A financial product recommendation method and system for a supply chain financial platform

Abstract

This invention provides a method and system for recommending financial products to supply chain finance platforms. The method includes: determining financial supply chain relationships, enterprise behavior data, and a wide table of enterprise master data for upstream and downstream enterprises based on multi-dimensional enterprise data from the supply chain finance platform; extracting enterprise attribute features for each upstream and downstream enterprise from the determined financial supply chain relationships, enterprise behavior data, and the wide table of enterprise master data using a feature extraction algorithm; inputting the extracted enterprise attribute features into an enterprise behavior prediction model to output the behavioral probability of each upstream and downstream enterprise in a specific future period; inputting the enterprise attribute features and behavioral probabilities into a product matching model pre-stored with various financial product attributes to output the matching results between various financial products and each upstream and downstream enterprise, thereby recommending specific financial products to upstream and downstream enterprises based on the matching results. This invention can accurately and personally recommend financial services.

Description

Technical Field

[0001] This invention relates to the field of big data analysis and intelligent recommendation technology, and in particular to a method and system for recommending financial products for supply chain finance platforms. Background Technology

[0002] With the rapid development of supply chain finance, corporate financial service recommendations have gradually become a core capability for improving the conversion rate and user experience of supply chain finance platforms. The essence of providing corporate financial services is to offer suitable financial products to enterprises. Existing financial product recommendations mainly fall into the following categories: ① Rule-based financial service recommendations: Recommendation rules are set based on static attributes of upstream and downstream enterprises in the supply chain (including core enterprises and their suppliers and distributors), such as registered capital and industry category, and recommendations are triggered through transaction amount thresholds or fixed marketing cycles; ② Financial service recommendations based on simple time series analysis: Traditional time series analysis methods such as moving averages are used to predict the demand for financial services (such as financing needs) of upstream and downstream enterprises in the supply chain, or logistic regression models are used to make periodic service recommendations based on historical financing time points.

[0003] However, existing recommendation schemes have the following shortcomings: ① The single data dimension leads to the inability to achieve accurate recommendations: Various supply chain finance platforms have accumulated a large amount of enterprise transaction data, financing behavior data, and supply chain relationship data during the management process. However, existing technologies often use single-dimensional data when formulating recommendation schemes, failing to make full use of the complex relationship network and dynamic behavior data (such as cloud credit transfer behavior) in the financial supply chain. This also makes the generated recommendation schemes lack a deep understanding of the financial supply chain environment, resulting in low service recommendation accuracy; ② The degree of recommendation personalization is low: Existing schemes are unable to provide differentiated financial services based on the specific role and real-time status of enterprises.

[0004] Therefore, there is an urgent need for a financial product recommendation method that can fully utilize the multi-dimensional and complex data stored on supply chain finance platforms to improve the accuracy of financial product recommendations and achieve personalized recommendations. Summary of the Invention

[0005] In view of this, embodiments of the present invention provide a method and system for recommending financial products to supply chain finance platforms, so as to accurately and personally recommend financial services to upstream and downstream enterprises in the supply chain.

[0006] One aspect of the present invention provides a method for recommending financial products for a supply chain finance platform, the method comprising the following steps: Based on the multi-dimensional enterprise data of the supply chain finance platform, we can determine the financial supply chain relationship, enterprise behavior data and enterprise owner data wide table for upstream and downstream enterprises in the supply chain. The feature extraction algorithm is used to extract the enterprise attribute features of each upstream and downstream enterprise in the established financial supply chain relationship, enterprise behavior data, and enterprise master data wide table; among which, the enterprise attribute features include enterprise semi-static attribute features, enterprise dynamic attribute features, and enterprise static attribute features. The extracted enterprise attribute features are input into the enterprise behavior prediction model, and the output is the probability of the behavior of each upstream and downstream enterprise in the supply chain in a specific future period. The product matching model, which pre-stores the attributes and behavioral probabilities of various financial products, inputs enterprise attribute characteristics and behavioral probabilities. The output is the matching results of various financial products with each upstream and downstream enterprise in the supply chain. Based on the matching results, specific financial products are recommended to upstream and downstream enterprises in the supply chain.

[0007] In some embodiments of the present invention, recommending specific financial products to upstream and downstream enterprises in the supply chain based on matching results includes: obtaining a recommendation execution strategy based on the behavioral probability of each upstream and downstream enterprise in the supply chain and the matching results through a recommendation prediction model, and realizing recommendation outreach according to the recommendation execution strategy; wherein, the recommendation execution strategy is used to indicate whether to recommend a specific type of financial product, the recommendation time, and the recommendation channel to specific upstream and downstream enterprises in the supply chain.

[0008] In some embodiments of the present invention, after the recommendation is reached, the method further includes: acquiring reach record data, and determining the recommendation effect corresponding to the recommendation execution strategy based on the reach record data; wherein the evaluation indicators of the recommendation effect include one or more of response rate, conversion rate and satisfaction.

[0009] In some embodiments of the present invention, the method further includes: using the recommendation effect as a reward, the model parameters of the enterprise behavior prediction model, the product matching model, and the recommendation prediction model as states, and the enterprise attribute features and recommendation execution strategies as actions, updating the enterprise behavior prediction model, the product matching model, and the recommendation prediction model through a reinforcement learning algorithm.

[0010] In some embodiments of the present invention, the matching results between various financial products and each upstream and downstream enterprise in the supply chain are obtained in the following manner: The integrated enterprise attribute characteristics of each enterprise group are determined based on the enterprise attribute characteristics of each upstream and downstream enterprise in the supply chain; wherein each upstream and downstream enterprise in the supply chain belongs to an enterprise group. Determine the matching degree between the integrated enterprise attribute characteristics of each enterprise group and the various financial product attributes pre-stored in the product matching model, and use it as the preference strength of each enterprise group for various financial products; based on the behavioral probability of each upstream and downstream enterprise in the supply chain and the similarity between pairs of enterprises in the enterprise group to which the enterprise belongs, determine the collaborative filtering score of each upstream and downstream enterprise in the supply chain for various financial products, and then determine the collaborative filtering score of each enterprise group for various financial products. For each enterprise group, the matching results between the enterprise group and various financial products are determined based on preference strength and collaborative filtering scores, thereby obtaining the matching results of all enterprises in the enterprise group with various financial products.

[0011] In some embodiments of the present invention, financial supply chain relationships, corporate behavior data, and enterprise master data wide tables for upstream and downstream enterprises in the supply chain are determined based on multi-dimensional enterprise data from a supply chain finance platform, including: The financial supply chain relationship is determined based on the supply chain hierarchy configuration between upstream and downstream enterprises in the supply chain finance platform. Based on the operational logs of the supply chain finance platform, the enterprise behavior data of each upstream and downstream enterprise in the supply chain is determined; and The business data from the supply chain finance platform is used to create a wide table of enterprise master data for each upstream and downstream enterprise in the supply chain.

[0012] In some embodiments of the present invention, the semi-static attribute features of an enterprise include the enterprise's supply chain location information, the dynamic attribute features of an enterprise include one or more of the following: electronic credit certificate activity, payment cycle, and financing frequency, and the static attribute features of an enterprise include enterprise credit assessment. The enterprise behavior prediction model is built upon long short-term memory networks and attention mechanisms; and Financial product attributes include basic identification attributes, scope of application attributes, and business rule attributes.

[0013] Another aspect of the present invention provides a financial product recommendation system for a supply chain finance platform, including a processor, a memory, and a computer program / instructions stored in the memory. The processor is used to execute the computer program / instructions, and when the computer program / instructions are executed, the system implements the steps of the method described in any of the above embodiments.

[0014] Another aspect of the present invention provides a computer-readable storage medium having a computer program / instructions stored thereon, which, when executed by a processor, implement the steps of the method described in any of the above embodiments.

[0015] Another aspect of the present invention provides a computer program product, including a computer program / instructions that, when executed by a processor, implement the steps of the method described in any of the above embodiments.

[0016] The financial product recommendation method and system proposed in this invention for supply chain finance platforms can acquire financial supply chain relationships, enterprise behavior data, and enterprise master data wide tables based on the multi-dimensional and complex data stored on the supply chain finance platform. Based on the acquired data, it can predict the financial service needs of upstream and downstream enterprises in the supply chain, and then recommend more suitable financial products to enterprises based on the characteristics of the financial products. The method proposed in this application can achieve accurate and personalized financial service recommendations.

[0017] 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 description, 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.

[0018] 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

[0019] 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 a financial product recommendation method for a supply chain finance platform according to an embodiment of the present invention.

[0020] Figure 2 This is a schematic diagram of feature extraction in one embodiment of the present invention.

[0021] Figure 3 This is a schematic diagram of a process for predicting the probability of corporate behavior in one embodiment of the present invention.

[0022] Figure 4 This is a flowchart illustrating a financial product recommendation method for a supply chain finance platform according to another embodiment of the present invention.

[0023] Figure 5 This is a schematic diagram of the architecture of a financial product recommendation system for a supply chain finance platform according to one embodiment of the present invention. Detailed Implementation

[0024] 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.

[0025] 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.

[0026] 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.

[0027] In the following description, embodiments of the invention will be illustrated with reference to the accompanying drawings. In the drawings, the same reference numerals represent the same or similar parts, or the same or similar steps.

[0028] Existing financial service recommendation technologies typically recommend financial products to upstream and downstream enterprises in the supply chain based on manually set rules or demand forecasts. However, they often use single-dimensional data and cannot fully understand the service needs of upstream and downstream enterprises based on enterprise user data on the supply chain finance platform, resulting in insufficient personalization of the recommendation schemes.

[0029] Based on this, this application constructs multi-dimensional features for supply chain scenarios by integrating enterprise supply chain relationship data, enterprise behavior data (such as data related to financial transaction behavior) and enterprise static attribute data, according to the business characteristics of the financial supply chain. It uses an enterprise behavior prediction model to predict the probability of behavior of upstream and downstream enterprises in the supply chain (specifically the probability of financing business behavior, especially the probability of financing demand) and combines it with a product matching model to achieve accurate and personalized matching of financial products.

[0030] The method proposed in this application is applicable to scenarios where supply chain finance platforms accurately recommend financial services to upstream and downstream enterprises in the supply chain. Figure 1 This is a flowchart illustrating a financial product recommendation method for a supply chain finance platform according to one embodiment of this application, such as... Figure 1 As shown, the method includes steps S110 to S140, as detailed below: Step S110: Based on the multi-dimensional enterprise data from the supply chain finance platform, determine the financial supply chain relationships, enterprise behavior data (historical behavior data), and enterprise master data wide tables (also known as enterprise dimension fusion wide tables) for upstream and downstream enterprises in the supply chain, providing a unified data source for the feature engineering in step S120. The multi-dimensional enterprise data includes supply chain hierarchical configurations (external configurations), enterprise operation logs, and business data generated by enterprises operating on the supply chain finance platform. This multi-dimensional enterprise data can be stored on the supply chain finance platform and can be updated according to configuration cycles or event-driven updates.

[0031] More specifically, the execution process of step S110 can be as follows: Based on the supply chain hierarchy configuration and / or trade transaction relationships between upstream and downstream enterprises in the supply chain finance platform (which can be obtained from enterprise operation logs), determine the financial supply chain relationship between upstream and downstream enterprises (which may involve all upstream and downstream enterprise users of the supply chain finance platform); determine the enterprise behavior data of each upstream and downstream enterprise based on the operation logs of the supply chain finance platform; and form a wide table of enterprise master data for each upstream and downstream enterprise based on the business data of the supply chain finance platform. Here, the supply chain hierarchy configuration indicates the positional order of upstream and downstream enterprises relative to the core enterprise in the supply chain network; the trade transaction relationship represents the actual trade relations between upstream and downstream enterprises; the financial supply chain relationship can be used to indicate the business connections between upstream and downstream enterprises, and the financial supply chain relationship can also be referred to as a multi-level supply chain network; enterprise behavior data refers to the recordable and analyzable operational traces and interaction information generated by upstream and downstream enterprises during their use of the supply chain finance platform, reflecting information such as enterprise activity, financial service preferences, credit habits, and potential risks. The operational and interactive behaviors of enterprises on the supply chain finance platform can be categorized into account... The business data of the supply chain finance platform should include basic enterprise information (such as the enterprise's legal identity, basic business information, management structure, equity and capital structure, and qualifications and licenses) and enterprise credit data, which are used to indicate the objective identity profile of the enterprise. In addition, the enterprise master data wide table is a core data asset that integrates the enterprise's static attributes, credit data and derived tags with the enterprise's unique identifier as the core, forming a unified, comprehensive and standardized enterprise-level data view. That is, the enterprise master data wide table in this application may not include the enterprise's semi-static attributes and dynamic indicator data.

[0032] As an example, financial supply chain relationships can be represented in the form of a graph. For instance, when using a directed weighted graph to represent financial supply chain relationships, an upstream or downstream enterprise in the supply chain can be defined as a network node, and the business connections between enterprises can be represented as directed network edges. The direction of the directed network edges is determined by the flow of funds, and the strength of the business relationship is indicated for each edge based on transaction density and / or dependence on the core enterprise. Enterprise behavioral data can be structured and presented in chronological order, such as using enterprise behavioral sequences. In a way that can be expressed, Corresponding companies At any moment ( This is a behavioral event, and the representation of the enterprise behavior sequence allows subsequent feature extraction algorithms to utilize the order and rhythm of behaviors in the supply chain scenario (such as the sequential relationship of "rights confirmation - financing - repayment"), which differs from the common approach of using only statistical features. In addition, financial supply chain relationships may include information in the following fields: the core enterprise to which each upstream and downstream enterprise belongs, the supply chain level (such as level 0 / 1 / 2 / 3), the hierarchical path, the relationship establishment time, and the relationship status (whether it exists); enterprise behavior data may include the identification information of each upstream and downstream enterprise, the type of behavioral event (such as holding, transferring, applying for financing, disbursement, repayment, login, and logout), timestamps, and transaction amounts. This application does not limit the specific field information contained in the financial supply chain relationships, enterprise behavior data, and enterprise master data wide tables.

[0033] Step S120: Utilize feature extraction algorithms to extract enterprise attribute features for each upstream and downstream enterprise in the established financial supply chain relationships, enterprise behavior data, and enterprise master data wide table. Step S120 can extract computable features from the data source provided in Step S110 for use by subsequent core algorithms.

[0034] More specifically, the features extracted in step S120 can be broadly categorized into three types: enterprise attribute features, including semi-static enterprise attribute features, dynamic enterprise attribute features, and static enterprise attribute features. Static enterprise attributes are inherent, unchanging basic attribute information of the enterprise. Semi-static enterprise attributes are attributes determined by external configurations or business relationships, relatively stable but adjustable (not inherent enterprise attributes). Dynamic enterprise attributes can continuously change with business activities (with time). (Updated), capable of reflecting the real-time operational status of enterprises. Therefore, this application is designed for upstream and downstream enterprises in the supply chain. Enterprise attribute characteristics may include enterprise semi-static attribute characteristics (This can be considered as being comprised of all upstream and downstream enterprises in the supply chain) Attributes and characteristics in supply chain networks Composition), Enterprise Dynamic Attributes and Characteristics and the static attributes and credit characteristics of enterprises .like Figure 2 As shown, the semi-static attribute features of an enterprise are extracted from the financial supply chain relationship (therefore, the semi-static attribute features of an enterprise can also be called supply chain network features), the dynamic attribute features of an enterprise are extracted from the enterprise behavior data (therefore, the dynamic attribute features of an enterprise can also be called dynamic behavior features), and the static attributes and credit features of an enterprise are extracted from the wide table of the enterprise master data. The static attributes and credit features of an enterprise specifically include the static attribute features of an enterprise and the credit features of an enterprise.

[0035] In some embodiments of this invention, the static attribute features of an enterprise are features corresponding to enterprise attributes that do not change rapidly over time (but can be updated over a relatively long period). These may include fields such as enterprise type, registered capital, and enterprise size. Enterprise credit features may include enterprise credit rating and enterprise risk score, among other enterprise credit assessment information. The dynamic attribute features of an enterprise may include one or more fields such as the activity level of electronic credit instruments (such as cloud credit), payment cycle, and financing frequency. If the financial supply chain relationship is determined using supply chain hierarchy configuration, the semi-static attribute features of an enterprise may include fields such as the enterprise's supply chain location, industry classification, and customer level. If the financial supply chain relationship is determined using trade transaction relationships, the semi-static attribute features of an enterprise may include information such as transaction density and dependence on core enterprises. This application does not specifically limit the field information included in the enterprise attribute features; they can be designed according to recommended requirements.

[0036] When using corporate behavior data to determine the dynamic attribute characteristics of a company, data can be extracted through a sliding window of fixed or variable length. Corporate behavior data including multiple behavioral events can be encoded into behavioral characteristics under multiple statistical periods (corresponding to the sliding window). When the behavioral events are tagged with financial products, they can be jointly encoded to obtain the dynamic attribute characteristics of the company.

[0037] As an example, this application describes the method for determining some field information in enterprise attribute characteristics, but this application is not limited to the following method, and those skilled in the art can design the calculation process themselves: A company's position in the supply chain can be coded at discrete levels based on the supply chain hierarchy, such as the supply chain position of a core enterprise. The supply chain position of Tier 1 suppliers The supply chain position of Tier 2 suppliers This process continues, ensuring that a company's position in the supply chain corresponds to its hierarchical position within the supply chain, and that if the same company... If multiple hierarchical paths exist, the lowest level is selected; if the enterprise level is unknown, a default level is assigned. Upstream and downstream enterprises in the supply chain Dependence on core enterprises and transaction density It can be determined based on trade transactions between enterprises, specifically, (When there is no transaction with the core enterprise) ), or It can also be calculated using logarithmic scaling. ; Electronic credit certificate activity By analyzing upstream and downstream enterprises in the supply chain This is obtained by counting events or summarizing transaction amounts in enterprise behavior data that are tagged with electronic credit certificates (taking Yunxin as an example). or ; Payment cycle (Unit: days) is calculated by taking the median or quantile of the time difference between "electronic credit certificate receipt / confirmation → financing or redemption"; if there is no historical enterprise behavior data, the industry default value is used. Financing frequency This is calculated by normalizing the number of successful financing applications or loan disbursements within the statistical period (using quarterly or annual normalization units); Enterprise credit ratings can be determined through enterprise master data wide tables or external credit reporting interfaces. They can be set to multiple discrete credit levels, such as A / B / C / D or score levels, and must be encoded as one-hot or ordered values. Enterprise Risk Score Scores can be generated based on the output of the internal risk control model or rules of the supply chain finance platform. The higher the value, the better for the company. The greater the risk.

[0038] The feature extraction algorithm in step S120 can be a traditional machine learning algorithm such as statistical methods (e.g., principal component analysis or t-SNE) or a deep learning algorithm based on convolutional network architecture or Transformer neural network. This application does not specifically limit the type of feature extraction algorithm.

[0039] After extracting features exhibiting different temporal variations, temporal and spatial alignment can be performed on the enterprise's semi-static, dynamic, and static attribute features to ensure that the final recommendation decision is based on a unified spatiotemporal feature representation. This alignment can be applied to the upstream and downstream enterprises in the supply chain obtained in step S120. Features , and The features are then stitched together and / or subjected to dimensionality reduction for unified use in downstream demand forecasting, product matching, and marketing decisions. Dimensionality reduction methods may include mapping to fixed-dimensional features using a fully connected network, autoencoders, and random projections, etc., and this invention is not limited to these methods. Furthermore, this application may also standardize or bin the continuous features obtained in step S120 to improve the stability of subsequent models.

[0040] Step S130: Extract the enterprise attribute features (which can be used below) Indicates enterprise The enterprise attribute characteristics are input into the pre-trained enterprise behavior prediction model, and the output is the behavior probability of each upstream and downstream enterprise in the supply chain in a specific future period (which may be the probability of the enterprise engaging in financing behavior). The specific future period can be a specific time period after the current time (at which point step S130 obtains the behavior probability distribution within the time period) or a specific point in time.

[0041] In some embodiments of the present invention, such as Figure 3 As shown, this application can use a corporate behavior prediction model built based on Long Short-Term Memory (LSTM) networks and attention mechanisms to model multi-dimensional features and output the prediction results of corporate financial service demand. Specifically, it models corporate attribute features... Input a Long Short-Term Memory (LSTM) network (which can be unidirectional or bidirectional) to obtain enterprise data. At each time step Hidden state : ; Using the attention mechanism to process the hidden state sequence of the LSTM output Calculate attention weights , ( and (as learnable parameters), thus obtaining the context vector. ;Will and After concatenation, the data is passed through a fully connected layer and an activation function (using the fully connected layer and activation function as the prediction head) to obtain the enterprise... Future period Behavioral probability ,in, and These represent the weights and biases, respectively. This represents the computation process of the fully connected layer and the activation function. .

[0042] As an example, this application can also improve the prediction accuracy of the enterprise behavior prediction model constructed by the Long Short-Term Memory Network-Attention Mechanism in the following ways, which are different from directly applying the general LSTM-Attention Mechanism model: ① By displaying information on the enterprise entity dimension (such as whether it is a core enterprise, level and industry category, etc.) and product dimension (historical recommendation effect of financial products such as electronic credit certificates or financing) in the enterprise behavior data, the LSTM can learn the differences in behavioral patterns under "different levels and different products" through the dynamic attribute characteristics of enterprises; ② Time-encoding enhancement is performed on key periods such as "before the expiration of electronic credit certificates" and "before repayment after financing" in the time dimension, so that the LSTM can fit the cash flow cycle; ③ Attention weights This can be interpreted as the importance of the "demand signal period"; when training the enterprise behavior prediction model, weak supervision (such as label propagation or auxiliary loss) can be applied to the time step before the occurrence of historical demand, so that the enterprise behavior prediction model pays more attention to the behavioral segments that are strongly related to the demand for financial services (such as the upcoming expiration of cloud credit, after a large transaction, etc.), thus distinguishing it from the uniform attention of general sequence models; ④ After the attention mechanism, by adding a multi-task head, the label of the enterprise's demand level for financial services can be predicted at the same time, thereby improving the representation quality through multi-task learning.

[0043] For example, the training process of the enterprise behavior prediction model can be as follows: Based on the multi-dimensional enterprise data of the supply chain finance platform in historical periods, obtain the corresponding enterprise attribute features, and use "whether the behavior of using financial products (such as whether financing behavior occurred)" or "whether marketing conversion occurred" as binary classification labels in that historical period. The time window is aligned with the time when demand occurs to avoid future information leakage. Cross-entropy is used as the loss function (if there are multiple task heads, the weighted sum is used as the total loss function) to iterate the model parameters and obtain the pre-trained enterprise behavior prediction model.

[0044] The above-described enterprise behavior prediction model using LSTM and attention mechanism is merely an example. This application may use graph neural networks or temporal convolutional networks to replace LSTM, or other types of time series prediction models as enterprise behavior prediction models. This invention is not limited thereto.

[0045] Step S140: Connect each upstream and downstream enterprise in the supply chain Enterprise Attributes and every upstream and downstream enterprise in the supply chain Behavioral probability The system inputs a product matching model pre-stored with attributes of various financial products and outputs matching results between these financial products and upstream and downstream enterprises in the supply chain. Based on these matching results, specific financial products are recommended to these enterprises. For each type of financial product, the product attributes may include one or more of the following: basic identification attributes, scope of application attributes, business rule attributes, risk control attributes, and operational configuration attributes. For example, basic identification attributes may include fields such as product code, product name, product type, and lifecycle status (including fundraising stage, operation stage, maturity redemption stage, liquidation stage, and termination stage); scope of application attributes may include fields such as applicable level and applicable industry; business rule attributes may include fields such as quota rules and term rules; risk control attributes may include fields such as risk level and guarantee method; and operational configuration attributes may include fields such as fee structure, cooperating funding party, and quota utilization method. This application does not specifically limit the field information of the product attributes pre-stored in the product matching model; users can choose to include or exclude fields based on recommendation needs.

[0046] As an example, if the product matching model stores too few product attributes (such as only including basic product attributes) to support the product matching process, then input... and Simultaneously, certain product attribute information is input based on the characteristics of the enterprise. The matching results obtained in step S140 can be presented in the form of a product recommendation list, and various financial products can be sorted in the product recommendation list according to the degree of matching. In addition, for the input data of the product matching model, historical usage records of various financial products can also be obtained from the supply chain finance platform (which may include product identification, the enterprise's level in the supply chain, the industry to which the enterprise belongs, and the enterprise's product usage feedback, etc.). The product matching model can adopt graph recommendation or sequence recommendation models, etc., and this invention does not specifically limit the network architecture of the product matching model.

[0047] This application can sequentially identify each upstream and downstream enterprise in the supply chain by repeatedly executing step S140. Matching results with financial products can also be achieved by setting up enterprise groups (which can be understood as sets of similar enterprises). After determining the matching results of each upstream and downstream enterprise group in the supply chain with the financial product, the matching results of the enterprise group are considered as the matching results of all enterprises included in the enterprise group with the financial product. Each upstream and downstream enterprise in the supply chain can belong to an enterprise group, and the enterprises included in each enterprise group are different. The design of enterprise groups can limit "similar enterprises" to a group with comparable businesses, which is different from global collaborative filtering. Moreover, pairwise similarity can be calculated for all enterprise users of the supply chain finance platform, or only for enterprises at the same supply chain level or under the same core enterprise. Enterprise groups can then be formed based on enterprise similarity (such as setting the two enterprises with the highest enterprise similarity in the same enterprise group) to reflect the preference similarity of "same scenario, same role". The formula for calculating enterprise similarity can be: It can also be based on the enterprise and Other similarity calculation methods can be used for enterprise attribute characteristics, and this invention is not limited to these. Furthermore, if two enterprises have similar conversion patterns for the same type of financial products in their historical behavioral data, a behavioral similarity item (such as...) can be added when calculating enterprise similarity. (Linear or nonlinear fusion).

[0048] In some embodiments of the present invention, this application can design and utilize a hybrid matching algorithm that integrates collaborative filtering and content matching to calculate enterprise similarity and enterprise preference in the context of the supply chain, thereby determining the matching results of various enterprise groups and various financial products. The specific calculation process is as follows: Step S01: Based on the enterprise attribute characteristics of each upstream and downstream enterprise in the supply chain Determine the integrated enterprise attribute characteristics of each enterprise group, such as by analyzing the enterprise groups. Enterprise groups are obtained by weighted fusion or feature extraction of the enterprise attribute characteristics of the included enterprises. Integration of corporate attributes and characteristics .

[0049] Step S02: Determine the matching degree between the integrated enterprise attribute characteristics of each enterprise group and the various financial product attributes pre-stored in the product matching model, and use this as the preference strength of each enterprise group for various financial products. Determining the preference strength aligns the "applicable level" and "applicable core enterprise circle" of financial products with the enterprise attribute characteristics. Products that do not meet the level or subject constraints can be filtered or downweighted, achieving filtering and ranking based on the supply chain context. Further, based on the behavioral probabilities of each upstream and downstream enterprise in the supply chain and the similarity between any two enterprises within the enterprise group to which that enterprise belongs, determine the collaborative filtering score for each upstream and downstream enterprise in the supply chain for various financial products. This, in turn, determines the collaborative filtering score for each enterprise group for various financial products. That is, for each enterprise group, based on the behavioral probabilities of the upstream and downstream enterprises in the supply chain related to various financial products and the similarity between any two enterprises within the enterprise group, determine the collaborative filtering score for that enterprise group for various financial products.

[0050] As an example, before calculating the preference strength, the various financial product attributes pre-stored in the product matching model can be adopted using features belonging to the same space as the enterprise attribute features. express( For type (The characteristics corresponding to the attributes of financial products). Calculation and The degree of matching between them can be used to obtain enterprise groups. For type The intensity of preference for financial products, such as and The formula for calculating the matching degree between them can be expressed as: , It can employ matching computation methods such as inner product, bilinear, or small neural networks. Indicates enterprise group Domestic enterprises and enterprises Taking the similarity between enterprises as an example, enterprises For type The formula for calculating the collaborative filtering score of financial products can be expressed as follows: ,in, In In addition to enterprises any other company besides Indicates enterprise With type The probability of behavior related to financial products. This is determined by analyzing enterprise groups. By calculating the weighted average of the collaborative filtering scores of all upstream and downstream enterprises in the supply chain for various financial products, a score for the enterprise group can be obtained. Collaborative filtering and scoring of various financial products .

[0051] Step S03: For each enterprise group, determine the matching results between the enterprise group and various financial products based on preference strength and collaborative filtering scores, thereby obtaining the matching results of all enterprises within the enterprise group with various financial products (the matching results of each upstream and downstream enterprise in the supply chain within the enterprise group with various financial products are the matching results of the enterprise group with financial products). For example, the weighted sum of preference strength and collaborative filtering scores can be used to obtain the enterprise group pair of type... The formula for matching financial products can be expressed as: . It can be configurable or learned through validation sets, such as adjusting the corresponding weights of preference strength and collaborative filtering scores based on behavioral probabilities: appropriately increasing the weights when the financing demand is high. To respond more quickly to explicit needs, increase the weighting when financing needs are low. To uncover potential preferences; adjustments can also be made based on the richness of historical behavioral data of the enterprise (more dependent on data when there is less data). ).

[0052] Existing financial service recommendation schemes often rely on fixed periods or simple event triggers, failing to optimize the timing of outreach based on the real-time status and cash flow rhythm of enterprises, resulting in inaccurate marketing timing judgments. To address this issue, this application utilizes a recommendation prediction model to determine feasible recommendation strategies when recommending specific financial products to upstream and downstream enterprises in the supply chain based on the matching results obtained in step S140. Specifically, the behavioral probability of each upstream and downstream enterprise in the supply chain and the matching results for each enterprise can be input into the recommendation prediction model, which outputs a recommendation execution strategy, and the recommendation outreach is implemented according to the strategy. The recommendation prediction model can preset multiple recommendation channels (such as in-app messages, SMS, or verbal notifications from account managers) to allow for appropriate channel selection. Furthermore, after outreach is achieved, outreach record data can be acquired, and the recommendation effect corresponding to the recommendation execution strategy can be determined based on this data, providing a monitoring signal for subsequent model updates. Reach record data can be obtained from reach devices or supply chain finance platforms. It refers to the full-link factual data recorded after content is pushed according to the recommendation execution strategy. It is used to describe the complete trajectory of "in what scenario, to which enterprise, what type of financial product was recommended, and what the result was". In other words, reach record data can include digital traces of the execution strategy and factual records of user feedback.

[0053] In some embodiments of this invention, the recommendation prediction model can be constructed based on a reinforcement learning algorithm. In this case, the recommendation prediction model can use the recommendation effect corresponding to the recommendation execution strategy as the reward (the reward may also include the enterprise's average reach interval and average reach number, etc.), use the output data of the enterprise behavior prediction model (such as the behavior probability of upstream and downstream enterprises in the supply chain) and the matching result as the state (the state may also include enterprise attribute features), and use the recommendation execution strategy as the action, thereby maximizing the recommendation effect of subsequent recommendation schemes (such as long-term conversion and satisfaction). Moreover, this application can also introduce market data (such as whether the current time is a weekday or holiday, the beginning or end of the month, the quota and interest rate of similar financial products in the market, etc.) and the current life cycle status of all financial products in the supply chain finance platform into the reinforcement learning algorithm environment. The reinforcement learning algorithm mentioned above is only an example, and this application does not specifically limit the network architecture and model type of the recommendation prediction model. For example, a contextual bandit can also be used to construct the recommendation prediction model.

[0054] The states and actions mentioned above in the recommendation prediction model can be represented either discretized or continuously. For example, discrete actions can be represented as... The discrete levels, while continuous actions can include continuous reach intensity values ​​and continuously selected recommendation intensities. When the recommendation effect is used as the reward, the reward function can be expressed as... ,in, , and These represent state, action, and reward, respectively. Indicates conversion rate ( ), For response rate, This indicates satisfaction levels. Different reward weights can be set for companies at different levels or with varying degrees of dependence on core enterprises (e.g., higher reward weights are given to companies with higher dependence and greater conversion value). Weighting enables recommendation methods to pursue a balance between long-term benefits and user experience within the supply chain structure.

[0055] As an example, Considered ( ),according to Sorting can yield companies The system matches various financial products and allows for the creation of recommendation and prediction models based on... ( When recommending financial products to each upstream and downstream enterprise in the supply chain (based on the matching results with various financial products), only the Top-K financial products (e.g., the three financial products with the highest matching scores) are selected and recommended to the enterprise. Recommendation. Similar to product matching models, the behavioral probabilities and matching results of each supply chain enterprise can be sequentially input into the recommendation prediction model. Alternatively, a mapping relationship can be constructed between behavioral probabilities and matching results and upstream and downstream enterprises in the supply chain, and the behavioral probabilities and matching results of all enterprises can be input into the recommendation prediction model together.

[0056] The recommendation execution strategy is used to indicate whether to recommend a specific type of financial product to specific upstream and downstream enterprises in the supply chain. When determining to recommend to a particular enterprise, the recommendation execution strategy may also include fields such as recommendation time and recommendation channel. Therefore, when generating a recommendation execution strategy, corresponding recommendation schemes can be generated for each upstream and downstream enterprise in the supply chain, or corresponding recommendation schemes can be generated for various types of financial products; however, this invention is not limited to these.

[0057] In some embodiments of this invention, the application can evaluate the effectiveness of financial service recommendations using response rate, conversion rate, and / or satisfaction rate, optimize marketing strategies through recommendation prediction models, and ultimately achieve personalized financial service recommendations using enterprise behavior prediction models, product matching models, and recommendation prediction models. Specifically, the response rate indicates the percentage of enterprise users who receive the recommendation message and browse / click on it (response rate = number of enterprises that respond / total number of recommended enterprises); the conversion rate indicates the percentage of enterprises that adopt financial products within a specific period after receiving the recommendation message (e.g., the conversion rate could be the percentage of recommended enterprises that ultimately complete a financing transaction); and the satisfaction rate indicates the risk of enterprises receiving the recommendation message having negative feedback (e.g., unsubscribing or complaining), i.e., the percentage of recommended enterprises satisfied with the recommended products and services.

[0058] The enterprise behavior prediction model, product matching model, and recommendation prediction model mentioned in this application support incremental learning or periodic full retraining. For example, the enterprise behavior prediction model, product matching model, and recommendation prediction model can be updated periodically or based on recommendation performance. The model update frequency and the triggering conditions for model updates can be designed according to requirements. This application can use reinforcement learning algorithms to achieve continuous model updates, forming... Figure 4 The closed loop of "recommendation execution - effect monitoring - model update" is shown. The model of this application can be updated by reinforcement learning. Specifically, the recommendation effect is used as the reward, the model parameters of the enterprise behavior prediction model, product matching model and recommendation prediction model are used as the state, and the enterprise attribute features and recommendation execution strategy are used as the actions. The enterprise behavior prediction model, product matching model and recommendation prediction model are updated by reinforcement learning algorithm.

[0059] This application does not limit the specific type of reinforcement learning algorithm; for example, deep Q-networks, Actor-Critic algorithms, and Proximal Policy Optimization (PPO) can be used, as well as neural networks that approximate Q(s,a) or policy π(a|s). This application can employ online learning for model updates, or a combination of offline pre-training and online fine-tuning to balance stability and adaptability. Furthermore, this application can write the state-action-reward sequence into the experience pool or training pipeline to drive model updates.

[0060] As an example, when only one enterprise behavior prediction model, product matching model, and recommendation prediction model are designed, the above-mentioned rewards, states, and actions can all be relative to the comprehensive value of all upstream and downstream enterprises in the supply chain (which can be obtained through weighted averaging, etc.). That is, the recommendation effect of all upstream and downstream enterprises in the supply chain is used as the reward, the model parameters of the enterprise behavior prediction model, product matching model, and recommendation prediction model are used as the state, and the enterprise attribute characteristics and recommendation execution strategies of all upstream and downstream enterprises in the supply chain are used as the action. This application can also design dedicated enterprise behavior prediction models, product matching models, and recommendation prediction models for various financial products or for each upstream and downstream enterprise in the supply chain, forming a large model library, and the corresponding model can be called when executing steps S130 and S140.

[0061] Compared to existing recommendation methods, this application explicitly models and integrates the transaction relationships among multiple business entities and dynamic fund flows (such as cloud computing) into the entire process of feature engineering, prediction models, product matching, and marketing decision-making. The method proposed in this application has the following significant advantages: ① More comprehensive data utilization: It comprehensively utilizes multi-dimensional enterprise data from the supply chain, including supply chain relationship data, enterprise behavior data (including cloud communication data), and industry characteristic data. In subsequent feature modeling, demand forecasting, product matching, timing optimization, recommendation execution and feedback, and model update processes, it explicitly models hierarchical levels, core enterprise dependencies, and cash flow cycles. Compared with existing solutions that only use basic enterprise information, it can more comprehensively and accurately understand enterprise needs.

[0062] ② Improved prediction accuracy: By using an LSTM-attention model designed for enterprise behavior sequences and cash flow cycles, as well as feature and attention supervision in supply chain scenarios, it is possible to capture demand timing and key influencing factors.

[0063] ③ Establish an intelligent mechanism for determining the best marketing timing, which greatly improves the recommendation effect: The recommendation prediction model based on reinforcement learning can incorporate demand level and reach history into the status and reward, realize intelligent reach timing and channel selection, and provide response rate, conversion rate and other comparative data with fixed period or rule reach through reach record data.

[0064] ④ Provide highly personalized enterprise financial service recommendation solutions: From feature engineering to recommendation execution strategy decision-making, it relies on the supply chain level, enterprise master data wide table and enterprise behavior data, so as to provide differentiated recommendations based on the role and real-time status of upstream and downstream enterprises in the supply chain, which is different from general recommendation solutions.

[0065] Corresponding to the above method, the present invention also provides a financial product recommendation system for supply chain finance platforms. This system includes a computer device comprising a processor and a memory. The memory stores computer programs / instructions, and the processor executes the computer programs / instructions stored in the memory. When the computer programs / instructions are executed by the processor, the system performs the steps of the method described above. This financial product recommendation system can be installed on the same side as the supply chain finance platform, assisting the platform in recommending suitable financial services to its users (i.e., upstream and downstream enterprises in the supply chain).

[0066] like Figure 5 As shown, the financial product recommendation system mentioned in this application may include a data acquisition and preprocessing module, a feature engineering module, an intelligent recommendation core module, and a recommendation execution and feedback module. The data acquisition and preprocessing module is used to execute step S110, which can form a financial supply chain relationship, enterprise behavior data, and enterprise master data wide table through supply chain relationship graph construction, enterprise behavior data collection, and multi-source data fusion. The feature engineering module is used to execute step S120, which can extract the static / dynamic / network attribute features of enterprises. The intelligent recommendation core module is used to execute steps S130 and S140 and generate recommendation execution strategies. It can realize demand prediction, product matching, and timing optimization through an enterprise behavior prediction model based on LSTM-attention mechanism, a product matching model containing a hybrid matching algorithm, and a recommendation prediction model based on reinforcement learning algorithm, respectively. The recommendation execution and feedback module is used to reach the recommendation execution strategy through multiple channels, monitor the recommendation effect, and update the model. The process of collecting multi-dimensional enterprise data by the supply chain finance platform can be done using stream processing or batch processing. Feature extraction algorithms and model services (enterprise behavior prediction model, product matching model and recommendation prediction model) can be containerized and elastically scalable. Furthermore, the storage of financial supply chain relationships can use graph databases to support more complex network feature calculations.

[0067] This invention also provides a computer-readable storage medium storing a computer program / instructions 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.

[0068] This invention also provides a computer program product storing a computer program / instructions thereon, which, when executed by a processor, implements the steps of the aforementioned edge computing server deployment method. The computer program product can be a tangible product, such as random access memory (RAM), 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 product known in the art.

[0069] 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.

[0070] 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.

[0071] 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.

[0072] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, various modifications and variations of the embodiments of the present invention are possible. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for recommending a financial product for a supply chain finance platform, characterized in that, The method includes the following steps: Based on the multi-dimensional enterprise data of the supply chain finance platform, we can determine the financial supply chain relationship, enterprise behavior data and enterprise owner data wide table for upstream and downstream enterprises in the supply chain. The feature extraction algorithm is used to extract the enterprise attribute features of each upstream and downstream enterprise in the supply chain from the determined financial supply chain relationship, enterprise behavior data and enterprise master data wide table; wherein, the enterprise attribute features include enterprise semi-static attribute features, enterprise dynamic attribute features and enterprise static attribute features; The extracted enterprise attribute features are input into the enterprise behavior prediction model, and the output is the probability of the behavior of each upstream and downstream enterprise in the supply chain in a specific future period. The enterprise attribute features and the behavioral probabilities are input into a product matching model that pre-stores the attributes of various financial products. The model outputs the matching results of various financial products with each upstream and downstream enterprise in the supply chain, and then recommends specific financial products to upstream and downstream enterprises in the supply chain based on the matching results.

2. The method according to claim 1, characterized in that, The step of recommending specific financial products to upstream and downstream enterprises in the supply chain based on the matching results includes: obtaining a recommendation execution strategy based on the behavioral probability of each upstream and downstream enterprise in the supply chain and the matching results through a recommendation prediction model, and realizing recommendation outreach according to the recommendation execution strategy; wherein, the recommendation execution strategy is used to indicate whether to recommend specific types of financial products, recommendation time, and recommendation channels to specific upstream and downstream enterprises in the supply chain.

3. The method according to claim 2, characterized in that, After achieving recommendation outreach, the method further includes: acquiring outreach record data, and determining the recommendation effect corresponding to the recommendation execution strategy based on the outreach record data; wherein the evaluation indicators of recommendation effect include one or more of response rate, conversion rate, and satisfaction.

4. The method according to claim 3, characterized in that, The method further includes: using recommendation performance as a reward, the model parameters of the enterprise behavior prediction model, product matching model, and recommendation prediction model as states, and enterprise attribute features and recommendation execution strategies as actions, updating the enterprise behavior prediction model, product matching model, and recommendation prediction model through a reinforcement learning algorithm.

5. The method according to claim 1, characterized in that, The matching results between the various financial products and each upstream and downstream enterprise in the supply chain were obtained in the following way: The integrated enterprise attribute characteristics of each enterprise group are determined based on the enterprise attribute characteristics of each upstream and downstream enterprise in the supply chain; wherein each upstream and downstream enterprise in the supply chain belongs to an enterprise group. Determine the matching degree between the integrated enterprise attribute characteristics of each enterprise group and the various financial product attributes pre-stored in the product matching model, and use it as the preference strength of each enterprise group for various financial products; based on the behavioral probability of each upstream and downstream enterprise in the supply chain and the similarity between pairs of enterprises in the enterprise group to which the enterprise belongs, determine the collaborative filtering score of each upstream and downstream enterprise in the supply chain for various financial products, and then determine the collaborative filtering score of each enterprise group for various financial products. For each enterprise group, the matching results between the enterprise group and various financial products are determined based on the preference strength and the collaborative filtering score, thereby obtaining the matching results of all enterprises in the enterprise group with various financial products.

6. The method according to claim 1, characterized in that, The multi-dimensional enterprise data based on the supply chain finance platform determines the financial supply chain relationships, enterprise behavior data, and enterprise master data wide tables for upstream and downstream enterprises in the supply chain, including: The financial supply chain relationship is determined based on the supply chain hierarchy configuration between upstream and downstream enterprises in the supply chain finance platform. Based on the operational logs of the supply chain finance platform, the enterprise behavior data of each upstream and downstream enterprise in the supply chain is determined; and The business data from the supply chain finance platform is used to create a wide table of enterprise master data for each upstream and downstream enterprise in the supply chain.

7. The method according to claim 1, characterized in that, The enterprise's semi-static attribute features include the enterprise's supply chain location information; the enterprise's dynamic attribute features include one or more of the following: electronic credit certificate activity, payment cycle, and financing frequency; and the enterprise's static attribute features include enterprise credit assessment. The enterprise behavior prediction model is constructed based on long short-term memory networks and attention mechanisms. as well as Financial product attributes include basic identification attributes, scope of application attributes, and business rule attributes.

8. A financial product recommendation system for a supply chain finance platform, comprising a processor, a memory, and a computer program / instructions stored in the memory, characterized in that, The processor is configured to execute the computer program / instructions, and when the computer program / instructions are executed, the system implements the steps of the method as described in any one of claims 1 to 7.

9. A computer-readable storage medium having a computer program / instructions stored thereon, characterized in that, When the computer program / instructions are executed by the processor, they implement the steps of the method as described in any one of claims 1 to 7.

10. A computer program product comprising a computer program / instructions, characterized in that, When the computer program / instructions are executed by the processor, they implement the steps of the method as described in any one of claims 1 to 7.

Images