Intelligent allocation system for cross-border payment channels based on transaction type

Abstract

The present invention belongs to the technical field of cross-border payment settlement analysis and management, and discloses a cross-border payment channel intelligent allocation system based on transaction type; comprising: a type identification module: used for collecting transaction order data, extracting features from the transaction order data to obtain transaction order features, using the transaction order features as input of a type identification model, and obtaining a transaction type label; a risk assessment module: used for collecting risk data, extracting features from the collected risk data to obtain risk assessment features, using the risk assessment features as input of a risk analysis model, and obtaining a risk assessment score; based on real-time data collection technology and stream computing technology, the payment process is monitored in real time, and when the risk level exceeds a preset risk threshold, an early warning mechanism is immediately triggered; the present invention dynamically adjusts the selection strategy of the payment channel according to factors such as real-time market conditions and performance changes of the payment channel, and provides a precise payment channel allocation plan.

Description

Technical Field

[0001] The present invention relates to the field of cross-border payment settlement analysis and management technology, and more specifically, to a cross-border payment channel intelligent allocation system based on transaction type. Background Art

[0002] As globalization deepens, cross-border trade and investment activities have experienced explosive growth, and the demand for cross-border payments has risen sharply. Cross-border payments involve complex factors such as currency exchange, financial regulatory systems, and diverse payment habits and cultural backgrounds in different countries or regions, which makes cross-border payments face many challenges and pain points.

[0003] Traditional cross-border payment channel allocation relies on manual experience or simple fixed rules; manual experience allocation methods are highly subjective, inefficient, and difficult to adapt to large-scale and diversified cross-border transaction needs. Fixed rules are difficult to adapt quickly, which may cause enterprises or individuals to face higher costs, lower efficiency, and even security loopholes in the cross-border payment process. Cross-border payment risks are complex and diverse and have far-reaching impacts. For example, exchange rate fluctuations will lead to uncertainty in payment amounts. If the exchange rate fluctuation factor is not fully considered in the selection of payment channels, the transaction parties may suffer additional exchange losses. With the increasing variety of cross-border transaction types, such as B2B large-scale trade, B2C small-scale retail, C2C personal transfers, etc., the requirements for payment channels for each transaction type are significantly different. However, the existing payment channel allocation mechanism is difficult to accurately meet the differentiated needs of different transaction types, resulting in a poor overall cross-border payment experience and hindering the further development of cross-border trade and exchanges.

[0004] The Chinese patent application with publication number CN118014581A discloses a settlement analysis management method, system and storage medium based on cross-border payment channels, including: the customer sends a transaction request, the platform receives and forwards it to the service provider, the service provider generates an electronic bill and feeds it back to the platform, and the platform feeds it back to the customer and generates a settlement order for storage. At the same time, important data of each link, external security attack data, etc. are collected, and the importance of data and the danger of attack are evaluated respectively, and a risk assessment model formula is established to form a risk assessment standard for data. In terms of transaction process, this invention realizes the orderly transmission and storage of information between the platform, service providers and customers, making transaction information processing efficient and standardized. At the data collection and evaluation level, by collecting multi-link data and security attack data, the importance of data and the danger of attack can be accurately assessed. The established risk assessment model and standard help to prevent risks in advance and ensure transaction security. Overall, it can improve the stability and reliability of the transaction system, enhance the trust of all parties, promote the healthy and stable development of transaction business, and provide strong data support and security guarantee for enterprise operation management and decision-making.

[0005] Although the above method can meet most scenarios, research and practical application of the above method and existing technology have found that the above method and existing technology have at least the following defects:

[0006] Insufficient attention is paid to transaction types, and it is impossible to accurately match the most suitable payment channel according to the specific characteristics of the transaction, such as different modes such as B2B and B2C, and the nature of goods or services, resulting in poor adaptability. There is a lack of comprehensive and systematic evaluation and response measures for the risks that different transaction types may face in different payment channels, such as exchange rate risk and credit risk. The evaluation of other performance indicators of payment channels, such as payment success rate and system stability, is not comprehensive and in-depth enough. Moreover, the analysis is based on preset models and algorithms, lacking the ability to dynamically learn and adapt to real-time data, and has limited adaptability to the complex and changing cross-border payment environment.

[0007] In view of this, the present invention proposes a cross-border payment channel intelligent allocation system based on transaction type to solve the above problems. Summary of the invention

[0008] In order to overcome the above-mentioned defects of the prior art and to achieve the above-mentioned purpose, the present invention provides the following technical solution: a cross-border payment channel intelligent allocation system based on transaction type, comprising:

[0009] The first collection module is used to collect transaction order data; the transaction order data includes contract text data and commodity text data; the contract text data includes the subject identity code and transaction amount in the contract text; the commodity text data includes the commodity name text and commodity detail description text in the commodity description text, and the commodity detail description text includes the commodity name text, commodity type text, commodity quantity text and commodity validity period text;

[0010] The first analysis module is used to extract commodity text data and obtain commodity text features;

[0011] Type recognition module: used to use transaction order data and product text features as inputs to the type recognition model to obtain transaction type labels;

[0012] The second collection module: used to collect risk data;

[0013] Risk assessment module: used to use risk data as input to the risk analysis model to obtain risk assessment scores;

[0014] The third acquisition module: used to collect threshold analysis data;

[0015] Threshold analysis module: used to use the threshold analysis data as the input of the threshold analysis model to obtain the amount threshold, and the amount threshold is adaptively adjusted according to the threshold analysis data collected in real time;

[0016] The fourth collection module is used to collect channel analysis data of each payment channel;

[0017] The channel analysis module is used to use the channel analysis data as the input of the channel analysis model to obtain the channel score of each payment channel, and to concatenate the obtained channel scores of each payment channel to obtain a channel score set;

[0018] Channel selection module: used to use the transaction type labels, risk assessment scores, amount thresholds and channel score sets of different payment channels as inputs to enhance the channel selection model to obtain the best payment channel.

[0019] Furthermore, the threshold analysis data includes user payment data, market commodity data, historical exchange rate data and channel statistical data of payment channels; user payment data is the average payment amount of x times; market commodity data is the average daily renewal amount of p competitors in the same industry in the market; historical exchange rate data is the average exchange rate change of both parties to the transaction within a preset time; channel statistical data includes handling fee rate and payment limit; handling fee rate is the proportion of handling fees charged by different payment channels for each transaction; payment limit is the limit amount of different payment channels on single transaction and daily transaction amount.

[0020] Furthermore, the channel analysis data includes transaction time data, historical transaction data, busyness data and channel performance data; the transaction time data includes the current user time zone and seller time zone; the historical transaction data includes the number of transactions and cumulative amounts of different payment channels within a preset time; the busyness data includes the length of the transaction processing queue in each payment channel; the channel performance data includes transaction success rate, processing time, security performance data and stability data; the transaction success rate is the ratio of the number of successful transactions to the total number of transactions for each payment channel within a preset time; the processing time is the sum of the time required for the payment channel to complete the payment from the time the user sends a payment request to the time the seller receives the payment amount; the security performance data includes the number of security occurrences and security technical data; the number of security occurrences is the number of security incidents that occur during the payment process; the security technical data is the number of types of payment methods adopted by the payment channel; the stability data is the number of failures and maintenance time of the payment channel within a preset time, and the maintenance time is the average duration of maintenance for each failure.

[0021] Furthermore, the data of the enhanced channel selection model is initialized as follows:

[0022] Get the transaction type labels, risk assessment scores, amount thresholds, and channel score sets of different payment channels, concatenate them to obtain the state space, and use vector Indicates that the optional payment channels are aggregated into the same set to obtain the action space, and the vector Indicates that a Q value table is created for the state space and action space, and the reward function is defined as follows:

[0023] ;

[0024] in, is the reward value; Is the transaction success rate; For processing time; is the handling fee rate; , and They are the weights corresponding to the transaction success rate, processing time and handling fee rate respectively;

[0025] Methods for creating Q-value tables corresponding to state space and action space include:

[0026] The number of rows in the Q value table corresponds to the size of the state space, and the number of columns corresponds to the size of the action space. The Q value of each element in the Q value table is Initialize with a random number or 0. Indicates in status Take action The expected cumulative reward value is value.

[0027] Further, the transaction type labels, risk assessment scores, amount thresholds, and channel score sets of the different payment channels are used as inputs of the enhanced channel selection model to obtain the initially selected payment channels;

[0028] The training method of the enhanced channel selection model includes:

[0029] Pre-collect M groups of channel training data, the channel training data including channel data features and payment channels corresponding to the channel data features; the channel data features include transaction type labels, risk assessment scores, amount thresholds and channel score sets;

[0030] Each set of channel training data is used as input of the enhanced channel selection model, the enhanced channel selection model uses the payment channel corresponding to each set of channel data features as output, and uses the actual payment channel corresponding to each set of channel data features as the prediction target; minimizing the sum of prediction errors of all payment channels is used as the training target; the enhanced channel selection model is trained until the sum of prediction errors reaches convergence and the training is stopped; the enhanced channel selection model is a deep neural network model;

[0031] The loss function value of the enhanced channel selection model is the sum of mean square error and reward loss;

[0032] Prediction error formula:

[0033] ;

[0034] ;

[0035] ;

[0036] in, For reward-related loss items; is a positive integer coefficient; is the reward value; is a positive number that makes the distribution non-zero; m is the mth group of channel data features; M is the number of groups of channel data features; is the predicted payment channel corresponding to the mth group of channel data features; is the payment channel corresponding to the mth group of channel data features.

[0037] Further, the transaction is executed according to the payment channel initially selected by the enhanced channel selection model, and the reward corresponding to the action is calculated according to the transaction result and the reward function. , based on rewards Update the loss function value of the enhanced channel selection model; according to the reward , use the reinforcement learning strategy to update the Q value in the Q value table. The update formula is as follows:

[0038] ;

[0039] in, Q value after the current state update; is the Q value before the current state is updated; is the learning rate; For updated rewards; is the discount factor; is the current state; The currently selected payment channel; For the next state; It is the optimal payment channel among all possible payment channels in the next state; It is the maximum Q value corresponding to all possible payment channels in the next state, that is, the maximum expected cumulative reward that can be obtained in all optional payment channels after transferring from the current state to the next state;

[0040] The Q-value table is trained and updated by using a reinforcement learning strategy. The Q-value in the Q-value table is continuously updated. When the update amplitude of the Q-value, that is, the difference between the updated Q-value and the Q-value before the update, reaches a preset value, the update is stopped, and the action corresponding to the element with the largest Q-value in the Q-value table at this time is selected, and the payment channel corresponding to the action is obtained. The selected payment channel is used as the best payment channel finally selected.

[0041] Furthermore, the risk data includes user risk data, seller risk data and product risk data; the user risk data includes historical payment data, user credit rating and account security data; the historical payment data includes the number of overdue payments by users, the number of payment failures, the number of payments per unit time and the average payment amount per unit time; the account security data includes the number of times the user account has been stolen, the number of password modifications and the number of abnormal logins; the seller risk data includes the seller's credit rating, business stability and compliance records; the business stability includes the debt-to-asset ratio and cash flow change data, and the cash flow change data includes cash flow inflow data and cash flow outflow data; the compliance record is the number of times the seller has been punished by the regulatory authorities for violations; the product risk data includes product sales volume, product feedback data and product update volume within a preset time, and the product update volume is the number of new products launched; wherein the product feedback data includes product cancellation volume, positive review rate, medium review rate and negative review rate; the product cancellation volume is the number of users who have ordered the current product and have canceled this product.

[0042] Furthermore, the method for obtaining the subject identity code includes:

[0043] If the transaction party includes an enterprise, the enterprise's unified social credit code is obtained as the enterprise identity code. If the transaction party includes an individual, the user's ID number or other valid ID number is obtained as the individual identity code; the identity codes of both parties to the transaction are concatenated as the subject identity code.

[0044] Furthermore, the method for obtaining the commodity text feature includes:

[0045] Obtain the product name text and product details description text in the product description text, use the Chinese word segmentation tool to segment the product name text and product details description text respectively, remove duplicate words, obtain the first u words in the product name text and the first v words in the product details description text respectively, use the pre-trained word vector model to convert the obtained words into word vectors, use the word vectors as the input of the text extraction model, and output the product text features;

[0046] The training methods for text extraction models include:

[0047] Collect a group A of text training data in advance. The text training data includes word vectors and product text features corresponding to the word vectors.

[0048] Each set of text training data is used as input of a text extraction model. The text extraction model uses the commodity text features corresponding to each set of word vectors as output and the actual commodity text features corresponding to each set of word vectors as prediction targets. Minimizing the sum of prediction errors of all commodity text features is used as a training target. The text extraction model is trained until the sum of prediction errors reaches convergence and training is stopped. The text extraction model is a deep neural network model.

[0049] The text extraction model loss function value is the mean square error.

[0050] Prediction error formula: , a is the ath group of word vectors; A is the number of groups of word vectors; is the product text feature corresponding to the a-th group of word vectors; is the actual product text feature corresponding to the a-th group of word vectors.

[0051] Furthermore, the training method of the type recognition model includes:

[0052] Collecting group B order training data in advance, the order training data includes transaction order features and transaction type labels, wherein the transaction order features include transaction order data and commodity text features;

[0053] Each set of order training data is used as input of a type recognition model. The type recognition model uses the transaction type label corresponding to each set of transaction order data and the commodity text features as output, and uses the actual transaction type label corresponding to each set of transaction order data and the commodity text features as a prediction target; minimizing the sum of prediction errors of all transaction type labels is used as a training target; the type recognition model is trained until the sum of prediction errors reaches convergence and training is stopped; the type recognition model is a deep neural network model;

[0054] The type recognition model loss function value is the mean square error;

[0055] Prediction error formula: , b is the bth group of transaction order features; B is the number of groups of transaction order features; is the predicted transaction type label corresponding to the transaction order features of group b; is the actual transaction type label corresponding to the transaction order features of group b.

[0056] The technical effects and advantages of the cross-border payment channel intelligent allocation system based on transaction type of the present invention are as follows:

[0057] The present invention takes the transaction type as one of the key factors, and intelligently selects the most suitable cross-border payment channel according to the specific needs of different transaction types. At the same time, it comprehensively considers multiple performance indicators of the payment channel, comprehensively evaluates various risks of users, sellers and commodities, and can effectively reduce cross-border payment risks; it also considers factors such as payment success rate, system stability, payment speed, channel availability, etc., so as to more comprehensively evaluate the performance of the payment channel and provide users with more reliable and efficient payment channel selection; it also combines the use of artificial intelligence, machine learning and other technologies, continuously learns and accumulates matching experience of different transaction types and payment channels, and can dynamically adjust the selection strategy of the payment channel according to real-time market conditions, exchange rate fluctuations, performance changes of payment channels and other factors, and provide a more intelligent and accurate payment channel allocation plan. BRIEF DESCRIPTION OF THE DRAWINGS

[0058] Figure 1 This is a system block diagram mentioned in Embodiment 1 of the present invention;

[0059] Figure 2 This is a schematic diagram of the risk warning process mentioned in Example 1 of the present invention;

[0060] Figure 3 This is a schematic diagram of the enhanced channel selection model mentioned in Example 1 of the present invention;

[0061] Figure 4 This is a schematic diagram of the enhanced channel selection model mentioned in Example 2 of the present invention. DETAILED DESCRIPTION

[0062] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0063] Example 1

[0064] See also Figure 1 As shown, the cross-border payment channel intelligent allocation system based on transaction type described in this embodiment includes:

[0065] The first collection module is used to collect transaction order data; the transaction order data includes contract text data and commodity text data; the contract text data includes the subject identity code and transaction amount in the contract text.

[0066] Methods for obtaining the subject identity code include:

[0067] If the transaction party includes an enterprise, the enterprise's unified social credit code is obtained as the enterprise identity code. If the transaction party includes an individual, the user's ID number or other valid ID number is obtained as the individual identity code; the identity codes of both parties to the transaction are concatenated as the subject identity code.

[0068] The collection of subject identity codes can determine the nature of the transaction subject. For example, if the code shows that both parties are corporate codes, then it is a B2B transaction; if the code shows that one party is a corporate code and the other party is a personal identity code, then it is a B2C or C2B transaction; if the code shows that both parties are personal identity codes, then it is a C2C transaction. Different types of transaction subjects often correspond to different transaction models and types of goods or services, which helps to preliminarily determine the type of transaction. Transaction subjects with different identities are suitable for different payment channels. For corporate users, they may prefer commonly used payment methods such as bank transfers and commercial bills, because these payment channels can provide more complete financial records and fund supervision functions. Individual users are more suitable for convenient payment methods such as e-wallets and credit cards.

[0069] Collecting transaction amounts can help identify transaction types. For example, transactions with non-fixed amounts and charged by usage or time are service transactions, such as communication traffic charges. Different payment channels can be selected according to different transaction amounts. For example, large-value transactions usually require payment channels with high security and a complete fund settlement process, such as the bank's large-value payment system or a third-party payment platform with fund guarantees, to reduce payment risks. For small-value transactions, payment methods with low handling fees and simple operations can be selected, such as the small-value password-free payment function in mobile payments, to improve payment efficiency.

[0070] The commodity text data includes the commodity name text and the commodity details description text in the commodity description text, which can be directly obtained through the transaction platform; the commodity name text can intuitively reflect the items or services involved in the transaction, such as "traffic package" indicates that it is a traffic renewal transaction. The commodity details description text helps to identify the transaction type more accurately. The commodity details description text contains the commodity name text, commodity type text, commodity quantity text and commodity validity period text; commodity type text such as the type of traffic (such as 4G / 5G traffic), commodity quantity text such as the amount of traffic (such as 10GB, 50GB, etc.), commodity validity period text such as the validity period of traffic (such as valid within 30 days after successful renewal) and other information can be determined to be a service transaction. Some commodities or services are suitable for specific payment channels due to their characteristics. For example, for traffic renewal services, it is more appropriate to use online payment channels for instant payment directly, so that service authorization can be obtained quickly. At the same time, the transaction frequency of commodities will also affect the choice of payment channels. If it is a frequently purchased commodity, such as a traffic package, users may choose a quick mobile payment method.

[0071] The first analysis module is used to extract the commodity text data to obtain commodity text features, which include commodity name, commodity type, commodity quantity and commodity validity period. The method of obtaining commodity text features includes:

[0072] Get the product name text and product details description text in the product description text, use the Chinese word segmentation tool to segment the product name text and the product details description text respectively, remove duplicate words, and get the first u words in the product name text and the first v words in the product details description text respectively (u can be preferably 15, indicating the first 15 words in the product name; v can be preferably 100, indicating the first 100 words in the product details description text). Use the pre-trained word vector model to convert the obtained words into word vectors, use the word vectors as the input of the text extraction model, and output the product text features.

[0073] The training methods for text extraction models include:

[0074] A group of text training data is collected in advance. The text training data includes word vectors and commodity text features corresponding to the word vectors.

[0075] Each group of text training data is used as the input of the text extraction model. The text extraction model uses the product text features corresponding to each group of word vectors as output, and the actual product text features corresponding to each group of word vectors as prediction targets; minimizing the sum of prediction errors of all product text features is used as the training target; the text extraction model is trained until the sum of prediction errors reaches convergence, and the training is stopped; the text extraction model is a deep neural network model.

[0076] The text extraction model loss function value is the mean square error.

[0077] Prediction error formula: , a is the ath group of word vectors; A is the number of groups of word vectors; is the product text feature corresponding to the a-th group of word vectors; is the actual product text feature corresponding to the a-th group of word vectors.

[0078] Type recognition module: The transaction order data and product text features are used as inputs to the type recognition model to obtain transaction type labels. Transaction type labels include B2B, B2C, and C2C.

[0079] The training methods for the type recognition model include:

[0080] A group B of order training data is collected in advance, and the order training data includes transaction order features and transaction type labels, wherein the transaction order features include transaction order data and commodity text features.

[0081] Each group of order training data is used as the input of the type recognition model. The type recognition model takes the transaction type label corresponding to each group of transaction order data and the commodity text features as the output, and the actual transaction type label corresponding to each group of transaction order data and the commodity text features as the prediction target; minimizing the sum of the prediction errors of all transaction type labels is used as the training target; the type recognition model is trained until the sum of the prediction errors reaches convergence and the training is stopped; the type recognition model is a deep neural network model.

[0082] The type recognition model loss function value is the mean square error;

[0083] Prediction error formula: , b is the bth group of transaction order features; B is the number of groups of transaction order features; is the predicted transaction type label corresponding to the transaction order features of group b; is the actual transaction type label corresponding to the transaction order features of group b.

[0084] like Figure 2 As shown, the second collection module is used to collect risk data; the risk data includes user risk data, seller risk data and product risk data;

[0085] User risk data is obtained through trading platform statistics, including historical payment data, user credit rating and account security data; historical payment data includes the number of overdue payments, the number of payment failures, the number of payments per unit time (such as 1 year) and the average payment amount per unit time; account security data includes the number of times the user account has been stolen, the number of password changes and the number of abnormal logins;

[0086] Historical payment data can assess the user's ability to pay. If the user's historical payment record shows that he often makes high-value payments and the payment frequency is stable, this indicates that he has strong payment ability; on the contrary, if the payment amount is low and the payment interval is long, it may mean that the payment ability is limited. This assessment helps the risk assessment module predict the probability of successful payment of users in future transactions. By assessing the user's payment ability, you can choose a payment channel that matches the user. For example, for users with strong payment ability and who are accustomed to using electronic payment, you can recommend electronic payment channels with fast payment processing capabilities and high payment success rates (such as fast mobile payment methods, etc.). For users with limited payment ability or unstable payment habits, you can provide some payment channel options with flexible payment methods (such as installment payments, prepaid payments, etc.).

[0087] User credit rating is a quantitative assessment of user credit risk. Users with high credit ratings usually have good credit records, indicating that they have been able to fulfill debts and contractual obligations on time in past transactions and financial activities, and their probability of default is low. The risk assessment module can assign corresponding risk levels to users based on user credit ratings in order to more accurately assess transaction risks. For users with high credit ratings, relatively flexible and simple payment channels can be selected. For example, users with good user credit ratings can enjoy convenient services such as password-free payment and fast payment channels. For users with low user credit ratings, in order to reduce risks, they may need to choose payment channels with strict identity verification, fund guarantee or credit assessment mechanisms, such as some third-party payment platforms that require additional identity proof or credit authorization.

[0088] Account security data is the basis for identifying account security risks. If a user's account has security problems, such as being stolen for fraudulent transactions, then in future transactions, the account is more likely to be threatened again. The risk assessment module can use this data to incorporate account security risks into the overall risk assessment system. Based on the user's account security data, you can choose a payment channel with a high security level. If the user's account security risk is high, such as frequent abnormal logins, you should choose a payment channel with multiple authentication (such as passwords, fingerprint recognition, SMS verification codes, etc.), secure encryption technology (such as SSL / TLS encryption protocol) and real-time risk monitoring functions. These payment channels can effectively prevent account theft and ensure the safety of funds.

[0089] Seller risk data is obtained based on the seller's public financial reports or financial database queries, including the seller's credit rating, business stability and compliance records; business stability includes asset-liability ratio and cash flow change data, and cash flow change data includes cash flow inflow data and cash flow outflow data; compliance records are the number of times the seller has been punished by the regulatory authorities for violations.

[0090] Seller credit ratings provide a quantitative measure of the seller's default risk. Sellers with high credit ratings usually have a good record of fulfilling contracts in past transactions, such as on-time delivery, providing goods or services that meet quality standards, and promptly handling after-sales issues. The risk assessment module can assign corresponding default risk levels to sellers based on these seller credit ratings, thereby more accurately assessing transaction risks. For sellers with high credit ratings, you can choose a payment channel that is relatively flexible and has a simple transaction process. Because such sellers have lower credit risks, payment channel providers have a high degree of trust in them. For sellers with lower credit ratings, in order to reduce risks, you may need to choose a payment channel with a fund guarantee function, such as a third-party payment platform. Under this payment channel, the buyer's funds are first paid to the platform, and the funds will be transferred to the seller only after confirming that the seller has fulfilled the transaction obligations, thereby protecting the buyer's rights and interests.

[0091] By analyzing the business stability of the seller, the risk assessment module can assess the seller's ability to continue operating. Stable sellers are more likely to fulfill their contractual obligations in the long-term transaction process, while unstable sellers may be unable to supply goods or provide services normally due to financial difficulties, market competition pressure or internal management problems, thereby increasing transaction risks. Stable sellers can choose a relatively wide range of payment channels, including some fixed payment methods under long-term cooperation agreements, such as regular bank transfer settlements between enterprises. These payment channels are usually based on trust and stable transaction relationships between the two parties. However, for unstable sellers, in order to cope with possible business changes, such as tight funds and business transformation, it is more appropriate to choose flexible payment channels. For example, a comprehensive payment platform that supports multiple payment methods (such as cash, card swiping and electronic payment, etc.), or a payment channel that can provide flexible payment options such as installment payment and deferred payment, can better meet the needs of both parties when the seller's business fluctuates.

[0092] Compliance records can help identify legal risks faced by sellers. If sellers violate industry regulations, this will increase transaction uncertainty and potential legal liability. The risk assessment module can determine whether the seller has legal risks based on compliance records and include them in the overall risk assessment. The seller's compliance record directly affects the legitimacy of the payment channel. If the seller has compliance issues, formal payment channels will usually refuse to cooperate with them. Therefore, when choosing a payment channel, it is necessary to ensure that the seller's business activities comply with legal and regulatory requirements and choose a legal and compliant payment channel.

[0093] Product risk data includes product sales volume, product feedback data (such as product cancellation volume, positive review rate, neutral review rate, and negative review rate, etc.) and product update volume within a preset time (the number of new products launched); product cancellation volume refers to the number of users who have ordered the current product and canceled the order;

[0094] The sales volume of a product can directly reflect the degree of acceptance of the product in the market; stable sales volume means that the product has a relatively stable market demand and its risk is relatively low; while products with large sales fluctuations may be affected by multiple factors such as market factors, product quality or competitive environment, and the risk assessment module can identify them as high-risk products with uncertainty. Feedback data such as product cancellation volume, positive review rate, medium review rate and negative review rate are key indicators for measuring user satisfaction; high cancellation volume and negative review rate indicate that the product may have quality problems, functions that do not meet expectations or services that are not in place, and these products should be considered high-risk categories in risk assessment. The number of product updates in the past three months and the corresponding feedback data are helpful in judging the product's sustainable development capabilities. Frequent and effective updates can enhance product competitiveness and reduce risks; but if the feedback after the update is poor, such as new problems with the updated product function or users do not approve of the updated content, this will increase product risks.

[0095] For products with lower risks (such as stable sales and high praise rate), you can choose a relatively flexible payment channel with simple procedures. For high-risk products (such as low sales and many negative reviews), in order to reduce consumers' payment risks, you may need to choose a payment channel with a guarantee function. Product updates will also affect the choice of payment channels. If the product is updated frequently and the feedback is good, the payment channel should be adaptable to support new functions or new service charges. For example, for a product that continuously launches new value-added services, the payment channel needs to be able to easily handle multiple payment methods such as subscription fees and value-added service fees. For products with poor feedback after updates, the payment channel may need to consider suspending certain automatic renewal functions that may cause disputes to avoid consumer losses and subsequent disputes.

[0096] Risk assessment module: used to use risk data as input to the risk analysis model to obtain a risk assessment score. The larger the risk assessment score, the higher the payment risk, and the smaller the risk assessment score, the lower the payment risk. The risk assessment module compares the risk score with the preset risk threshold. When the risk score exceeds the preset risk threshold, the early warning mechanism is triggered and relevant information is recorded to take countermeasures. If the risk score does not exceed the preset risk threshold, no measures are taken.

[0097] The training methods of risk analysis models include:

[0098] A group D of risk training data is collected in advance, and the risk training data includes risk data and risk assessment scores corresponding to the risk data.

[0099] Each set of risk training data is used as the input of the risk analysis model. The risk analysis model takes the risk assessment score corresponding to each set of risk data as the output, and the actual risk assessment score corresponding to each set of risk data as the prediction target; minimizing the sum of the prediction errors of all risk assessment scores is used as the training target; the risk analysis model is trained until the sum of the prediction errors reaches convergence and the training is stopped; the risk analysis model is a deep neural network model.

[0100] The loss function value of the risk analysis model is mean square error;

[0101] Prediction error formula: , d is the dth group of risk data; D is the number of risk data groups; is the predicted risk assessment score corresponding to the risk data of group d; is the actual risk assessment score corresponding to the risk data of group d.

[0102] The third collection module is used to collect threshold analysis data; the threshold analysis data includes user payment data, market commodity data, historical exchange rate data and channel statistics of payment channels; user payment data is the average payment amount of x times.

[0103] Collecting user payment data can evaluate the user's ability to pay. For example, if the average payment amount of the user is low, it can be inferred that the user is less receptive to high-amount payments. Then the threshold analysis module can determine a reasonable amount threshold, and payments below this threshold may be considered to be more acceptable to the user. It is also possible to further select payment channels that match the user's ability to pay. For example, for users with limited payment ability, some payment channels that support installment payments or prepaid payments can be provided to meet their needs.

[0104] Market commodity data is the average daily renewal amount of p competitors in the same industry in the market. Market commodity data provides the price distribution of products or services in the same industry; by analyzing this data, the threshold analysis module can determine the position of the seller's commodity price in the market. With reference to market transaction data, you can choose a payment channel that is widely accepted and competitive in the market.

[0105] Historical exchange rate data is the average exchange rate change between the two parties in a preset period of time. Historical exchange rate data is crucial for cross-border transactions. Exchange rate fluctuations can affect the prices of goods or services denominated in different currencies, which in turn affects the payment amount threshold. Payment channels that can provide good exchange rate services based on historical exchange rate data. Some payment channels may provide more favorable currency exchange rates or provide exchange rate locking functions, which are very helpful in reducing cross-border transaction costs and exchange rate risks.

[0106] Channel statistics include handling fee rates and payment limits; handling fee rates are the proportion of handling fees charged by different payment channels for each transaction, obtained through trading platform statistics; payment limits are the limits on single transactions, daily transaction amounts, etc. by different payment channels. By analyzing channel statistics, the threshold analysis module can calculate the costs of different payment channels within different amount ranges. By comprehensively considering channel statistics, you can choose a payment channel that balances risks and benefits.

[0107] Threshold analysis module: used to use threshold analysis data as input of the threshold analysis model to obtain the amount threshold paid by the user. The amount threshold is adaptively adjusted according to the threshold analysis data collected in real time.

[0108] The training methods of the threshold analysis model include:

[0109] F groups of threshold training data are collected in advance, and the threshold training data include threshold analysis data and amount thresholds corresponding to the threshold analysis data.

[0110] Each set of threshold training data is used as the input of the threshold analysis model. The threshold analysis model takes the monetary amount threshold corresponding to each set of threshold analysis data as the output, and takes the actual monetary amount threshold corresponding to each set of threshold analysis data as the prediction target; minimizing the sum of prediction errors of all monetary amount thresholds is used as the training target; the threshold analysis model is trained until the sum of prediction errors reaches convergence and the training is stopped; the threshold analysis model is a deep neural network model.

[0111] The threshold analysis model loss function value is the mean square error;

[0112] Prediction error formula: , f is the fth group of threshold analysis data; F is the number of groups of threshold analysis data; The monetary threshold corresponding to the threshold analysis data of group f; The actual amount threshold corresponding to the fth group of threshold analysis data.

[0113] The fourth collection module is used to collect channel analysis data for each payment channel. The channel analysis data includes transaction time data, historical transaction data, busyness data and channel performance data. The transaction time data includes the current user time zone and seller time zone, which are obtained through networking. By collecting the user time zone and seller time zone, the transaction heat and busyness of the payment channel in different time periods can be analyzed, thereby optimizing the payment channel arrangement.

[0114] Historical transaction data includes the number of transactions and cumulative amounts of different payment channels within a preset period of time, which are obtained through statistics of the transaction platform. Collecting historical transaction data can provide an intuitive understanding of the popularity of each payment channel. Based on the number of transactions and cumulative amounts of different payment channels in the historical transaction data, it is also possible to determine the applicability of each payment channel to specific business scenarios, and further select payment channels that perform well in business scenarios based on business characteristics.

[0115] The busyness data includes the length of the transaction processing queue in each payment channel, which is obtained by connecting to the payment provider corresponding to the payment channel. Collecting busyness data can intuitively show the channel processing pressure and provide a reference for the allocation and optimization of payment channels.

[0116] Channel performance data is obtained through trading platform statistics, including transaction success rate, processing time, security performance data and stability data; transaction success rate is the ratio of the number of successful transactions to the total number of transactions within a preset time for each payment channel; processing time is the sum of the time required for the payment channel to complete the payment from the time the user sends a payment request to the time the seller receives the payment amount; for example, the time interval from the consumer initiating a payment request to the payment system returning the payment result, and the time the merchant receives the payment, etc., excluding the time consumed by non-payment channels such as user consumption time. Security performance data includes the number of security events and security technology data; the number of security events is the number of security events such as fraud, theft, and information leakage during the payment process; security technology data is the number of payment methods used by the payment channel (such as payment through passwords, fingerprints, and facial recognition, the number of types is 3); stability data is the number of failures and maintenance time of the payment channel within a preset time, and the maintenance time is the average length of maintenance for each failure. Collecting channel performance data plays an important role in selecting payment channels. The transaction success rate can intuitively show the reliability of the payment channel, and channels with a high success rate can reduce the risk of transaction failure. Processing time can determine payment efficiency, and channels with fast processing can improve user experience. Security performance data is the key to ensuring fund security, and channels with high security factors can avoid fund losses. Stability data reflects whether the channel can continue to work stably, and stable channels can ensure smooth transactions. Combining these data, you can select payment channels that better meet business needs, have good user experience, and are safe and stable.

[0117] The channel analysis module is used to use the channel analysis data as the input of the channel analysis model to obtain the channel score of each payment channel, and to concatenate the channel scores of each payment channel to obtain a channel score set.

[0118] The training methods of the channel analysis model include:

[0119] H groups of analysis training data are collected in advance, and the analysis training data includes channel analysis data and channel scores corresponding to the channel analysis data. The channel analysis data includes transaction time data, historical transaction data, busyness data and channel performance data.

[0120] Each set of analysis training data is used as the input of the channel analysis model. The channel analysis model takes the channel score corresponding to each set of channel analysis data as the output, and the actual channel score corresponding to each set of channel analysis data as the prediction target; minimizing the sum of the prediction errors of all channel scores is used as the training target; the channel analysis model is trained until the sum of the prediction errors reaches convergence and the training is stopped; the channel analysis model is a deep neural network model.

[0121] The loss function value of the channel analysis model is mean square error;

[0122] Prediction error formula: , h is the hth group of channel analysis data; H is the number of groups of channel analysis data; is the predicted channel score corresponding to the hth group of channel analysis data; is the actual channel score corresponding to the hth group of channel analysis data.

[0123] Channel selection module: used to take the transaction type label, risk assessment score, amount threshold and channel score set as input to strengthen the channel selection model to obtain the best payment channel.

[0124] When building an enhanced channel selection model, in order to ensure that the model can run accurately and obtain reliable results, the enhanced channel selection model first needs to initialize the data.

[0125] Reference Figure 3 , the data initialization of the enhanced channel selection model is as follows:

[0126] Get the transaction type labels, risk assessment scores, amount thresholds, and channel score sets of different payment channels, concatenate them to obtain the state space, and use vector Indicates that the optional payment channels are aggregated into the same set to obtain the action space, and the vector Indicates that a Q value table is created for the state space and action space, and the reward function is defined as follows:

[0127] ;

[0128] in, is the reward value; is the transaction success rate; For processing time; is the handling fee rate; , and They are the weights corresponding to transaction success rate, processing time and handling fee rate respectively.

[0129] Methods for creating Q-value tables corresponding to state space and action space include:

[0130] The number of rows in the Q value table corresponds to the size of the state space, and the number of columns corresponds to the size of the action space. The Q value of each element in the Q value table is Initialize it to a random number or 0. Indicates in status Take action Expected cumulative reward.

[0131] The transaction type label, risk assessment score, amount threshold and channel score set are used as inputs of the enhanced channel selection model to obtain the preliminarily selected payment channels.

[0132] The training methods for strengthening the channel selection model include:

[0133] M groups of channel training data are collected in advance, and the channel training data include channel data features and payment channels corresponding to the channel data features; the channel data features include transaction type labels, risk assessment scores, amount thresholds and channel score sets.

[0134] Each set of channel training data is used as the input of the enhanced channel selection model. The enhanced channel selection model takes the payment channel corresponding to each set of channel data features as output, and the actual payment channel corresponding to each set of channel data features as the prediction target; minimizing the sum of prediction errors of all payment channels is used as the training goal; the enhanced channel selection model is trained until the sum of prediction errors reaches convergence and the training is stopped; the enhanced channel selection model is a deep neural network model.

[0135] The loss function value of the enhanced channel selection model is the sum of the mean square error and the reward loss;

[0136] Prediction error formula:

[0137] ;

[0138] ;

[0139] ;

[0140] in, For reward-related loss items; is a positive integer coefficient; is the reward value; is a positive number that makes the distribution non-zero; m is the mth group of channel data features; M is the number of groups of channel data features; is the predicted payment channel corresponding to the mth group of channel data features; is the payment channel corresponding to the mth group of channel data features. When , the reward loss term decreases as the reward increases, thereby reducing the impact of prediction error loss when the reward is high; when , the reward loss term will increase as the reward decreases, aggravating the impact of the prediction error loss to punish the output of the channel selection model in this scenario.

[0141] Execute transactions based on the payment channel initially selected by the reinforcement channel selection model, and calculate the reward corresponding to the action based on the transaction results and the reward function , based on rewards Update the loss function value of the enhanced channel selection model.

[0142] According to the reward , use the reinforcement learning strategy to update the Q value in the Q value table. The update formula is as follows:

[0143] ;

[0144] in, Q value after the current state update; is the Q value before the current state is updated; is the learning rate, which is used to control the update step size and is usually between 0 and 1; For updated rewards; A discount factor used to measure the importance of future rewards, usually between 0 and 1; is the current state; The currently selected payment channel; For the next state; It is the optimal payment channel among all possible payment channels in the next state; It is the maximum Q value corresponding to all possible payment channels in the next state, that is, the maximum expected cumulative reward that can be obtained in all optional payment channels after transferring from the current state to the next state.

[0145] By using the reinforcement learning strategy to train and update the Q-value table, the Q-value in the Q-value table is continuously updated to obtain more accurate expected cumulative rewards to select payment channels, and finally optimize the strategy to maximize long-term cumulative rewards. When the Q-value update amplitude (i.e., the difference between the updated Q-value and the Q-value before the update) reaches the preset value (such as 0.01), it indicates that the Q-value table tends to be stable, and the update stops. The action corresponding to the element with the largest Q-value in the Q-value table at this time is selected, and the payment channel corresponding to the action is obtained. The selected payment channel is used as the best payment channel for the final selection.

[0146] Example 2

[0147] Reference Figure 4 This example provides another method for strengthening the channel selection model to select the output payment channel, including:

[0148] According to the payment channel output by the enhanced channel selection model, the corresponding action is selected based on the ε-greedy strategy. Here, the ε-greedy strategy can be used to randomly select a payment channel based on a certain probability, or to select the payment channel output by the enhanced channel selection model. The advantage of using the ε-greedy strategy is to balance exploration and utilization. On the one hand, new and potentially better payment channels or channel features can be discovered through random exploration; on the other hand, using the payment channel output by the enhanced channel selection model can increase the transaction success rate, reduce costs, and improve payment efficiency, while ensuring the stability and security of transactions to a certain extent.

[0149] Methods for selecting corresponding actions based on the ε-greedy strategy include:

[0150] Step 1: Set the initial exploration rate , for example, 0.8. This means that there is an 80% probability of randomly selecting a number of payment channels, and a 20% probability of selecting the payment channel output by the enhanced channel selection model.

[0151] Step 2: Get the currently defined state space.

[0152] Step 3: Generate a random number , ranging from 0 to 1.

[0153] if , then randomly select a payment channel:

[0154] Randomly select a payment channel from the set of available payment channels. For example, suppose there are payment channels A, B, C, and D to choose from, and randomly select one of them, such as payment channel A or payment channels A and B.

[0155] if , then select the payment channel output by the enhanced channel selection model:

[0156] The optimal payment channel is selected based on the Q value calculated for the current payment channel. Assume that in the current state, according to the Q value calculation, the value of payment channel C is the highest. In this case, payment channel C is selected.

[0157] Step 4: Execute the selected payment channel to complete the transaction. The results of this transaction can be recorded, which will be used for subsequent strategy updates and Q value calculation adjustments.

[0158] Step 5: As the number of transactions increases, gradually reduce the exploration rate using linear or exponential decay The purpose of reducing the exploration rate is to gradually use more payment channels that have been calculated to obtain Q values ​​after fully exploring various payment channels in the early stage, so as to improve the stability and efficiency of the strategy. However, a certain exploration probability is retained at the same time to cope with problems such as changes in payment channel performance.

[0159] The above is only a specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art can easily think of changes or substitutions within the technical scope disclosed by the present invention, which should be included in the protection scope of the present invention. Therefore, the protection scope of the present invention shall be based on the protection scope of the claims.

[0160] Finally: The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A cross-border payment channel intelligent allocation system based on transaction type, characterized by: include: The first collection module is used to collect transaction order data; the transaction order data includes contract text data and commodity text data; the contract text data includes the subject identity code and transaction amount in the contract text; the commodity text data includes the commodity name text and commodity detail description text in the commodity description text, and the commodity detail description text includes the commodity name text, commodity type text, commodity quantity text and commodity validity period text; The first analysis module is used to extract commodity text data and obtain commodity text features; Type recognition module: used to use transaction order data and product text features as inputs to the type recognition model to obtain transaction type labels; The second collection module: used to collect risk data; Risk assessment module: used to use risk data as input to the risk analysis model to obtain risk assessment scores; The third acquisition module: used to collect threshold analysis data; Threshold analysis module: used to use the threshold analysis data as the input of the threshold analysis model to obtain the amount threshold, and the amount threshold is adaptively adjusted according to the threshold analysis data collected in real time; The fourth collection module is used to collect channel analysis data of each payment channel; The channel analysis module is used to use the channel analysis data as the input of the channel analysis model to obtain the channel score of each payment channel, and to concatenate the obtained channel scores of each payment channel to obtain a channel score set; Channel selection module: used to take the transaction type labels, risk assessment scores, amount thresholds and channel score sets of different payment channels as inputs to strengthen the channel selection model to obtain the best payment channel.

2. The cross-border payment channel intelligent allocation system based on transaction type according to claim 1 is characterized in that: The data initialization of the enhanced channel selection model is as follows: Get the transaction type labels, risk assessment scores, amount thresholds, and channel score sets of different payment channels, concatenate them to obtain the state space, and use vector Indicates that the optional payment channels are aggregated into the same set to obtain the action space, and the vector Indicates that a Q value table is created for the state space and action space, and the reward function is defined as follows: ； in, is the reward value; Is the transaction success rate; For processing time; is the handling fee rate; , and They are the weights corresponding to the transaction success rate, processing time and handling fee rate respectively; Methods for creating Q-value tables corresponding to state space and action space include: The number of rows in the Q value table corresponds to the size of the state space, and the number of columns corresponds to the size of the action space. The Q value of each element in the Q value table is Initialize with a random number or 0. Indicates in status Take action The expected cumulative reward value is value.

3. The cross-border payment channel intelligent allocation system based on transaction type according to claim 2 is characterized in that: The transaction type labels, risk assessment scores, amount thresholds, and channel score sets of the different payment channels are used as inputs of the enhanced channel selection model to obtain a preliminarily selected payment channel; The training method of the enhanced channel selection model includes: Pre-collect M groups of channel training data, the channel training data including channel data features and payment channels corresponding to the channel data features; the channel data features include transaction type labels, risk assessment scores, amount thresholds and channel score sets; Each set of channel training data is used as input of the enhanced channel selection model, the enhanced channel selection model uses the payment channel corresponding to each set of channel data features as output, and uses the actual payment channel corresponding to each set of channel data features as the prediction target; minimizing the sum of prediction errors of all payment channels is used as the training target; the enhanced channel selection model is trained until the sum of prediction errors reaches convergence and the training is stopped; the enhanced channel selection model is a deep neural network model; The loss function value of the enhanced channel selection model is the sum of mean square error and reward loss; Prediction error formula: ； ； ； in, For reward-related loss items; is a positive integer coefficient; is the reward value; is a positive number that makes the distribution non-zero; m is the mth group of channel data features; M is the number of groups of channel data features; is the predicted payment channel corresponding to the mth group of channel data features; is the payment channel corresponding to the mth group of channel data features.

4. The cross-border payment channel intelligent allocation system based on transaction type according to claim 3 is characterized in that: Execute the transaction according to the payment channel initially selected by the enhanced channel selection model, and calculate the reward corresponding to the action based on the transaction result and the reward function , based on rewards Update the loss function value of the enhanced channel selection model; according to the reward , use the reinforcement learning strategy to update the Q value in the Q value table. The update formula is as follows: ； in, Q value after the current state update; is the Q value before the current state is updated; is the learning rate; For updated rewards; is the discount factor; is the current state; The currently selected payment channel; For the next state; It is the optimal payment channel among all possible payment channels in the next state; It is the maximum Q value corresponding to all possible payment channels in the next state, that is, the maximum expected cumulative reward that can be obtained in all optional payment channels after transferring from the current state to the next state; The Q-value table is trained and updated by using a reinforcement learning strategy. The Q-value in the Q-value table is continuously updated. When the update amplitude of the Q-value, that is, the difference between the updated Q-value and the Q-value before the update, reaches a preset value, the update is stopped, and the action corresponding to the element with the largest Q-value in the Q-value table at this time is selected, and the payment channel corresponding to the action is obtained. The selected payment channel is used as the best payment channel finally selected.

5. The cross-border payment channel intelligent allocation system based on transaction type according to claim 1 is characterized in that: The risk data includes user risk data, seller risk data and product risk data; the user risk data includes historical payment data, user credit rating and account security data; the historical payment data includes the number of overdue payments, the number of payment failures, the number of payments per unit time and the average payment amount per unit time; the account security data includes the number of times the user account has been stolen, the number of password changes and the number of abnormal logins; The seller risk data includes the seller's credit rating, business stability and compliance records; the business stability includes the debt-to-asset ratio and cash flow change data, and the cash flow change data includes cash flow inflow data and cash flow outflow data; the compliance record is the number of times the seller has been punished by the regulatory authorities for violations; the product risk data includes product sales volume, product feedback data and product updates within a preset time, and the product update volume is the number of new products launched; among them, the product feedback data includes product cancellation volume, positive review rate, medium review rate and negative review rate; the product cancellation volume is the number of users who have ordered the current product and have canceled this product.

6. The cross-border payment channel intelligent allocation system based on transaction type according to claim 1 is characterized in that: The method for obtaining the subject identity code includes: If the transaction party includes an enterprise, the enterprise's unified social credit code is obtained as the enterprise identity code. If the transaction party includes an individual, the user's ID number or other valid ID number is obtained as the individual identity code; the identity codes of both parties to the transaction are concatenated as the subject identity code.

7. The cross-border payment channel intelligent allocation system based on transaction type according to claim 1 is characterized in that: The method for obtaining the commodity text feature includes: Obtain the product name text and product details description text in the product description text, use the Chinese word segmentation tool to segment the product name text and product details description text respectively, remove duplicate words, obtain the first u words in the product name text and the first v words in the product details description text respectively, use the pre-trained word vector model to convert the obtained words into word vectors, use the word vectors as the input of the text extraction model, and output the product text features; The training methods for text extraction models include: Collect a group A of text training data in advance. The text training data includes word vectors and product text features corresponding to the word vectors. Each set of text training data is used as input of a text extraction model. The text extraction model uses the commodity text features corresponding to each set of word vectors as output and the actual commodity text features corresponding to each set of word vectors as prediction targets. Minimizing the sum of prediction errors of all commodity text features is used as a training target. The text extraction model is trained until the sum of prediction errors reaches convergence and training is stopped. The text extraction model is a deep neural network model. The text extraction model loss function value is the mean square error; Prediction error formula: , a is the ath group of word vectors; A is the number of groups of word vectors; is the product text feature corresponding to the a-th group of word vectors; is the actual product text feature corresponding to the a-th group of word vectors.

8. The cross-border payment channel intelligent allocation system based on transaction type according to claim 1 is characterized in that: The threshold analysis data includes user payment data, market commodity data, historical exchange rate data and channel statistical data of payment channels; user payment data is the average payment amount of x times; market commodity data is the average daily renewal amount of p competitors in the same industry in the market; historical exchange rate data is the average exchange rate change between the two parties to the transaction within a preset time; channel statistical data includes handling fee rate and payment limit; handling fee rate is the proportion of handling fees charged by different payment channels for each transaction; payment limit is the limit amount of different payment channels on single transaction and daily transaction amount.

9. The cross-border payment channel intelligent allocation system based on transaction type according to claim 8 is characterized in that: The channel analysis data includes transaction time data, historical transaction data, busyness data and channel performance data; the transaction time data includes the current user time zone and seller time zone; the historical transaction data includes the number of transactions and cumulative amounts of different payment channels within a preset time; the busyness data includes the length of the transaction processing queue in each payment channel; the channel performance data includes transaction success rate, processing time, security performance data and stability data; the transaction success rate is the ratio of the number of successful transactions to the total number of transactions for each payment channel within a preset time; the processing time is the sum of the time required for the payment channel to complete the payment from the time the user sends a payment request to the time the seller receives the payment amount; the security performance data includes the number of security occurrences and security technical data; the number of security occurrences is the number of security incidents that occur during the payment process; the security technical data is the number of types of payment methods used by the payment channel; the stability data is the number of failures and maintenance time of the payment channel within a preset time, and the maintenance time is the average duration of maintenance for each failure.

10. The cross-border payment channel intelligent allocation system based on transaction type according to claim 1, characterized in that: The training method of the type recognition model includes: Collecting group B order training data in advance, the order training data includes transaction order features and transaction type labels, wherein the transaction order features include transaction order data and commodity text features; Each set of order training data is used as input of a type recognition model. The type recognition model uses the transaction type label corresponding to each set of transaction order data and the commodity text features as output, and uses the actual transaction type label corresponding to each set of transaction order data and the commodity text features as a prediction target; minimizing the sum of prediction errors of all transaction type labels is used as a training target; the type recognition model is trained until the sum of prediction errors reaches convergence and training is stopped; the type recognition model is a deep neural network model; The type recognition model loss function value is the mean square error; Prediction error formula: , b is the bth group of transaction order features; B is the number of groups of transaction order features; is the predicted transaction type label corresponding to the transaction order features of group b; is the actual transaction type label corresponding to the transaction order features of group b.

Images