Factoring service hierarchical access method and device

Abstract

According to the factoring service hierarchical access method and device provided by the invention, multi-source heterogeneous data of a target enterprise is collected through a standardized interface, dimensions such as financial indexes, system performance, public opinion dynamics and compliance texts are covered, and data quality and safety are ensured through a professional data preprocessing process. A risk grading model of a fusion framework is adopted: text semantic features are extracted by utilizing a BERT model subjected to field fine tuning, structured data are analyzed in combination with the numerical understanding capability of a GPT series model, multi-source feature alignment is realized through a cross-modal attention mechanism, a comprehensive feature vector is generated based on a multi-head attention network, and the risk grading model of the fusion framework is obtained. And finally, outputting a risk level by the lightweight full-connection network, and forming a corresponding access strategy based on the risk level. According to the method, actual measurement-free evaluation is realized, a traditional test period needing a plurality of weeks is compressed to a plurality of hours, the model adaptability is continuously optimized through an online learning mechanism, multi-dimensional risk evaluation is realized, and test resource consumption and time cost are greatly reduced.

Description

Technical Field

[0001] This invention relates to the field of factoring business management technology, and in particular to a method and apparatus for hierarchical access to factoring business. Background Technology

[0002] In supply chain finance platforms, performance evaluation of newly connected enterprises or subsystems is a crucial step in ensuring the overall service quality and operational security of the platform. Existing technical solutions primarily employ two main approaches: First, relying on manual review and on-site inspections, where professionals conduct offline assessments of enterprise qualifications and system architecture. This method is not only inefficient and costly, but also significantly influenced by subjective factors, making it difficult to scale. Second, using automated testing methods based on rule engines. These methods match and judge structured data using preset static rules (such as financial indicator thresholds and interface response time standards). While this method improves efficiency to some extent, its rules lack flexibility and cannot effectively handle the risk and performance information contained in unstructured data such as enterprise sentiment, system logs, and contract texts, resulting in a single evaluation dimension and incomplete coverage. Furthermore, the updates and iterations of existing rule engines rely on manual intervention, making it difficult to adapt in real-time to rapidly changing market environments and complex and diverse business scenarios (such as cross-border factoring and intangible asset factoring), easily leading to evaluation lags and misjudgments. Importantly, existing solutions typically require setting up simulated testing environments and constructing large amounts of test data for stress testing that can last for weeks, resulting in inherent drawbacks such as high resource consumption and long evaluation cycles. Therefore, current technologies have not yet provided an efficient, accurate, comprehensive, and intelligent access performance evaluation solution that requires no actual testing, failing to meet the urgent needs of modern supply chain finance platforms for efficient access and dynamic risk management. Summary of the Invention

[0003] In view of this, embodiments of the present invention provide a tiered access method and apparatus for factoring business, so as to eliminate or improve one or more defects existing in the prior art and solve the problems of low efficiency, high cost and narrow evaluation dimensions in the enterprise accessibility test of supply chain finance platforms.

[0004] One aspect of the present invention provides a tiered access method for factoring business, the method comprising the following steps: For target enterprises, multi-source heterogeneous business data is obtained based on standardized interfaces. The multi-source heterogeneous business data includes structured financial and business data, system technical performance data, unstructured enterprise public opinion data, operational status data and business compliance data, semi-structured interface logs and industry policy document data. The multi-source heterogeneous business data is cleaned, standardized, vectorized, and anonymized. The risk grading model is pre-trained using the multi-source heterogeneous business data as input to output risk level evaluation results. The risk grading model includes a feature extraction unit, a feature fusion unit, and a grading unit. The feature extraction unit uses a large model based on BERT to extract text features from textual data in the multi-source heterogeneous business data, and uses a large model based on GPT to extract numerical features from numerical data in the multi-source heterogeneous business data. It also aligns the textual and numerical features based on a cross-modal attention mechanism. The feature fusion unit constructs a feature fusion network based on a multi-head attention mechanism to extract a comprehensive feature vector from the textual and numerical features. The grading unit is a fully connected neural network that takes the comprehensive feature vector as input and outputs a probability distribution for multiple risk levels as the risk level evaluation result. Based on the risk level assessment results, a factoring business access plan is generated and executed for the target enterprise according to preset rules. The factoring business access plan includes whether access is allowed and factoring business is provided, rectification requirements, and the business scope configured according to the risk level.

[0005] In some embodiments, obtaining multi-source heterogeneous business data for a target enterprise object based on a standardized interface further includes: A real-time data acquisition channel was built using the Flume system and Kafka platform, and the data formats of multiple data sources were compatible based on the Schema Registry service.

[0006] In some embodiments, the multi-source heterogeneous business data is cleaned, standardized, vectorized, and anonymized, including: Optical character recognition is used to extract text from the image data in the multi-source heterogeneous business data to obtain the corresponding text data. The Spark SQL module is used to fill missing text data in the multi-source heterogeneous business data with default identifiers and to filter abnormal text based on preset regular expressions; the missing numerical data in the multi-source heterogeneous business data is filled with the median and numerical anomalies are identified and removed using the IQR rule. The text data is standardized using UTF-8 encoding, and the numerical data is standardized using Z-Score. The text data is segmented using WordPiece to generate an index sequence, and the numerical data is converted into tensor format to complete vectorization. The sensitive corporate information can be added with Laplace noise using privacy-checking methods, or the sensitive corporate information can be replaced based on regular expressions.

[0007] In some embodiments, the pre-training step of the risk grading model includes: Obtain a training sample set containing multiple samples. Each sample contains multi-source heterogeneous business sample data collected for the sample enterprise object, and add the corresponding risk level as a label. The initial risk grading model, comprising a feature extraction unit, a feature fusion unit, and a grading unit, is pre-trained using the training sample set. The feature extraction unit uses a large model based on BERT to extract textual features from textual data in the multi-source heterogeneous business sample data, and a large model based on GPT to extract numerical features from numerical data in the same data. It also aligns the textual and numerical features using a cross-modal attention mechanism. The feature fusion unit constructs a feature fusion network based on a multi-head attention mechanism to extract a comprehensive feature vector from the textual and numerical features. The grading unit is a fully connected neural network that takes the comprehensive feature vector as input and outputs a predicted risk level. It updates the parameters of the initial risk grading model based on the deviation between the predicted risk level and the label, thus obtaining the risk grading model.

[0008] In some embodiments, the hierarchical unit employs a fully connected neural network, including an input layer, a hidden layer, and an output layer. The hidden layer contains at least three fully connected layers, and each fully connected layer has a ReLU activation function layer and a Dropout layer. The output layer employs a Softmax activation function layer.

[0009] In some embodiments, the method further includes: setting a dynamic iteration unit to collect the multi-source heterogeneous business data from newly connected enterprise objects to construct an incremental dataset; and performing incremental learning when the incremental dataset reaches a preset number of records. A threshold optimization unit is set up to iterate the risk level threshold based on the industry average level, so as to adjust the risk level of newly connected enterprise objects during the incremental learning process.

[0010] In some embodiments, the method further includes: setting up an interpretability analysis unit and using the SHAP algorithm to generate a feature contribution heatmap to demonstrate the impact of each feature on the risk level assessment results.

[0011] On the other hand, the present invention also provides a factoring business tiered access device, including a processor, a memory, and a computer program or instructions stored in the memory. The processor is used to execute the computer program or instructions, and when the computer program or instructions are executed, the device implements the steps of the above method.

[0012] On the other hand, the present invention also provides a computer-readable storage medium having a computer program or instructions stored thereon, which, when executed by a processor, implement the steps of the above-described method.

[0013] On the other hand, the present invention also provides a computer program product, including a computer program or instructions that, when executed by a processor, implement the steps of the above-described method.

[0014] The factoring business tiered access method and device described in this invention collects multi-source heterogeneous data from target enterprises through standardized interfaces, covering dimensions such as financial indicators, system performance, public opinion dynamics, and compliance texts. A professional data preprocessing process ensures data quality and security. A risk tiering model with a fusion architecture is employed: a domain-adjusted BERT model extracts semantic features from the text, combined with the numerical understanding capabilities of the GPT series models to parse structured data, a cross-modal attention mechanism aligns multi-source features, a multi-head attention network generates a comprehensive feature vector, and finally, a lightweight fully connected network outputs the risk level, forming a corresponding access strategy based on the risk level. This invention achieves test-free evaluation, compressing the traditional testing cycle of several weeks to several hours. Through an online learning mechanism, the model's adaptability is continuously optimized, enabling multi-dimensional risk assessment and significantly reducing testing resource consumption and time costs.

[0015] Furthermore, this invention improves the generalization capability for complex factoring businesses by dynamically adjusting risk levels and combining incremental updates to accurately identify risks in target enterprises. By leveraging SHAP interpretability technology to identify the contribution of each feature, risk factors can be precisely located.

[0016] Additional advantages, objects, and features of the invention will be set forth in part in the description which follows, and will also become apparent in part to those skilled in the art upon studying the text, or may be learned by practice of the invention. The objects and other advantages of the invention can be realized and obtained by means of the structures specifically pointed out in the description and drawings.

[0017] Those skilled in the art will understand that the objectives and advantages achievable with the present invention are not limited to those specifically described above, and that the above and other objectives achievable with the present invention will become clearer from the following detailed description. Attached Figure Description

[0018] The accompanying drawings, which are included to provide a further understanding of the invention and form part of this application, are not intended to limit the scope of the invention. In the drawings: Figure 1 This is a flowchart illustrating the factoring business hierarchical access method according to an embodiment of the present invention.

[0019] Figure 2 This is an architecture diagram of a large model-driven, test-free, intelligent hierarchical access performance testing system according to an embodiment of the present invention.

[0020] Figure 3 This is an access flowchart of the large model-driven, test-free, intelligent hierarchical access performance testing system according to an embodiment of the present invention. Detailed Implementation

[0021] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the embodiments and accompanying drawings. Here, the illustrative embodiments and descriptions of this invention are used to explain the invention, but are not intended to limit the invention.

[0022] It should also be noted that, in order to avoid obscuring the invention with unnecessary details, only the structures and / or processing steps closely related to the solution according to the invention are shown in the accompanying drawings, while other details that are not closely related to the invention are omitted.

[0023] It should be emphasized that the term "including / comprises" as used herein refers to the presence of a feature, element, step, or component, but does not exclude the presence or addition of one or more other features, elements, steps, or components.

[0024] Supply chain finance platforms are a core infrastructure for addressing the difficulties and high costs of financing for SMEs. Their efficient operation relies on rapid and accurate performance evaluation of newly connected enterprises / subsystems. Traditional access performance testing methods mainly fall into two categories: one is manual review plus on-site inspection, which relies on professionals to conduct offline assessments of enterprise qualifications and system architecture, resulting in low efficiency, high costs, and strong subjectivity; the other is automated testing based on rule engines, which uses preset static rules (such as debt-to-equity ratio thresholds and interface response time standards) to judge structured data, but it is difficult to adapt to the complex and ever-changing supply chain environment.

[0025] With the development of artificial intelligence technology, large-scale models have demonstrated powerful advantages in fields such as natural language processing, multi-source data fusion, and deep pattern recognition, providing a new technical path for performance evaluation of supply chain finance access. However, existing technologies have not fully leveraged the core capabilities of large-scale models and lack end-to-end solutions for "test-free" scenarios. They are unable to effectively handle unstructured data such as news and public opinion, contract texts, and system logs, and also struggle to achieve real-time iteration and dynamic adaptation of evaluation models. This results in biased and delayed evaluation results, failing to meet the dual requirements of efficient platform access and risk management.

[0026] In view of this, the present invention provides a tiered access method for factoring business, such as... Figure 1 As shown, the method includes the following steps S101~S104: Step S101: Obtain multi-source heterogeneous business data for the target enterprise based on standardized interfaces. Multi-source heterogeneous business data includes structured financial and business data, system technical performance data, unstructured enterprise public opinion data, operational status data and business compliance data, and semi-structured interface logs and industry policy documents.

[0027] Step S102: Perform data cleaning, standardization, vectorization, and data anonymization on multi-source heterogeneous business data.

[0028] Step S103: Input the pre-trained risk grading model from the multi-source heterogeneous business data to output the risk level evaluation result; the risk grading model includes a feature extraction unit, a feature fusion unit, and a grading unit. The feature extraction unit uses a large model based on BERT to extract text features from textual data in the multi-source heterogeneous business data, and uses a large model based on GPT to extract numerical features from numerical data in the multi-source heterogeneous business data, and aligns the text features and numerical features based on a cross-modal attention mechanism; the feature fusion unit constructs a feature fusion network based on a multi-head attention mechanism to extract a comprehensive feature vector from the text features and numerical features; the grading unit is a fully connected neural network that takes the comprehensive feature vector as input and outputs the probability distribution for multiple risk levels as the risk level evaluation result.

[0029] Step S104: Based on the risk level assessment results, generate and execute a factoring business access plan for the target enterprise according to the preset rules. The factoring business access plan includes whether access is allowed and factoring business is provided, rectification requirements, and the business scope configured according to the risk level.

[0030] In step S101, in order to build the data foundation for the test-free assessment, a real-time data collection system that can seamlessly connect with various data sources in the supply chain finance ecosystem is constructed.

[0031] Specifically, the standardized interface is not a single form, but a collection of interfaces including RESTful APIs and SDKs (supporting Python / Java / Go), designed to be compatible with various environments from public clouds and private clouds to on-premises deployment systems. The collected multi-source heterogeneous business data is systematically divided into three categories: The first category is structured data, such as specific indicators in corporate financial statements (debt-to-equity ratio, profit margin), business records (order volume, frequency of financing applications), and system technical performance data (API interface response time, TPS, server CPU / memory utilization); the second category is unstructured data, which is a blind spot that traditional methods struggle to handle, including corporate sentiment data crawled from news and social media, log text containing exception stack information generated during system runtime, and business compliance documents such as contracts and vouchers; the third category is semi-structured data, such as API call logs in JSON / XML format and industry policy documents in HTML / PDF format. In terms of technical implementation, a real-time data acquisition channel is built using the Flume system and Kafka platform, and the data formats of multiple data sources are compatible based on the Schema Registry service to support the real-time access of tens of thousands of data entries per second, ensuring that the original information required for the assessment is comprehensive and timely.

[0032] In step S102, the multi-source heterogeneous business data undergoes data cleaning, standardization, vectorization, and data anonymization, including steps S1021~S1025: Step S1021: Use optical character recognition to extract text from the image data in the multi-source heterogeneous business data to obtain the corresponding text data.

[0033] Step S1022: Based on the Spark SQL module, missing text data in multi-source heterogeneous business data is filled with default identifiers and abnormal text is filtered based on preset regular expressions; missing numerical data in multi-source heterogeneous business data is filled with the median and numerical anomalies are identified and removed using the IQR rule.

[0034] Step S1023: Standardize text data using UTF-8 encoding and numerical data using Z-Score standardization.

[0035] Step S1024: After word segmentation of the text data using WordPiece, generate an index sequence, convert the numerical data into tensor format, and complete the vectorization.

[0036] Step S1025: Add Laplace noise to the sensitive corporate information using privacy-checking methods, or replace the sensitive corporate information based on regular expressions.

[0037] First, the data cleaning phase utilizes distributed computing frameworks such as Spark SQL to handle missing and outlier values: for numerical data, such as sudden TPS spikes, the IQR interquartile range method is used for identification and correction; for textual data, such as disorganized logs, irrelevant characters and spam are filtered out using pre-defined regular expression rules. Next, the standardization phase aims to eliminate the influence of units of measurement. Numerical data, such as response time, is Z-score standardized to conform to a distribution with a mean of 0 and a standard deviation of 1; textual data is uniformly encoded using UTF-8. Then, the vectorization phase transforms the data into a model-readable numerical form: numerical data is directly converted into tensors; textual data is processed by word segmenters such as WordPiece to form an index sequence, ready for use in large models. Finally, the crucial data anonymization step uses differential privacy technology with added Laplace noise to process sensitive information (such as legal person ID cards and bank account numbers), and regular expressions are used to de-identify and replace phone numbers and email addresses, ensuring data security and compliance in subsequent analysis.

[0038] In step S103, the risk grading model comprises three collaboratively working units: First, the feature extraction unit does not simply use a general-purpose large model, but rather adopts a "pre-training + fine-tuning" strategy. Specifically, for textual data (public opinion, logs, contracts), a BERT-base model finely tuned on a massive corpus of over 500,000 supply chain finance data entries is used for deep semantic understanding, outputting a 768-dimensional text feature vector containing contextual information, enabling it to accurately capture key risk signals such as overdue payments and defaults. For numerical data (financial indicators, performance monitoring), the powerful numerical understanding and contextual association capabilities of the GPT series models are utilized to transform discrete numerical values ​​such as "daily average order volume of 1 million" and "P99 response time of 500ms" into semantically related 512-dimensional numerical feature vectors. Subsequently, a cross-modal attention mechanism is used to semantically align the textual and numerical features, establishing connections between information from different modalities.

[0039] Second, the feature fusion unit constructs a network based on a multi-head attention mechanism, dynamically calculates the importance weights of different feature sources (for example, assigning higher weights to abnormal log features than to regular transaction data), and stabilizes the training process through residual connections and layer normalization techniques, ultimately generating a 1024-dimensional comprehensive feature vector.

[0040] Third, the hierarchical unit employs a fully connected neural network, taking the comprehensive feature vector as input. It prevents overfitting through hidden layers including Dropout layers, and finally, the Softmax output layer provides the risk level, such as the probability distribution of four performance levels: S / A / B / C. To balance accuracy and efficiency, this module can introduce knowledge distillation technology, transferring knowledge from a large model to a lightweight classifier. While maintaining an accuracy of ≥92%, it compresses the inference time for a single data point to less than 100 milliseconds. Furthermore, an online learning mechanism based on incremental learning is designed, enabling the model to continuously and adaptively optimize.

[0041] In some embodiments, the pre-training step of the risk grading model includes steps S1031-S1032: Step S1031: Obtain a training sample set containing multiple samples. Each sample contains multi-source heterogeneous business sample data collected for the sample enterprise object, and adds the corresponding risk level as a label.

[0042] Step S1032: The initial risk grading model, which includes a feature extraction unit, a feature fusion unit, and a grading unit, is pre-trained using a training sample set. The feature extraction unit uses a large model based on BERT to extract sample text features from textual data in multi-source heterogeneous business sample data, and uses a large model based on GPT to extract sample numerical features from numerical data in multi-source heterogeneous business sample data. The sample text features and sample numerical features are aligned based on a cross-modal attention mechanism. The feature fusion unit constructs a feature fusion network based on a multi-head attention mechanism to extract a comprehensive feature vector from the sample text features and sample numerical features. The grading unit is a fully connected neural network that takes the comprehensive feature vector as input and outputs a risk level prediction value. The loss is established based on the deviation between the risk level prediction value and the label to update the parameters of the initial risk grading model to obtain the risk grading model.

[0043] Steps S1031 to S1032 together constitute the supervised pre-training process of the risk grading model. Its core lies in using labeled historical data to construct an initial model that can accurately map multi-source heterogeneous data to risk levels. Specifically, in step S1031, a high-quality training sample set needs to be constructed. Each sample contains multi-source heterogeneous business data from the sample companies, corresponding to that in step S101, and is critically labeled with a risk level tag, such as S / A / B / C, which serves as a true benchmark and is determined by domain experts or verified through long-term business performance. Step S1032 then trains the initial risk grading model end-to-end based on this dataset: First, a large model based on BERT extracts deep textual features from textual sample data. Simultaneously, a large model based on GPT processes numerical sample data to generate numerical features, and a cross-modal attention mechanism is used to align and interact the two features in the semantic space. Subsequently, the feature fusion unit dynamically weights and fuses these aligned features through its multi-head attention network to generate a comprehensive feature vector containing multi-source information. Finally, the grading unit predicts the risk level based on this vector and iteratively updates the entire model parameters from feature extraction to classification decision using the backpropagation algorithm by calculating loss functions such as cross-entropy loss between the predicted value and the true label, until the model converges, thus obtaining a pre-trained risk grading model with accurate risk discrimination capabilities.

[0044] In some embodiments, the hierarchical unit employs a fully connected neural network, including an input layer, a hidden layer, and an output layer. The hidden layer contains at least three fully connected layers, each of which has a ReLU activation function layer and a Dropout layer. The output layer employs a Softmax activation function layer.

[0045] In some embodiments, the method further includes: setting up a dynamic iteration unit to collect multi-source heterogeneous business data from newly connected enterprise objects to construct an incremental dataset; and performing incremental learning when the incremental dataset reaches a preset number of records.

[0046] The dynamic iterative unit continuously collects multi-source heterogeneous business data from newly joined enterprises, constructing a real-time growing incremental dataset. Its key technology is not simply data accumulation, but rather the automatic triggering of a model fine-tuning process based on incremental learning algorithms (such as the Ftrl optimizer) when the incremental data reaches a preset scale (e.g., every 1000 new valid samples). This process updates the parameters of the pre-trained risk grading model only based on the new data, without initiating a costly full retraining. This allows the model to quickly absorb newly emerging risk patterns and business characteristics in the market with low resource overhead, achieving online evolution and real-time optimization of the model's recognition capabilities.

[0047] A threshold optimization unit is set up to iterate the risk level classification thresholds based on industry averages, adjusting the risk level of newly added enterprise objects during incremental learning. Calibration is performed at the macro-standard level, using continuously updated industry benchmark data, such as the average performance level of similar enterprises, and optimization strategies such as genetic algorithms to dynamically adjust the specific thresholds for each risk level (S, A, B, C). This mechanism ensures that the classification standards are not static but can flexibly fluctuate with changes in industry development and platform business strategies, thus maintaining the relative objectivity and fairness of the assessment results. Ultimately, the dynamic iteration unit and the threshold optimization unit work together; the former optimizes the model's discriminative ability based on micro-data characteristics, while the latter calibrates the judgment scale based on macro-classification standards. Together, they form a complete, self-optimizing intelligent assessment ecosystem, significantly improving the system's accuracy, robustness, and business adaptability in long-term operation.

[0048] Step S104 transforms the intelligent analysis results into specific, executable business instructions, forming a decision-making closed loop. The preset rules are a set of clearly defined and configurable business strategy libraries. Specifically: If the evaluation result is S-level or A-level, the system will automatically generate an "Allow Access" instruction and configure the highest priority business scope (such as full-category factoring business, optimal financing amount, and green approval channel) for S-level enterprises, and the standard business scope for A-level enterprises. If the result is B-level, the system will trigger a "Conditional Access" process. On the one hand, it will generate restrictive business plans, such as only opening basic factoring business and setting a low limit. On the other hand, based on the interpretability analysis of the aforementioned model, such as the feature contribution heatmap output by the SHAP algorithm, it will automatically generate a structured "Rectification Requirement Notice," clearly pointing out performance shortcomings or risk points, such as "Due to the existence of a 'database connection pool exhausted' record in the system log, please expand to 2000 connections." After rectification, the enterprise can initiate a second evaluation. If the result is C-level, the system will generate an "Reject Access" instruction and issue a decision letter with key veto factors. All decisions and execution statuses are recorded and fed back into the model iteration process, forming a closed loop of "evaluation-decision-feedback-optimization" to continuously improve the platform's risk management intelligence and operational efficiency.

[0049] In some embodiments, the method further includes: setting up an interpretability analysis unit to generate a feature contribution heatmap using the SHAP algorithm to demonstrate the impact of each feature on the risk level assessment result. This unit begins working after the risk grading model outputs the probability distribution of S / A / B / C levels. The SHAP algorithm, based on the Shapley value principle in game theory, calculates a quantified contribution score for each original or derived feature that constitutes the comprehensive feature vector input to the model. This score precisely reveals whether each feature has a positive impact on the final rating result, such as pushing the rating from B to A, or a negative impact, such as pulling the rating from A to B. Finally, these contributions are presented to platform operators or risk control experts in an intuitive visualization form: a feature contribution heatmap.

[0050] The introduction of this technology transforms the model's decision-making logic from invisible numerical calculations into a transparent, readable, and traceable business language. Its technological effects are multifaceted and profound: First, it significantly enhances the reliability and credibility of assessment results, enabling decision-makers to understand why a particular level was assigned and thus dare to execute critical business instructions such as access, rectification, or rejection based on these results. Second, it provides assessed enterprises with clear, specific, and actionable optimization directions. For example, the report can explicitly state that "the rating was downgraded by one level due to the 'database connection pool exhaustion' log feature," allowing enterprises to precisely optimize database configurations rather than blindly investigating. Finally, these analytical results also provide model development and operation teams with a basis for diagnosing model behavior, helping them discover potential biases in feature engineering or training data, thereby driving the continuous optimization and iteration of the risk assessment model itself.

[0051] On the other hand, the present invention also provides a factoring business tiered access device, including a processor, a memory, and a computer program or instructions stored in the memory. The processor is used to execute the computer program or instructions, and when the computer program or instructions are executed, the device implements the steps of the above method.

[0052] On the other hand, the present invention also provides a computer-readable storage medium having a computer program or instructions stored thereon, which, when executed by a processor, implement the steps of the above-described method.

[0053] On the other hand, the present invention also provides a computer program product, including a computer program or instructions that, when executed by a processor, implement the steps of the above-described method.

[0054] The present invention will now be described with reference to a specific embodiment: This embodiment provides a large-model-driven, test-free intelligent hierarchical access performance testing system suitable for supply chain finance platforms. Through multi-source data fusion, deep feature extraction from a large model, and intelligent hierarchical decision-making, it achieves access performance evaluation without the need for actual testing. (Refer to...) Figure 2 The system comprises five core modules and four supporting mechanisms, with the following specific architecture: 1. Data Acquisition and Preprocessing Module 1.1 Data Acquisition Unit: Adopting a multi-source heterogeneous data access architecture, it supports data acquisition from multiple data sources, including public cloud, private cloud, hybrid cloud, and local systems, covering three major categories of data: A. Structured data: Enterprise financial statements (balance sheet, profit and loss statement), transaction records (order volume, number of financing applications), performance monitoring indicators (API response time, TPS, server resource utilization); B. Unstructured data: corporate public opinion news, social media comments, system operation logs (anomaly stack text, operation and maintenance change records), contract texts, and PDF format vouchers; C. Semi-structured data: XML / JSON format interface logs, industry policy documents, and cross-border trade customs declaration data.

[0055] Specifically, Flume+Kafka is used to build a real-time data acquisition channel, supporting high-concurrency data access with a peak of 10,000 records per second; data format standardization and adaptation are achieved through Schema Registry, which is compatible with the differences in data format from different data sources.

[0056] 1.2 Data Preprocessing Unit: Performs a four-step process of "cleaning - standardization - vectorization - de-identification": Data cleaning: Implement missing value imputation (median for numeric values, default identifier for text values) and outlier removal (identify numerical outliers using IQR rules and filter text garbage data using regular expressions) based on Spark SQL. Standardization: Numerical data is standardized using Z-Score (X' = (X - mu) / sigma), and text data is standardized using UTF-8 encoding to unify the data format; Vectorization: Numerical data is directly converted into tensor format, text data is segmented using WordPiece to generate an index sequence, and image-based voucher data is extracted using CNN to extract visual feature vectors; Data anonymization: Differential privacy technology (adding Laplace noise) is used to process sensitive corporate information (such as legal representative ID number and bank account), and regular expression replacement is used to de-identify private data such as mobile phone number and email address, which complies with the requirements of the Personal Information Protection Law and the Financial Data Security Guidelines.

[0057] 2. Large Model Feature Extraction and Fusion Module 2.1 Feature Extraction Unit: Adopting a "pre-trained large model + fine-tuning adaptation" architecture, dedicated feature extraction schemes are designed for different data types. Text data (public opinion, logs, contracts): The BERT-base model is selected as the basic framework and fine-tuned through a corpus of supply chain finance (500,000+ industry texts) to output a 768-dimensional contextual semantic embedding vector, capturing risk signals (such as the semantic association of keywords "overdue" and "default") and business features (such as the semantics of key contract clauses) in the text. Numerical data (financial indicators, performance monitoring data): Using GPT-4's numerical understanding capabilities, discrete numerical values ​​are transformed into high-dimensional vectors with contextual relationships (e.g., combining the semantic relationship between "daily average order volume of 1 million" and "P99 response time of 500ms"), outputting a 512-dimensional numerical feature vector. Hybrid data (semi-structured + structured): By aligning textual features with numerical features through a cross-modal attention mechanism, semantic association extraction across data types can be achieved.

[0058] 2.2 Feature Fusion Unit: A feature fusion network is constructed based on a multi-head attention mechanism. The specific process is as follows: The dimensions of textual features, numerical features, and visual features are unified (mapped to 1024 dimensions); the weights of features from different data sources are dynamically allocated through attention weight calculation, such as giving higher weights to risk sentiment data than to regular transaction data; residual connections and LayerNorm layers are used to alleviate the gradient vanishing problem, and the fused 1024-dimensional comprehensive feature vector is output for subsequent hierarchical decision-making.

[0059] 3. Intelligent Hierarchical Engine Module 3.1 Hierarchical Model Unit: Employs a lightweight classifier and online learning architecture, with a fully connected neural network (FCN) at its core. The specific structure is as follows: Input layer: 1024-dimensional fused feature vector.

[0060] Hidden layers: 3 fully connected layers, using ReLU activation function, with a Dropout layer (dropout rate=0.2) in each layer to prevent overfitting.

[0061] Output layer: Employs the Softmax activation function to output a probability distribution of four performance levels: S, A, B, and C.

[0062] Technical optimization: Knowledge distillation technology is introduced to transfer the decision knowledge of large models to lightweight classifiers, thereby increasing the inference speed by 3 times (inference time per data point ≤100ms) while ensuring the classification accuracy (accuracy ≥92%).

[0063] 3.2 Dynamic iterative unit, supporting online learning and adaptive threshold adjustment.

[0064] Online learning is implemented using an incremental learning algorithm (Ftrl optimizer), which absorbs evaluation feedback data and market change data from newly connected enterprises in real time. The model is automatically fine-tuned for every 1,000 new samples, eliminating the need for full retraining. Threshold optimization is performed by dynamically adjusting the grading thresholds based on industry benchmark data and platform business needs using a genetic algorithm. For example, the S-level performance threshold iterates with the industry average level to ensure the objectivity and adaptability of the grading results.

[0065] 4. Results Feedback and Visualization Module 4.1 Result Output Unit: Pushes the grading results (S / A / B / C level) to the front end of the supply chain finance platform in real time, and outputs the grading basis (e.g., "Graded as B level due to P99 response time exceeding the standard") and key influencing factors (Top 5 influencing characteristics ranked).

[0066] 4.2 The interpretability analysis unit uses the SHAP (SHapley Additive exPlanations) algorithm to generate a feature contribution heatmap, which visually displays the positive / negative impact of each feature on the grading results. For example, "the text feature 'database connection pool exhausted' in the system log reduces the grading by 1 level".

[0067] 4.3 The visualization unit supports radar charts (performance score distribution across dimensions), trend charts (performance changes during model iteration), and topology diagrams (data lineage). It also provides PDF / Excel report export functionality to meet the needs of platform operation and maintenance and decision-making scenarios.

[0068] 5. Data Security and Governance Module 5.1 The data security unit adopts a full lifecycle security mechanism of "transmission-storage-use-destruction": the transmission process uses SM4 national cryptographic algorithm for encryption, the storage uses AES-256 encryption, the use process sets role-based access levels (administrator / tester / viewer), and the destruction process adopts a dual mechanism of data overwriting + physical deletion. 5.2 The data governance unit supports data catalog management, lineage analysis, anomaly monitoring, and impact analysis. It uses Spark to build data quality verification rules (such as data integrity and consistency verification), monitors anomalies in data transmission and processing in real time, and generates alarm logs.

[0069] The system described in this embodiment supports the following mechanisms: 1. Multi-terminal deployment adaptation mechanism: Supports deployment on public cloud, private cloud, hybrid cloud, and local system. It achieves elastic scaling of resources through containerization technology (Docker+K8s) to adapt to the deployment needs of supply chain finance platforms of different sizes. 2. Standardized Interface Mechanism: Provides RESTful API and SDK interfaces (supporting Python / Java / Go languages), compatible with mainstream supply chain finance platform architectures, and supports rapid connection and integration; 3. Scenario-based adaptation mechanism: Built-in feature weight configuration templates for special scenarios such as cross-border factoring, SME factoring, and intangible asset factoring. Users can customize scenario parameters to improve the evaluation accuracy of niche scenarios; 4. Resume interruption mechanism: Resume interruption is supported during data acquisition and model inference. When network interruption or system failure occurs, execution can continue from the breakpoint after recovery, avoiding data loss and duplicate calculations.

[0070] Reference Figure 3 The system execution flow described in this embodiment is as follows: 1. Multi-source data acquisition: Obtain structured, unstructured, and semi-structured data from various data sources through standardized interfaces; 2. Data Preprocessing: Perform cleaning, standardization, vectorization, and data anonymization operations to generate compliant and usable preprocessed data; 3. Large model feature processing: Extract deep features from various types of data through the fine-tuned large model, and fuse multi-source features using an attention mechanism; 4. Intelligent hierarchical decision-making: Input the fused features into a lightweight classifier and output S / A / B / C level performance ratings; 5. Results Feedback and Model Iteration: The hierarchical results and interpretability analysis are fed back to the application layer, and the feedback data is absorbed through the online learning module to dynamically optimize the model.

[0071] For example, taking the "performance evaluation of the access of a large supply chain finance platform to the 'one-stop supplier service subsystem'" as an example, the implementation process of the present invention is explained in detail to verify the feasibility and advantages of the "test-free" intelligent grading.

[0072] A large supply chain finance platform needs to conduct a performance evaluation of a newly integrated one-stop supplier service subsystem to determine its ability to handle high-concurrency business traffic from core enterprises. Traditional solutions require building a simulated test environment, constructing at least 100,000 test data points, and conducting a two-week stress test, which is resource-intensive and time-consuming. This embodiment adopts a "test-free" approach, completing the evaluation based on historical production environment data of the subsystem. The specific implementation steps are as follows: Step 1: Data Acquisition and Preprocessing Perform data collection, gathering production operation data from the past 30 days for this subsystem using Flume+Kafka, including: Structured data: average daily number of transaction orders, number of financing applications, average API response time and P95 / P99 percentile, TPS, server CPU / memory utilization, number of database connections, and number of slow queries.

[0073] Unstructured data: system operation logs, operation and maintenance change records.

[0074] Semi-structured data: JSON format cross-border trade interface logs.

[0075] Data preprocessing was performed, including: data cleaning to remove invalid and redundant data (approximately 5%) from the logs, and identifying and correcting 10 abnormal TPS data points using the IQR rule. Standardization was performed by applying Z-score standardization to numerical data such as response time and CPU utilization, and WordPiece segmentation was used for the log text. Vectorization was performed, generating 768-dimensional embedding vectors for text data using a fine-tuned BERT model, and converting numerical data into 512-dimensional tensors. Data anonymization was performed, using differential privacy anonymization to remove server IP addresses and maintenance personnel contact information from the logs.

[0076] Step 2: Large Model Feature Extraction and Fusion Feature extraction: The preprocessed text vectors and numerical vectors are input into the fine-tuned BERT-base model to extract deep semantic features and numerical correlation features. For example, the model identifies a strong correlation between the logs "P99 response time 450ms" and "database connection pool exhausted", thus capturing the system's concurrent processing bottleneck.

[0077] Feature fusion: The weights of each feature are calculated through a multi-head attention mechanism. For example, the weight of the abnormal log feature is 0.35, the weight of the response time feature is 0.25, and the weight of the CPU utilization feature is 0.2. The features are fused to generate a 1024-dimensional comprehensive feature vector, highlighting the core influencing factors.

[0078] Step 3: Intelligent Hierarchical Decision Making Hierarchical reasoning: The comprehensive feature vector is input into the lightweight FCN classifier, and the model outputs the probability distribution: 65% probability of level A, 30% probability of level B, 3% probability of level S, and 2% probability of level C, and finally it is judged as level A.

[0079] Threshold calibration: Based on industry benchmark data, the excellent threshold for P99 response time of similar subsystems is ≤300ms. The grading engine dynamically adjusts the threshold through a genetic algorithm and confirms that this subsystem does not meet the S-level standard because its P99 response time is slightly higher than the excellent threshold.

[0080] Output results: Grade A results and decision basis: The core performance indicators of the system meet the standards and can be connected to regular business; however, the P99 response time (450ms) is close to the threshold and there is a risk of database connection pool exhaustion, so the database configuration needs to be optimized.

[0081] Step 4: Results Feedback and Model Iteration Results feedback: The platform operations and maintenance team viewed the radar chart and SHAP heatmap through the visualization module to identify optimization directions.

[0082] Model Iteration: The evaluation data and feedback results of this subsystem are incorporated into the incremental learning sample library. The system automatically triggers model fine-tuning, updates the association weights between the "database connection pool exhaustion" feature and the response time feature, and improves the evaluation accuracy of subsequent similar subsystems.

[0083] Step 5: Secondary evaluation after optimization Based on the recommendations, the operations and maintenance team optimized the database connection pool configuration, expanding it to 2000 connections. The system collected production data for 7 days after the optimization and repeated steps 1-3 above. The second evaluation result was S-level, confirming that the subsystem could be connected to the core enterprise's high-concurrency business.

[0084] The beneficial effects of this embodiment include: Test-free and highly efficient: No need to set up a simulation test environment or construct test data. Evaluation is completed based on production / pre-production data, reducing the cycle from several weeks to several hours, improving efficiency by more than 90%.

[0085] Accurate and comprehensive: The large model supports multi-source heterogeneous data processing, integrates structured and unstructured data features, and achieves an evaluation accuracy of ≥92%, avoiding the one-sidedness of traditional methods.

[0086] Dynamic adaptation: Supports online learning and dynamic threshold optimization, adapts to market changes and special scenario requirements in real time, and has strong generalization ability.

[0087] Highly interpretable: The SHAP algorithm provides feature contribution analysis, clarifies the grading criteria and optimization direction, and facilitates rapid iteration.

[0088] Security and compliance: The end-to-end data anonymization and encryption mechanism meets the data security and privacy protection requirements of the financial industry.

[0089] Low cost and easy integration: The lightweight classifier reduces hardware resource consumption, and the standardized interface supports quick integration with existing platforms, reducing deployment costs by 60%.

[0090] Corresponding to the above method, the present invention also provides an apparatus / system including a computer device, the computer device including a processor and a memory, the memory storing computer instructions, the processor executing the computer instructions stored in the memory, and when the computer instructions are executed by the processor, the apparatus / system performs the steps of the method as described above.

[0091] This invention also provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the steps of the aforementioned edge computing server deployment method. The computer-readable storage medium can be a tangible storage medium, such as random access memory (RAM), main memory, read-only memory (ROM), electrically programmable ROM, electrically erasable programmable ROM, registers, floppy disks, hard disks, removable storage disks, CD-ROMs, or any other form of storage medium known in the art.

[0092] In summary, the factoring business tiered access method and device described in this invention collects multi-source heterogeneous data from target enterprises through standardized interfaces, covering dimensions such as financial indicators, system performance, public opinion dynamics, and compliance texts. A professional data preprocessing process ensures data quality and security. A risk tiering model with a fusion architecture is employed: a domain-adjusted BERT model extracts semantic features from the text, combined with the numerical understanding capabilities of the GPT series models to parse structured data, a cross-modal attention mechanism aligns multi-source features, a multi-head attention network generates a comprehensive feature vector, and finally, a lightweight fully connected network outputs the risk level, forming a corresponding access strategy based on the risk level. This invention achieves test-free evaluation, compressing the traditional testing cycle of several weeks to several hours. Through an online learning mechanism, it continuously optimizes model adaptability, achieving multi-dimensional risk assessment and significantly reducing testing resource consumption and time costs.

[0093] Furthermore, this invention improves the generalization capability for complex factoring businesses by dynamically adjusting risk levels and combining incremental updates to accurately identify risks in target enterprises. By leveraging SHAP interpretability technology to identify the contribution of each feature, risk factors can be precisely located.

[0094] Those skilled in the art will understand that the exemplary components, systems, and methods described in conjunction with the embodiments disclosed herein can be implemented in hardware, software, or a combination of both. Whether implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this invention. When implemented in hardware, it can be, for example, electronic circuits, application-specific integrated circuits (ASICs), appropriate firmware, plug-ins, function cards, etc. When implemented in software, the elements of this invention are programs or code segments used to perform the desired tasks. The programs or code segments can be stored in a machine-readable medium or transmitted over a transmission medium or communication link via data signals carried in a carrier wave.

[0095] It should be clarified that the present invention is not limited to the specific configurations and processes described above and shown in the figures. For the sake of brevity, detailed descriptions of known methods are omitted here. In the above embodiments, several specific steps are described and shown as examples. However, the method process of the present invention is not limited to the specific steps described and shown. Those skilled in the art can make various changes, modifications, and additions, or change the order of steps, after understanding the spirit of the present invention.

[0096] In this invention, features described and / or illustrated for one embodiment may be used in the same or similar manner in one or more other embodiments, and / or combined with or in place of features of other embodiments.

[0097] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. For those skilled in the art, various modifications and variations can be made to the embodiments of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for hierarchical access to a factoring service, characterized in that The method comprises the following steps: The target enterprise object is based on a standardized interface to obtain multi-source heterogeneous business data, which includes structured financial and business data, system technical performance data, unstructured enterprise public opinion data, operation status data and business compliance data, semi-structured interface logs and industry policy file data; The multi-source heterogeneous business data is cleaned, standardized, vectorized and desensitized; The input pre-trained risk grading model in the multi-source heterogeneous business data outputs a risk level evaluation result; the risk grading model includes a feature extraction unit, a feature fusion unit and a grading unit, the feature extraction unit uses a large model based on a BERT base to extract text features from the text data in the multi-source heterogeneous business data, and uses a large model based on a GPT base to extract numerical features from the numerical data in the multi-source heterogeneous business data, and aligns the text features and the numerical features based on a cross-modal attention mechanism; the feature fusion unit constructs a feature fusion network based on a multi-head attention mechanism to extract a comprehensive feature vector from the text features and the numerical features; the grading unit is a fully connected neural network, which takes the comprehensive feature vector as input and outputs a probability distribution for multiple risk levels as the risk level evaluation result; According to the risk level evaluation result, a factoring business access scheme is generated for the target enterprise according to a pre-set rule and is executed, the factoring business access scheme includes whether to allow access and provide factoring business, rectification requirements, and business scope configured according to the risk level.

2. The method of claim 1, wherein the access level is determined by the type of the transaction. For the target enterprise object, multi-source heterogeneous business data is obtained based on a standardized interface, which also includes: A real-time data collection channel is built using the Flume system and the Kafka platform, and the data formats of multiple data sources are compatible based on the Schema Registry service.

3. The method of claim 1, wherein the access level is determined by the type of the transaction. The multi-source heterogeneous business data is cleaned, standardized, vectorized and desensitized, which includes: Text extraction is performed on the image data in the multi-source heterogeneous business data using optical character recognition to obtain corresponding text data; The missing text data in the multi-source heterogeneous business data is filled with a default identifier based on the Spark SQL module, and abnormal text is filtered based on a pre-set regular expression; the missing numerical data in the multi-source heterogeneous business data is filled with a median value, and numerical anomalies are identified and removed by IQR rule; The text data is standardized using UTF-8 encoding, and the numerical data is standardized using Z-Score; After the text data is segmented by WordPiece, an index sequence is generated, and the numerical data is converted to a tensor format to complete the vectorization; Sensitive information of the enterprise is added with Laplace noise in the form of score privacy, or the sensitive information of the enterprise is replaced based on a regular expression.

4. The method of claim 1, wherein the access level is determined by the type of the transaction. The pre-training step of the risk grading model includes: Acquire a training sample set containing a plurality of samples, each sample containing a plurality of source heterogeneous business sample data collected for a sample enterprise object, and add a corresponding risk level as a label; Pre-train a risk grading initial model containing a feature extraction unit, a feature fusion unit and a grading unit using the training sample set, the feature extraction unit uses a large model based on a BERT base to extract sample text features from text data in the plurality of source heterogeneous business sample data, uses a large model based on a GPT base to extract sample numerical features from numerical data in the plurality of source heterogeneous business sample data, and aligns the sample text features and the sample numerical features based on a cross-modal attention mechanism; the feature fusion unit constructs a feature fusion network based on a multi-head attention mechanism to extract a sample comprehensive feature vector from the sample text features and the sample numerical features; the grading unit is a fully connected neural network, which takes the sample comprehensive feature vector as input and outputs a risk level prediction value, and updates the parameters of the risk grading initial model according to the deviation between the risk level prediction value and the label to obtain the risk grading model.

5. The method of claim 4, wherein the access level is determined by the supplier. The grading unit uses a fully connected neural network, including an input layer, a hidden layer and an output layer, the hidden layer contains at least 3 fully connected layers, each fully connected layer is provided with a ReLU activation function layer and a Dropout layer; the output layer uses a Softmax activation function layer.

6. The method of claim 5, wherein the access level is determined by the supplier. The method further comprises: setting a dynamic iteration unit to collect the plurality of source heterogeneous business data from newly accessed enterprise objects to construct an incremental data set; performing incremental learning when the incremental data set reaches a preset number; Setting a threshold optimization unit, iteratively adjusting the grading threshold of the risk level based on the industry average level to adjust the risk level of the newly accessed enterprise objects in the incremental learning process.

7. The method of claim 1, wherein the access level is determined by a predetermined rule. The method further comprises: setting an explainability analysis unit, using a SHAP algorithm to generate a feature contribution degree heat map to show the influence of each feature on the risk level evaluation result.

8. A factoring service hierarchical access apparatus comprising a processor, a memory, and a computer program or instructions stored on the memory, wherein, The processor is configured to execute the computer programs or instructions, and when the computer programs or instructions are executed, the device implements the steps of the method of any one of claims 1 to 7.

9. A computer readable storage medium having stored thereon a computer program or instructions, characterized in that, The computer programs or instructions are executed by the processor to implement the steps of the method of any one of claims 1 to 7.

10. A computer program product comprising computer programs or instructions, characterized in that, The computer programs or instructions are executed by the processor to implement the steps of the method of any one of claims 1 to 7. The computer programs or instructions are executed by the processor to implement the steps of the method of any one of claims 1 to 7.

Images