multimodal encoder to uniformly represent multi-source heterogeneous data, generating a multimodal risk control representation vector that integrates enterprise metadata. This effectively breaks down data silos and enables in-depth mining of cross-modal risk features. By performing multi-strategy retrieval in the supply chain finance risk control database and combining it with a multimodal semantic conflict detection mechanism to output risk control annotation results, the reuse efficiency and annotation consistency of historical annotated training samples are significantly improved, greatly reducing manual annotation costs. Furthermore, by constructing a multi-dimensional risk credibility assessment model that includes model uncertainty, label consistency, out-of-distribution degree, cross-modal consistency, and robustness against interference, the model performs weighted fusion and confidence calibration of multi-dimensional credibility indicators to generate a risk credibility score mapped to the frequency of experienced defaults. This effectively intercepts high-confidence erroneous judgments and reduces the default false alarm rate. Based on the risk credibility score, the samples are classified into high-credibility automatically confirmed samples, samples awaiting manual review, and high-risk priority processing samples. Relying on the human-machine collaboration mechanism, accurate review and label correction are achieved, generating high-quality labeled training samples. The model incremental training is triggered through closed-loop knowledge feedback, and the model parameters are optimized and updated in the inner and outer loops using a meta-learning strategy, so that the supply chain finance risk control model can continuously evolve in long-term operation and the risk identification accuracy can be continuously improved. Attached Figure Description

[0018] Figure 1 A flowchart illustrating the training method for a supply chain finance risk control model based on multimodal data provided in this embodiment of the invention; Figure 2A system block diagram of a supply chain finance risk control model training system based on multimodal data provided in an embodiment of the present invention; Figure 3 This is a schematic diagram of the hardware structure of an electronic device provided in an embodiment of the present invention. Detailed Implementation

[0019] 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 embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] like Figure 1 The diagram shown is a flowchart of a supply chain finance risk control model training method based on multimodal data provided in an embodiment of the present invention. Figure 1 The execution entity of the method shown can be a software and / or hardware device. The execution entity of this application can include, but is not limited to, at least one of the following: user equipment, network equipment, etc. User equipment can include, but is not limited to, computers, smartphones, personal digital assistants (PDAs), and the aforementioned electronic devices. Network equipment can include, but is not limited to, a single network server, a server group consisting of multiple network servers, or a cloud based on cloud computing consisting of a large number of computers or network servers. Cloud computing is a type of distributed computing, consisting of a super virtual computer composed of a group of loosely coupled computers. This embodiment does not limit this. Steps S1 to S5 are detailed as follows: S1, Construct a supply chain finance risk control model, and conduct supervised training on the supply chain finance risk control model based on a preset supply chain finance risk control database; The construction of the supply chain finance risk control model, based on a pre-set supply chain finance risk control database, includes supervised training of the model, specifically including: Obtain historical multimodal data samples and risk labels corresponding to the historical multimodal data samples from the supply chain finance risk control database; The historical multimodal data samples are input into the supply chain finance risk control model, and the features of the historical multimodal data samples are extracted by the multimodal encoder to generate a historical multimodal risk control representation vector. The historical multimodal risk control representation vector is mapped to an initial predicted risk label using the risk classifier in the supply chain finance risk control model. Based on the cross-entropy loss function, the loss value between the initial predicted risk label and the risk label corresponding to the historical multimodal data sample is calculated, and the model parameters of the multimodal encoder and the risk classifier are iteratively updated according to the loss value until the preset convergence condition is met, so as to obtain the pre-trained supply chain finance risk control model.

[0021] The supply chain finance risk control database is a pre-built data management system that stores historical multimodal data samples and their corresponding risk labels. The historical multimodal data samples are heterogeneous records from multiple sources, including transaction timelines, logistics IoT data, invoice and contract images, corporate financial statements, public opinion texts, and credit knowledge graphs, generated by enterprises in their past financing activities. Risk labels are classification markers assigned to each sample by risk control experts based on the enterprise's actual repayment performance and credit status, typically including categories such as normal, overdue, default, and fraud.

[0022] Next, the multimodal encoder consists of multiple single-modal encoders, including a temporal encoder, an image encoder, a table encoder, a text encoder, and a knowledge graph encoder. Each encoder processes the raw data of its corresponding modality and maps it to a shared semantic space of a unified dimension. The historical multimodal risk control representation vector is a high-dimensional feature vector generated by the multimodal encoder after extracting historical multimodal data samples, aligning them through cross-modal contrastive learning, and embedding enterprise metadata. This vector integrates the semantic features of different modal data and metadata such as enterprise identity, industry, size, payment terms, and guarantee methods. The risk classifier is a neural network mapping layer connected after the multimodal encoder, responsible for mapping the historical multimodal risk control representation vector to the predicted probabilities of each risk category and outputting the initial predicted risk label.

[0023] Understandably, the cross-entropy loss function measures the difference between the predicted probability distribution of the initial predicted risk labels and the actual risk label distribution. This difference is quantified as a loss value, serving as a supervisory signal to guide model parameter updates. The preset convergence condition refers to a pre-defined criterion for terminating iterative training. When the loss value no longer decreases significantly over multiple consecutive training cycles or the model's prediction accuracy on the validation set meets the expected requirements, the convergence condition is satisfied, training stops, and a fully trained pre-trained supply chain finance risk control model is obtained. This model possesses the basic ability to extract risk features from multimodal data and perform preliminary risk assessment.

[0024] By using the above methods, the supply chain finance risk control model is trained end-to-end with historical multimodal data samples and corresponding risk labels. This enables the multimodal encoder to extract features by mapping multi-source heterogeneous data to a shared semantic space. The risk classifier establishes an accurate mapping relationship from multimodal risk control representation vectors to risk categories, thereby obtaining a pre-trained model that can perform preliminary risk judgment on multimodal supply chain finance data.

[0025] S2, acquire supply chain finance multimodal data, extract features from the supply chain finance multimodal data through the multimodal encoder in the supply chain finance risk control model, and generate a multimodal risk control representation vector; The acquisition of multimodal data in supply chain finance, and the extraction of features from the multimodal encoder in the supply chain finance risk control model to generate a multimodal risk control representation vector, specifically includes: Obtain single-modal data of supply chain finance from the multimodal data of supply chain finance; Based on the data type of the single-modal supply chain finance data, the corresponding single-modal encoder is called from the multimodal encoder, and the single-modal encoder is used to extract features from the single-modal supply chain finance data to generate a single-modal supply chain finance modal feature map. Based on the cross-modal contrastive learning mechanism, positive and negative sample pairs of the feature maps of the supply chain finance modality in different modalities are constructed. With the constraint of aligning the semantic features of the positive sample pairs and dispersing the semantic features of the negative sample pairs, the single modality encoder is iteratively optimized so that the feature maps of the supply chain finance modality are mapped to a shared semantic space and fused to generate a basic unified vector. A preset enterprise meta-information vector is obtained, and an attention mechanism is used to calculate the dimensional modulation weight of the enterprise meta-information vector on the basic unified vector. The enterprise meta-information vector is then embedded into the basic unified vector according to the dimensional modulation weight to generate the multimodal risk control representation vector.

[0026] Among them, supply chain finance single-modal data is independent modal data extracted from multimodal supply chain finance data according to data type. Examples include individual transaction time-series data, logistics IoT data, invoice and contract images, corporate financial statements, public opinion texts, or credit knowledge graphs. Each type of data retains its original modal form before entering the model. The single-modal encoder is a dedicated feature extraction network designed for specific data types. The time-series encoder processes transaction and logistics time-series data, the image encoder processes invoice and contract images, the table encoder processes corporate financial statements, the text encoder processes public opinion texts, and the knowledge graph encoder processes credit knowledge graphs. Each encoder independently transforms the original data of its corresponding modality into a high-dimensional supply chain finance modal feature map.

[0027] Understandably, cross-modal contrastive learning is an unsupervised representation alignment strategy. Its core idea is to construct positive sample pairs from feature maps of different modalities within the same enterprise or transaction, and negative sample pairs from feature maps of different enterprises or transactions. The encoder is trained by narrowing the distance between positive sample pairs and widening the distance between negative sample pairs in the feature space, forcing semantically similar data from different modalities to map to similar positions within the same shared semantic space. This shared semantic space is a vector space of unified dimension, where data from different modalities are distributed according to semantic relationships, with semantically similar samples clustered and semantically different samples separated. The basic unified vector is a comprehensive vector obtained by fusing the aligned supply chain finance modal feature maps from each modality, possessing cross-modal semantic consistency.

[0028] Specifically, the enterprise meta-information vector is a representation of structured attributes such as enterprise identity, industry type, business scale, payment term, guarantee method, and enterprise relevance mapped into a continuous vector through an embedding layer. The attention mechanism is used to calculate the modulation weights of the enterprise meta-information vector on each dimension of the basic unified vector. Essentially, it dynamically adjusts the activation intensity of different dimensions in the basic unified vector based on enterprise attributes, ensuring that the final multimodal risk control representation vector contains both semantic information from business data and embeds a specific business profile of the enterprise.

[0029] S3, based on the multimodal risk control representation vector, perform multi-strategy retrieval and multimodal semantic conflict detection in the supply chain finance risk control database, and output the risk control labeling results; Based on the multimodal risk control representation vector, a multi-strategy retrieval is performed in the supply chain finance risk control database to generate preliminary risk labels and preliminary default probabilities, specifically including: Based on the multimodal risk control representation vector, the enterprise meta-information vector is matched layer by layer in the supply chain finance risk control database to select candidate samples that meet the preset layer matching conditions and generate a candidate sample set. In the candidate sample set, a similarity algorithm is used to perform precise similarity retrieval on the multimodal risk control representation vector to generate a first similarity value, and risk labels corresponding to candidate samples that meet the preset first similarity conditions are selected to generate a first risk label set; In the candidate sample set, based on the transaction pattern and payment period structure, a weighted similarity algorithm is used to perform cross-scenario generalization retrieval on the multimodal risk control representation vector to generate a second similarity value, and to select the risk labels corresponding to the candidate samples that meet the preset second similarity conditions to generate a second risk label set; Integrate the first risk tag set and the second risk tag set, and remove duplicates according to the preset risk tag categories to generate a candidate risk tag set. Normalize the similarity values ​​in the candidate risk tag set to generate the voting weight of the risk tag. Based on the risk label category, the voting weights of the risk labels in the same category are accumulated and aggregated to generate a risk label weighted probability. The risk label with the highest risk label weighted probability is taken as the initial risk label, and the risk label weighted probability is taken as the initial default probability.

[0030] The layer-by-layer matching of enterprise metadata vectors refers to comparing the enterprise metadata vectors of the sample to be retrieved with the corresponding layer-level labels of samples already stored in the supply chain finance risk control database, according to a hierarchical index structure based on industry, scale, payment term, and guarantee method. This quickly eliminates samples with significant differences in metadata structure, and initially narrows down the candidate sample set. The candidate sample set is a group of historical samples that have been retained after layer-by-layer screening and are similar to the sample to be retrieved in the metadata dimension.

[0031] Understandably, precise similarity retrieval uses cosine similarity and other metrics to calculate the overall semantic similarity of multimodal risk control representation vectors within a candidate sample set, constrained by the same industry, scale, and guarantee method. This generates a first similarity value, and candidate samples with first similarity values ​​higher than a preset first similarity threshold are selected, forming a first risk label set. Cross-scenario generalized retrieval, on the other hand, is not limited to the same industry constraint within the candidate sample set. Using transaction pattern similarity and payment term structure comparability as core matching criteria, it calculates a weighted similarity value by assigning weighted coefficients to dimensions such as transaction frequency, repayment cycle, and capital occupation characteristics. This generates a second similarity value, and risk labels corresponding to candidate samples with similar transaction structures are retrieved from historical samples across industries and scenarios, forming a second risk label set.

[0032] Furthermore, the voting weight is a weighted coefficient on a uniform scale, converted from the normalized similarity values ​​of each label in the first and second risk label sets, reflecting the contribution of each similar sample to the final risk assessment. The risk label weighted probability is a comprehensive score for the risk category obtained by summing and aggregating the voting weights of risk labels of the same category. The risk label with the highest probability is selected as the preliminary risk label, and this probability value itself serves as a quantitative estimate of the preliminary default probability.

[0033] By integrating the dual-path recall results of precise similarity retrieval and cross-scenario generalized retrieval, and using normalized similarity value as the voting weight for weighted aggregation, the over-reliance of a single retrieval strategy on homogeneous scenarios is avoided. This allows the generation of preliminary risk labels and preliminary default probabilities to utilize the reference value of historical samples in the same industry and draw on the experience of similar transaction patterns across industries, thereby improving the retrieval recall rate and the robustness of label inference.

[0034] Perform multimodal semantic conflict detection on the preliminary risk labels, specifically including: Obtain the supply chain finance modal feature map, input the supply chain finance modal feature map into a preset modal risk classifier, and output the modal prediction probability vector; Based on the modality prediction probability vector, the semantic conflict degree C of the preliminary risk label is calculated, and the corresponding calculation formula is as follows: In the formula, M represents the total number of modalities detected in multimodal semantic conflict detection; K represents the total number of categories of the initial risk label; This represents the probability value of the initial risk label of the k-th class in the modality prediction probability vector of the m-th modality; This represents the average probability that all M modalities are predicted to be the initial risk label of the kth class; If the semantic conflict degree is less than the preset conflict threshold, then the preliminary risk label is determined to have passed the multimodal semantic conflict detection, and the preliminary risk label is retained. If the semantic conflict degree is greater than or equal to the preset conflict threshold, it is determined that the preliminary risk label has failed the multimodal semantic conflict detection. The parameters of the multi-strategy retrieval process are adjusted to iteratively execute the multi-strategy retrieval and multimodal semantic conflict detection until the number of iterations reaches the preset maximum number of iterations. The preliminary risk label with the lowest semantic conflict degree is retained. The preliminary risk label and the preliminary default probability are used as the risk control labeling results.

[0035] The modality prediction probability vector is the prediction result vector independently output by each modality after the supply chain finance modality feature map is input into a pre-set modality risk classifier. Each dimension corresponds to the probability of the modality's tendency towards various risk labels. The pre-set modality risk classifier is a lightweight classifier pre-trained for each individual modality, used to independently evaluate the probability that a single modality's data points to each risk category. Semantic conflict degree measures whether the risk prediction distributions of different modalities for the same transaction are consistent. Its calculation process involves two steps: first, averaging the modality prediction probability vectors of all modalities by category to obtain the average probability of each risk category; then, calculating the normalized information entropy of this average probability distribution. The higher the entropy value, the greater the discrepancy in the predictions of different modalities and the more severe the semantic conflict.

[0036] Specifically, the preset conflict threshold is a manually set boundary value. When the semantic conflict degree is below this threshold, it is considered that the prediction tendencies of different modalities are basically consistent, and the preliminary risk label is valid at the semantic level. When the semantic conflict degree reaches or exceeds this threshold, it indicates that there is a significant judgment discrepancy between the modalities, and the credibility of the label is questionable. Iterative execution of adjusting multi-strategy retrieval parameters refers to readjusting retrieval parameters such as similarity thresholds or weighting coefficients in multi-strategy retrieval when multi-modal semantic conflict detection fails, regenerating preliminary risk labels, and re-performing conflict detection. After multiple attempts, the label with the lowest semantic conflict degree is selected as the final result, thereby suppressing the risk of misjudgment caused by incomplete consistency of cross-modal information.

[0037] In practical applications, assuming the initial risk label categories are "normal" and "default," and the preset conflict threshold is set to 0.50, when a company applies for financing, it inputs invoice images, transaction records, and public opinion texts into their respective preset modal risk classifiers, resulting in three modal prediction probability vectors. Modal 1 outputs a default probability of 0.92 and a normal probability of 0.08; Modal 2 outputs a default probability of 0.94 and a normal probability of 0.06; and Modal 3 outputs a default probability of 0.93 and a normal probability of 0.07. The calculated average probability for the default category is 0.93, and the average probability for the normal category is 0.07. The semantic conflict degree is 0.366, which is less than the preset conflict threshold of 0.50, indicating that the three modalities have a highly similar tendency to predict the default label. The initial risk label passes the multimodal semantic conflict detection and is retained. This label, along with the initial default probability, is output as the risk control labeling result.

[0038] S4. Construct a multi-dimensional risk credibility assessment model, input the risk control labeling results into the multi-dimensional risk credibility assessment model, output multi-dimensional credibility indicators, perform weighted fusion and confidence calibration on the multi-dimensional credibility indicators, and generate a risk credibility score. The construction of a multidimensional risk credibility assessment model involves inputting the risk control labeling results into the model, outputting multidimensional credibility indicators, and performing weighted fusion and confidence calibration on these indicators to generate a risk credibility score. Specifically, this includes: The multidimensional risk credibility assessment model is constructed, wherein the multidimensional risk credibility assessment model includes a model uncertainty assessment unit, a label consistency assessment unit, an out-of-distribution assessment unit, a cross-modal consistency assessment unit, and an anti-interference robustness assessment unit; The risk control labeling results and intermediate representation data generated during the risk control labeling process are input into the multidimensional risk credibility assessment model. The assessment unit calculates and outputs the model uncertainty index, label consistency index, out-of-distribution index, cross-modal consistency index, and anti-interference robustness index, respectively. The model uncertainty index, the label consistency index, the out-of-distribution index, the cross-modal consistency index, and the anti-interference robustness index are weighted and fused to generate an initial credibility score; Based on a preset calibration sample set, with the goal of minimizing the error between the initial credibility score and the experienced default frequency in the calibration sample set, the initial credibility score is probabilistically mapped and calibrated, mapping the initial credibility score to the numerical value of the experienced default frequency, thereby generating the risk credibility score.

[0039] The model uncertainty assessment unit quantifies the model's inherent uncertainty regarding the current prediction results. This includes both random uncertainty reflecting data noise and cognitive uncertainty reflecting insufficient model knowledge; combining these two aspects helps identify unreliable predictions caused by insufficient training coverage. The label consistency assessment unit analyzes the distribution concentration and entropy of risk labels in retrieved historical similar samples to measure the consistency between the current risk control labeling results and the labeling patterns of similar case groups. Lower consistency indicates less reliable labeling results. The out-of-distribution assessment unit quantifies the distance of the current sample from the training data distribution based on an energy function. Higher out-of-distribution levels indicate a greater likelihood that the current sample is in a knowledge blind spot of the model, resulting in lower prediction reliability. The cross-modal consistency assessment unit compares whether the risk labels for the same business predictions from different single modalities converge; greater discrepancies between modal predictions indicate a worse performance. The robustness assessment unit observes the magnitude of change in prediction results after applying small adversarial perturbations to the input data; smaller changes indicate stronger noise tolerance and more stable predictions.

[0040] Understandably, weighted fusion involves aggregating the five categories of indicators according to preset weighting coefficients to generate an initial credibility score that comprehensively reflects the overall credibility level. The calibration sample set is a pre-prepared set of samples labeled with actual default results. The empirical default frequency is the proportion of actual defaults statistically analyzed within the calibration sample set according to the score interval. Probability mapping and calibration, by fitting the functional relationship between the initial credibility score and the empirical default frequency, calibrates the score to a risk credibility score reflecting the true probability of default.

[0041] In practical application, a company's risk control labeling result is "overdue." This result and intermediate representation data are input into a multidimensional risk credibility assessment model. The model uncertainty assessment unit outputs a model uncertainty index of 0.32 after multiple random forward propagations. The label consistency assessment unit calculates the label entropy of similar samples and outputs a label consistency index of 0.81. The out-of-distribution assessment unit calculates an out-of-distribution index of 0.15 based on the energy function. The cross-modal consistency assessment unit compares the predicted distributions of each modality and outputs a cross-modal consistency index of 0.88. The robustness assessment unit applies a perturbation and then observes the prediction offset, outputting a robustness index of 0.79. The five indicators are weighted and fused according to preset weights to generate an initial credibility score of 0.73. After calibration and probability mapping of the sample set, this score is converted into a risk credibility score of 0.71, approximating the true probability of default for this sample.

[0042] S5. Classify the risk control labeling results according to the risk credibility score, generate high-quality labeled training samples through manual review and correction, and incrementally train the supply chain finance risk control model based on the high-quality labeled training samples to obtain the trained supply chain finance risk control model.

[0043] The process of classifying the risk control labeling results based on the risk credibility score and generating high-quality labeled training samples through manual review and correction specifically includes: The risk credibility score is compared with a preset first credibility threshold and a second credibility threshold, wherein the first credibility threshold is greater than the second credibility threshold; If the risk credibility score is greater than or equal to the first credibility threshold, the risk control labeling result is marked as a high-credibility automatic confirmation sample. If the risk credibility score is greater than or equal to the second credibility threshold and less than the first credibility threshold, then the risk control labeling result is marked as a sample to be manually reviewed, the risk credibility scores are sorted in ascending order, and a manual review queue is generated. If the risk credibility score is less than the second credibility threshold, the risk control labeling result is marked as a high-risk priority processing sample and placed at the top of the manual review queue. Based on the highly reliable automatically confirmed samples, the samples awaiting manual review after verification, and the high-risk priority processing samples, the high-quality labeled training samples are generated and written into the supply chain finance risk control database.

[0044] The first credibility threshold is the upper limit for distinguishing between highly credible automatically confirmed samples and samples requiring manual review, while the second credibility threshold is the lower limit for distinguishing between samples requiring manual review and high-risk priority processing samples. Both are pre-set values ​​based on business tolerance. High-credibility automatically confirmed samples are risk control annotation results whose risk credibility scores reach the first credibility threshold. The model is highly confident in its risk judgment and can directly accept and store them in the database without manual intervention. Samples requiring manual review are risk control annotation results whose risk credibility scores fall between the first and second credibility thresholds. The model's judgment on these samples has some uncertainty and requires review and verification by risk control experts.

[0045] Next, high-risk priority processing samples are risk control annotation results with a risk confidence score below the second confidence threshold. These samples have extremely low model confidence and may pose a serious risk of misjudgment. Therefore, they are placed at the top of the manual review queue for priority processing. The manual review queue is a list of samples awaiting review, arranged in ascending order of risk confidence score; the lower the score, the higher the ranking, ensuring that the samples with the lowest confidence and highest risk receive priority for manual review. High-quality annotated training samples are an annotated dataset formed by merging the manually reviewed and corrected labels with high-confidence automatically confirmed samples. Their label accuracy is significantly higher than that of purely model-automatic annotation results.

[0046] By using the above method, the risk control labeling results are classified into high-confidence automatically confirmed samples, samples awaiting manual review, and high-risk priority processing samples according to the risk confidence score. This allows manual review resources to be accurately focused on samples with low model confidence, thereby improving labeling efficiency and label quality.

[0047] The incremental training of the supply chain finance risk control model based on the high-quality labeled training samples to obtain the trained supply chain finance risk control model specifically includes: Monitor the cumulative number of the high-quality labeled training samples in the supply chain finance risk control database, and trigger incremental training when the cumulative number reaches a preset incremental trigger threshold; Based on the high-quality labeled training samples and the historical labeled training samples in the supply chain finance risk control database, an incremental training sample set is constructed. Based on the incremental training sample set, the underlying network parameters used to extract basic semantic features in the multimodal encoder are frozen, the incremental training sample set is input into the supply chain finance risk control model, and the predicted risk label is output through the risk classifier in the supply chain finance risk control model. The parameters of the high-level network in the multimodal encoder and the parameters of the risk classifier are updated in multiple steps using the cross-entropy loss between the predicted risk label and the risk label corresponding to the high-quality labeled training sample. Using the difference between the high-level network parameters and the risk classifier parameters after the multi-step inner loop adaptive update and the parameters before incremental training as the meta gradient, the outer loop optimization update is performed on the high-level network parameters and the risk classifier parameters to generate the trained supply chain finance risk control model.

[0048] The incremental trigger threshold is a preset threshold for the number of accumulated samples. When the number of high-quality labeled training samples accumulated in the supply chain finance risk control database reaches this threshold, the incremental training process is automatically initiated to avoid excessive training frequency due to sporadic sample re-entry. The incremental training sample set consists of newly added high-quality labeled training samples combined with historical labeled training samples, ensuring that the model does not forget historical knowledge while learning new samples. The bottom-level network parameters are the parameters of the first few layers in the multimodal encoder responsible for extracting general semantic features. These parameters are frozen during incremental training to protect the model's learned cross-modal basic representation capabilities from being overwritten by new samples. The high-level network parameters are the network layer parameters near the output in the multimodal encoder, responsible for combining basic semantic features into abstract representations related to risk discrimination. These parameters, along with the parameters of the risk classifier, are the objects to be updated during incremental training.

[0049] Understandably, multi-step inner-loop adaptive update involves performing multiple gradient descent updates on the parameters of the high-level network and the risk classifier within a single incremental training task, simulating the model's rapid adaptation to new samples. The meta-gradient uses the difference between the updated parameters and the parameters before the update as an optimization direction signal, guiding the outer-loop optimization update to absorb risk patterns from new samples while preserving historical knowledge. The outer-loop optimization update, based on the meta-gradient, makes the final adjustment to the parameters of the high-level network and the risk classifier, enabling the model to balance old and new knowledge and continuously evolve, generating the trained supply chain finance risk control model.

[0050] By using the above methods, the training timing is controlled by incremental trigger thresholds, the underlying network parameters are frozen to protect the basic semantic representation, and the meta-learning strategy is used to perform inner-loop adaptive updates and outer-loop optimizations on the parameters of the high-level network and the risk classifier. This allows the model to continuously absorb high-quality labeled samples while avoiding catastrophic forgetting, thus achieving closed-loop self-evolution.

[0051] The outer-loop optimization update of the high-level network parameters and the risk classifier parameters is performed using the following calculation formula: In the formula, This represents the set of parameters for the high-level network and the risk classifier after the outer ring optimization and update; This represents the set of parameters for the high-level network and the risk classifier before incremental training. This represents the preset outer loop learning rate; This represents the number of meta-learning tasks divided from the incremental training sample set; This represents the set of high-level network parameters and risk classifier parameters after multi-step inner loop adaptive update in the i-th meta-learning task; Let represent the meta-gradient of the i-th meta-learning task; Based on the updated high-level network parameters after outer ring optimization and the parameter set of the risk classifier The parameters of the high-level network and the parameters of the risk classifier in the supply chain finance risk control model are updated to generate the trained supply chain finance risk control model.

[0052] Understandably, the outer-loop optimization update is the final adjustment operation for the parameters of the high-level network and the risk classifier during incremental training, executed after the adaptive learning of multiple inner-loop tasks is completed. The meta-learning task involves dividing the incremental training sample set into several training sub-tasks based on small-sample scenarios. Each sub-task simulates an independent fast learning process. The difference between the parameter set after the inner-loop adaptive update and the parameter set before incremental training constitutes the meta-gradient, reflecting the direction and magnitude of parameter adjustments made by that task. The outer-loop learning rate is a preset hyperparameter controlling the step size of the outer-loop optimization update. By averaging and weighting the meta-gradients generated by all meta-learning tasks, and subtracting this weighted average from the parameter set before incremental training, the parameter set after the outer-loop optimization update is obtained, ultimately updating the corresponding model parameters of the supply chain finance risk control model.

[0053] In practical applications, assuming the parameter value of a high-level network before incremental training is 0.50, and the outer loop learning rate is set to 0.10, three meta-learning tasks are divided from the incremental training sample set. After performing multi-step inner loop adaptive updates for the first meta-learning task, the parameter value becomes 0.55, and the meta-gradient is 0.05; after the inner loop update for the second task, the parameter value becomes 0.52, and the meta-gradient is 0.02; after the inner loop update for the third task, the parameter value becomes 0.58, and the meta-gradient is 0.08. The average meta-gradient of all meta-learning tasks is calculated to be 0.05, resulting in a parameter value of 0.495 after the outer loop optimization update. This value is then written back to the corresponding high-level network parameter value of the supply chain finance risk control model to complete the update, generating the trained supply chain finance risk control model.

[0054] like Figure 2 The diagram shown is a system block diagram of a supply chain finance risk control model training system based on multimodal data provided in an embodiment of the present invention. The system includes: The model pre-training module is used to build a supply chain finance risk control model and to conduct supervised training of the supply chain finance risk control model based on a pre-set supply chain finance risk control database. The multimodal representation module is used to acquire multimodal data of supply chain finance, and to extract features from the multimodal data of supply chain finance through the multimodal encoder in the supply chain finance risk control model to generate multimodal risk control representation vectors. The retrieval and conflict detection module is used to perform multi-strategy retrieval and multi-modal semantic conflict detection in the supply chain finance risk control database based on the multimodal risk control representation vector, and output the risk control labeling results. The credibility assessment module is used to construct a multi-dimensional risk credibility assessment model. The risk control labeling results are input into the multi-dimensional risk credibility assessment model, and multi-dimensional credibility indicators are output. The multi-dimensional credibility indicators are weighted, fused, and calibrated to generate a risk credibility score. The incremental optimization module is used to classify the risk control labeling results according to the risk credibility score, generate high-quality labeled training samples through manual review and correction, and perform incremental training on the supply chain finance risk control model based on the high-quality labeled training samples to obtain the trained supply chain finance risk control model.

[0055] Figure 2 The apparatus of the illustrated embodiment can be used to perform corresponding actions. Figure 1 The steps in the method embodiments shown are implemented in a similar manner and have similar technical effects, and will not be repeated here.

[0056] An electronic device includes a memory and a processor, wherein the memory stores a computer program, and when the processor runs the computer program stored in the memory, the processor executes the steps of the supply chain finance risk control model training method based on multimodal data as described in any of the above.

[0057] like Figure 3 The diagram shown is a hardware structure schematic of an electronic device according to an embodiment of the present invention. The electronic device 30 includes: a processor 31, a memory 32, and a computer program; wherein... The memory 32 is used to store the computer program, and the memory may also be flash memory. The computer program is, for example, an application program or functional module that implements the above method.

[0058] Processor 31 is configured to execute the computer program stored in the memory to implement the various steps performed by the device in the above method. For details, please refer to the relevant descriptions in the preceding method embodiments.

[0059] Alternatively, the memory 32 can be either standalone or integrated with the processor 31.

[0060] When the memory 32 is a device independent of the processor 31, the device may further include: Bus 33 is used to connect the memory 32 and the processor 31.

[0061] A readable storage medium storing a computer program, which, when executed by a processor, is used to implement the steps of the supply chain finance risk control model training method based on multimodal data as described in any of the above claims.

[0062] The readable storage medium can be a computer storage medium or a communication medium. A communication medium includes any medium that facilitates the transfer of computer programs from one location to another. A computer storage medium can be any available medium accessible to a general-purpose or special-purpose computer. For example, a readable storage medium is coupled to a processor, enabling the processor to read information from and write information to the readable storage medium. Of course, the readable storage medium can also be a component of the processor. The processor and the readable storage medium can reside in an Application-Specific Integrated Circuit (ASIC). Alternatively, the ASIC can be located in a user equipment. Of course, the processor and the readable storage medium can also exist as discrete components in a communication device. The readable storage medium can be a read-only memory (ROM), random access memory (RAM), CD-ROM, magnetic tape, floppy disk, and optical data storage device, etc.

[0063] The present invention also provides a program product including executable instructions stored in a readable storage medium. At least one processor of the device can read the executable instructions from the readable storage medium, and the at least one processor executes the executable instructions to cause the device to implement the methods provided in the various embodiments described above.

[0064] In the embodiments of the above-described device, it should be understood that the processor can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), etc. The general-purpose processor can be a microprocessor or any conventional processor. The steps of the method disclosed in this invention can be directly manifested as execution by a hardware processor, or execution by a combination of hardware and software modules within the processor.

[0065] Through the above embodiments, this invention constructs a supply chain finance risk control model and conducts supervised training on the model based on a pre-set supply chain finance risk control database; acquires multimodal supply chain finance data, extracts features from the multimodal data using a multimodal encoder in the supply chain finance risk control model, and generates a multimodal risk control representation vector; based on the multimodal risk control representation vector, performs multi-strategy retrieval and multimodal semantic conflict detection in the supply chain finance risk control database, and outputs risk control annotation results; and constructs a multidimensional risk credibility assessment model, inputting the risk control annotation results into the multidimensional risk credibility assessment model. The evaluation model outputs multi-dimensional credibility indicators, which are then weighted, fused, and calibrated to generate a risk credibility score. Based on this score, the risk control labeling results are categorized, and high-quality labeled training samples are generated through manual review and correction. These high-quality labeled training samples are then used for incremental training of the supply chain finance risk control model, resulting in a trained model. This enables unified representation of multi-modal data and efficient reuse of historical samples, constructing a full-link, closed-loop, and self-evolving risk control model training system that significantly improves the high credibility of risk identification and cross-scenario adaptability.

[0066] This invention constructs a supply chain finance risk control model and conducts supervised training based on a pre-set supply chain finance risk control database. It utilizes a multimodal encoder to uniformly represent multi-source heterogeneous data, generating a multimodal risk control representation vector that integrates enterprise metadata. This effectively breaks down data silos and enables in-depth mining of cross-modal risk features. By performing multi-strategy retrieval in the supply chain finance risk control database and combining it with a multimodal semantic conflict detection mechanism to output risk control annotation results, the reuse efficiency and annotation consistency of historical annotated training samples are significantly improved, greatly reducing manual annotation costs. Furthermore, by constructing a multi-dimensional risk credibility assessment model that includes model uncertainty, label consistency, out-of-distribution degree, cross-modal consistency, and robustness against interference, the model performs weighted fusion and confidence calibration of multi-dimensional credibility indicators to generate a risk credibility score mapped to the frequency of experienced defaults. This effectively intercepts high-confidence erroneous judgments and reduces the default false alarm rate. Based on the risk credibility score, the samples are classified into high-credibility automatically confirmed samples, samples awaiting manual review, and high-risk priority processing samples. Relying on the human-machine collaboration mechanism, accurate review and label correction are achieved, generating high-quality labeled training samples. The model incremental training is triggered through closed-loop knowledge feedback, and the model parameters are optimized and updated in the inner and outer loops using a meta-learning strategy, so that the supply chain finance risk control model can continuously evolve in long-term operation and the risk identification accuracy can be continuously improved.

[0067] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for training a supply chain finance risk control model based on multimodal data, characterized in that, The method includes: A supply chain finance risk control model is constructed, and the model is trained under supervision based on a pre-set supply chain finance risk control database. Acquire multimodal data of supply chain finance, and extract features from the multimodal encoder in the supply chain finance risk control model to generate a multimodal risk control representation vector; Based on the multimodal risk control representation vector, multi-strategy retrieval and multimodal semantic conflict detection are performed in the supply chain finance risk control database, and risk control annotation results are output. A multidimensional risk credibility assessment model is constructed. The risk control labeling results are input into the multidimensional risk credibility assessment model, and multidimensional credibility indicators are output. The multidimensional credibility indicators are weighted, fused, and calibrated to generate a risk credibility score. The risk control labeling results are classified according to the risk credibility score, and high-quality labeled training samples are generated through manual review and correction. The supply chain finance risk control model is incrementally trained based on the high-quality labeled training samples to obtain the trained supply chain finance risk control model.

2. The method for training a supply chain finance risk control model based on multimodal data according to claim 1, characterized in that, The construction of the supply chain finance risk control model, based on a pre-set supply chain finance risk control database, includes supervised training of the model, specifically including: Obtain historical multimodal data samples and risk labels corresponding to the historical multimodal data samples from the supply chain finance risk control database; The historical multimodal data samples are input into the supply chain finance risk control model, and the features of the historical multimodal data samples are extracted by the multimodal encoder to generate a historical multimodal risk control representation vector. The historical multimodal risk control representation vector is mapped to an initial predicted risk label using the risk classifier in the supply chain finance risk control model. Based on the cross-entropy loss function, the loss value between the initial predicted risk label and the risk label corresponding to the historical multimodal data sample is calculated, and the model parameters of the multimodal encoder and the risk classifier are iteratively updated according to the loss value until the preset convergence condition is met, so as to obtain the pre-trained supply chain finance risk control model.

3. The method for training a supply chain finance risk control model based on multimodal data according to claim 1, characterized in that, The acquisition of multimodal data in supply chain finance, and the extraction of features from the multimodal encoder in the supply chain finance risk control model to generate a multimodal risk control representation vector, specifically includes: Obtain single-modal data of supply chain finance from the multimodal data of supply chain finance; Based on the data type of the single-modal supply chain finance data, the corresponding single-modal encoder is called from the multimodal encoder, and the single-modal encoder is used to extract features from the single-modal supply chain finance data to generate a single-modal supply chain finance modal feature map. Based on the cross-modal contrastive learning mechanism, positive and negative sample pairs of the feature maps of the supply chain finance modality in different modalities are constructed. With the constraint of aligning the semantic features of the positive sample pairs and dispersing the semantic features of the negative sample pairs, the single modality encoder is iteratively optimized so that the feature maps of the supply chain finance modality are mapped to a shared semantic space and fused to generate a basic unified vector. A preset enterprise meta-information vector is obtained, and an attention mechanism is used to calculate the dimensional modulation weight of the enterprise meta-information vector on the basic unified vector. The enterprise meta-information vector is then embedded into the basic unified vector according to the dimensional modulation weight to generate the multimodal risk control representation vector.

4. The method for training a supply chain finance risk control model based on multimodal data according to claim 1, characterized in that, Based on the multimodal risk control representation vector, a multi-strategy retrieval is performed in the supply chain finance risk control database to generate preliminary risk labels and preliminary default probabilities, specifically including: Based on the multimodal risk control representation vector, the enterprise meta-information vector is matched layer by layer in the supply chain finance risk control database to select candidate samples that meet the preset layer matching conditions and generate a candidate sample set. In the candidate sample set, a similarity algorithm is used to perform precise similarity retrieval on the multimodal risk control representation vector to generate a first similarity value, and risk labels corresponding to candidate samples that meet the preset first similarity conditions are selected to generate a first risk label set; In the candidate sample set, based on the transaction pattern and payment period structure, a weighted similarity algorithm is used to perform cross-scenario generalization retrieval on the multimodal risk control representation vector to generate a second similarity value, and to select the risk labels corresponding to the candidate samples that meet the preset second similarity conditions to generate a second risk label set; Integrate the first risk tag set and the second risk tag set, and remove duplicates according to the preset risk tag categories to generate a candidate risk tag set. Normalize the similarity values ​​in the candidate risk tag set to generate the voting weight of the risk tag. Based on the risk label category, the voting weights of the risk labels in the same category are accumulated and aggregated to generate a risk label weighted probability. The risk label with the highest risk label weighted probability is taken as the initial risk label, and the risk label weighted probability is taken as the initial default probability.

5. The method for training a supply chain finance risk control model based on multimodal data according to claim 4, characterized in that, Perform multimodal semantic conflict detection on the preliminary risk labels, specifically including: Obtain the supply chain finance modal feature map, input the supply chain finance modal feature map into a preset modal risk classifier, and output the modal prediction probability vector; Based on the modality prediction probability vector, the semantic conflict degree C of the preliminary risk label is calculated, and the corresponding calculation formula is as follows: In the formula, M represents the total number of modalities detected in multimodal semantic conflict detection; K represents the total number of categories of the initial risk label; This represents the probability value of the initial risk label of the k-th class in the modality prediction probability vector of the m-th modality; This represents the average probability that all M modalities are predicted to be the initial risk label of the kth class; If the semantic conflict degree is less than the preset conflict threshold, then the preliminary risk label is determined to have passed the multimodal semantic conflict detection, and the preliminary risk label is retained. If the semantic conflict degree is greater than or equal to the preset conflict threshold, it is determined that the preliminary risk label has failed the multimodal semantic conflict detection. The parameters of the multi-strategy retrieval process are adjusted to iteratively execute the multi-strategy retrieval and multimodal semantic conflict detection until the number of iterations reaches the preset maximum number of iterations. The preliminary risk label with the lowest semantic conflict degree is retained. The preliminary risk label and the preliminary default probability are used as the risk control labeling results.

6. The method for training a supply chain finance risk control model based on multimodal data according to claim 1, characterized in that, The construction of a multidimensional risk credibility assessment model involves inputting the risk control labeling results into the model, outputting multidimensional credibility indicators, and performing weighted fusion and confidence calibration on these indicators to generate a risk credibility score. Specifically, this includes: The multidimensional risk credibility assessment model is constructed, wherein the multidimensional risk credibility assessment model includes a model uncertainty assessment unit, a label consistency assessment unit, an out-of-distribution assessment unit, a cross-modal consistency assessment unit, and an anti-interference robustness assessment unit; The risk control labeling results and intermediate representation data generated during the risk control labeling process are input into the multidimensional risk credibility assessment model. The assessment unit calculates and outputs the model uncertainty index, label consistency index, out-of-distribution index, cross-modal consistency index, and anti-interference robustness index, respectively. The model uncertainty index, the label consistency index, the out-of-distribution index, the cross-modal consistency index, and the anti-interference robustness index are weighted and fused to generate an initial credibility score; Based on a preset calibration sample set, with the goal of minimizing the error between the initial credibility score and the experienced default frequency in the calibration sample set, the initial credibility score is probabilistically mapped and calibrated, mapping the initial credibility score to the numerical value of the experienced default frequency, thereby generating the risk credibility score.

7. The method for training a supply chain finance risk control model based on multimodal data according to claim 1, characterized in that, The process of classifying the risk control labeling results based on the risk credibility score and generating high-quality labeled training samples through manual review and correction specifically includes: The risk credibility score is compared with a preset first credibility threshold and a second credibility threshold, wherein the first credibility threshold is greater than the second credibility threshold; If the risk credibility score is greater than or equal to the first credibility threshold, the risk control labeling result is marked as a high-credibility automatic confirmation sample. If the risk credibility score is greater than or equal to the second credibility threshold and less than the first credibility threshold, then the risk control labeling result is marked as a sample to be manually reviewed, the risk credibility scores are sorted in ascending order, and a manual review queue is generated. If the risk credibility score is less than the second credibility threshold, the risk control labeling result is marked as a high-risk priority processing sample and placed at the top of the manual review queue; Based on the highly reliable automatically confirmed samples, the samples awaiting manual review after verification, and the high-risk priority processing samples, the high-quality labeled training samples are generated and written into the supply chain finance risk control database.

8. The method for training a supply chain finance risk control model based on multimodal data according to claim 1, characterized in that, The incremental training of the supply chain finance risk control model based on the high-quality labeled training samples to obtain the trained supply chain finance risk control model specifically includes: Monitor the cumulative number of the high-quality labeled training samples in the supply chain finance risk control database, and trigger incremental training when the cumulative number reaches a preset incremental trigger threshold; Based on the high-quality labeled training samples and the historical labeled training samples in the supply chain finance risk control database, an incremental training sample set is constructed. Based on the incremental training sample set, the underlying network parameters used to extract basic semantic features in the multimodal encoder are frozen, the incremental training sample set is input into the supply chain finance risk control model, and the predicted risk label is output through the risk classifier in the supply chain finance risk control model. The parameters of the high-level network in the multimodal encoder and the parameters of the risk classifier are updated in multiple steps using the cross-entropy loss between the predicted risk label and the risk label corresponding to the high-quality labeled training sample. Using the difference between the high-level network parameters and the risk classifier parameters after the multi-step inner loop adaptive update and the parameters before incremental training as the meta gradient, the outer loop optimization update is performed on the high-level network parameters and the risk classifier parameters to generate the trained supply chain finance risk control model.

9. The method for training a supply chain finance risk control model based on multimodal data according to claim 8, characterized in that, The outer-loop optimization update of the high-level network parameters and the risk classifier parameters is performed using the following calculation formula: In the formula, This represents the set of parameters for the high-level network and the risk classifier after the outer ring optimization and update; This represents the set of parameters for the high-level network and the risk classifier before incremental training. This represents the preset outer loop learning rate; This represents the number of meta-learning tasks divided from the incremental training sample set; This represents the set of high-level network parameters and risk classifier parameters after multi-step inner loop adaptive update in the i-th meta-learning task; Let represent the meta-gradient of the i-th meta-learning task; Based on the updated high-level network parameters after outer ring optimization and the parameter set of the risk classifier The parameters of the high-level network and the parameters of the risk classifier in the supply chain finance risk control model are updated to generate the trained supply chain finance risk control model.

10. A supply chain finance risk control model training system based on multimodal data, characterized in that, The system, applied to the supply chain finance risk control model training method based on multimodal data as described in any one of claims 1-9, comprises: The model pre-training module is used to build a supply chain finance risk control model and to conduct supervised training of the supply chain finance risk control model based on a pre-set supply chain finance risk control database. The multimodal representation module is used to acquire multimodal data of supply chain finance, and to extract features from the multimodal data of supply chain finance through the multimodal encoder in the supply chain finance risk control model to generate multimodal risk control representation vectors. The retrieval and conflict detection module is used to perform multi-strategy retrieval and multi-modal semantic conflict detection in the supply chain finance risk control database based on the multimodal risk control representation vector, and output the risk control labeling results. The credibility assessment module is used to construct a multi-dimensional risk credibility assessment model. The risk control labeling results are input into the multi-dimensional risk credibility assessment model, and multi-dimensional credibility indicators are output. The multi-dimensional credibility indicators are weighted, fused, and calibrated to generate a risk credibility score. The incremental optimization module is used to classify the risk control labeling results according to the risk credibility score, generate high-quality labeled training samples through manual review and correction, and perform incremental training on the supply chain finance risk control model based on the high-quality labeled training samples to obtain the trained supply chain finance risk control model.

Images