Artificial intelligence-based transaction processing method, device, and electronic equipment

Abstract

The present application provides a transaction processing method, device, electronic device and computer-readable storage medium based on artificial intelligence; it relates to the security technology in the field of artificial intelligence technology and cloud technology; the method includes: training a transaction risk control model based on transaction samples corresponding to transaction parties of labeled categories; performing category prediction processing on transaction samples corresponding to transaction parties of unlabeled categories based on the trained transaction risk control model to obtain the confidence of the predicted category; marking categories for transaction parties corresponding to confidences greater than a first confidence threshold and for transaction parties corresponding to confidences less than a second confidence threshold; and performing risk control processing on transactions involving the transaction parties to be tested based on the transaction risk control model retrained with transaction samples corresponding to transaction parties of labeled categories. Through the present application, effective protection of transactions can be achieved, and transaction processing efficiency and actual resource utilization can also be improved.

Description

Technical Field

[0001] The present application relates to artificial intelligence technology, and in particular to an artificial intelligence-based transaction processing method, device, electronic device, and computer-readable storage medium. Background Art

[0002] Artificial Intelligence (AI) is the theory, methods, techniques, and application systems that use digital computers or machines controlled by digital computers to simulate, extend, and expand human intelligence, perceive the environment, acquire knowledge, and use that knowledge to achieve optimal results. Machine Learning (ML), at the core of AI, specifically studies how computers can simulate or implement human learning behaviors to acquire new knowledge or skills and reorganize existing knowledge structures to continuously improve their performance.

[0003] Transaction processing is a key application of machine learning, widely used in scenarios such as online shopping and online transfers. In the solutions provided by related technologies, manual labeling is typically performed to label categories for certain transaction parties and train corresponding models, enabling the trained models to predict the categories of the transaction parties. However, due to the limited labeling capabilities of manual labeling (i.e., the limited number of transaction parties with tagged categories), the accuracy of the categories predicted by the trained models is low, which in turn makes it easy to incorrectly process transactions based on the wrong predicted categories, necessitating subsequent manual intervention to correct the incorrect processing.

[0004] Therefore, the transaction processing efficiency of the solutions provided by the related technologies is low and will lead to waste of resources (such as human resources and computing resources). Summary of the Invention

[0005] The embodiments of the present application provide an artificial intelligence-based transaction processing method, device, electronic device, and computer-readable storage medium, which can improve transaction processing efficiency and improve the actual utilization of resources.

[0006] The technical solution of the embodiment of the present application is implemented as follows:

[0007] The present invention provides an artificial intelligence-based transaction processing method, including:

[0008] Train the transaction risk control model based on transaction samples corresponding to the labeled categories of transaction parties;

[0009] Performing category prediction processing on transaction samples corresponding to unlabeled transaction parties according to the trained transaction risk control model to obtain confidence scores for the predicted categories of the unlabeled transaction parties;

[0010] Marking categories for transaction parties corresponding to a confidence level greater than a first confidence threshold and for transaction parties corresponding to a confidence level less than a second confidence threshold, respectively; wherein the first confidence threshold is greater than the second confidence threshold;

[0011] Retraining the transaction risk control model based on transaction samples corresponding to the labeled transaction parties;

[0012] According to the retrained transaction risk control model, category prediction processing is performed on the transaction samples to be tested corresponding to the transaction party to be tested, and risk control processing is performed on the transactions in which the transaction party to be tested participates based on the obtained prediction category.

[0013] The present invention provides an artificial intelligence-based transaction processing device, comprising:

[0014] A training module is used to train the transaction risk control model based on transaction samples corresponding to the labeled transaction parties;

[0015] A category prediction module, configured to perform category prediction processing on transaction samples corresponding to transaction parties of unlabeled categories based on the trained transaction risk control model, and obtain a confidence level of the predicted category of the transaction parties of the unlabeled categories;

[0016] a marking module, configured to mark a category for a transaction party corresponding to a confidence level greater than a first confidence threshold and a transaction party corresponding to a confidence level less than a second confidence threshold, respectively; wherein the first confidence threshold is greater than the second confidence threshold;

[0017] The training module is further configured to retrain the transaction risk control model based on transaction samples corresponding to transaction parties of the marked categories;

[0018] The risk control module is used to perform category prediction processing on the transaction samples to be tested corresponding to the transaction party to be tested based on the retrained transaction risk control model, and to perform risk control processing on the transactions in which the transaction party to be tested participates based on the obtained predicted category.

[0019] An embodiment of the present application provides an electronic device, including:

[0020] a memory for storing executable instructions;

[0021] The processor is used to implement the artificial intelligence-based transaction processing method provided in the embodiment of the present application when executing the executable instructions stored in the memory.

[0022] An embodiment of the present application provides a computer-readable storage medium storing executable instructions for causing a processor to execute instructions to implement the artificial intelligence-based transaction processing method provided in an embodiment of the present application.

[0023] The embodiments of the present application have the following beneficial effects:

[0024] Based on the transaction risk control model trained with transaction samples corresponding to labeled transaction parties, the transaction samples corresponding to unlabeled transaction parties are subjected to category prediction processing. Among the confidence levels of the predicted categories of the unlabeled transaction parties, the transaction parties corresponding to confidence levels greater than a first confidence threshold (lowest uncertainty) and those corresponding to confidence levels less than a second confidence threshold (highest uncertainty) are identified as the most valuable transaction parties and labeled accordingly. This improves the effectiveness of retraining the transaction risk control model, enabling the retrained transaction risk control model to accurately predict categories, further improving the efficiency of risk control processing for transactions, and also increasing the actual utilization of resources. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] Figure 1 This is a schematic diagram of the architecture of an artificial intelligence-based transaction processing system provided in an embodiment of the present application;

[0026] Figure 2 This is a schematic diagram of the architecture of a terminal device provided in an embodiment of the present application;

[0027] Figure 3A This is a flowchart of an artificial intelligence-based transaction processing method provided in an embodiment of the present application;

[0028] Figure 3B This is a flowchart of an artificial intelligence-based transaction processing method provided in an embodiment of the present application;

[0029] Figure 3C This is a flowchart of an artificial intelligence-based transaction processing method provided in an embodiment of the present application;

[0030] Figure 3D This is a flowchart of an artificial intelligence-based transaction processing method provided in an embodiment of the present application;

[0031] Figure 4 This is a flowchart of an artificial intelligence-based transaction processing method provided in an embodiment of the present application;

[0032] Figure 5 This is a flow chart of semi-supervised learning according to a learning method provided in an embodiment of the present application;

[0033] Figure 6 is a flowchart of semi-supervised learning according to another learning method provided in an embodiment of the present application;

[0034] Figure 7 This is a flow chart of semi-supervised learning by combining two learning methods according to an embodiment of the present application;

[0035] Figure 8 This is a flow chart of transaction processing in a payment platform provided in an embodiment of the present application. DETAILED DESCRIPTION

[0036] In order to make the purpose, technical solutions and advantages of this application clearer, the application will be further described in detail below with reference to the accompanying drawings. The described embodiments should not be regarded as limiting this application. All other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of this application.

[0037] In the following description, reference is made to “some embodiments”, which describes a subset of all possible embodiments, but it will be understood that “some embodiments” may be the same subset or different subsets of all possible embodiments and may be combined with each other without conflict.

[0038] In the following description, the terms "first\second" are only used to distinguish similar objects and do not represent a specific order for the objects. It can be understood that "first\second" can be interchanged with a specific order or sequence where permitted, so that the embodiments of the present application described herein can be implemented in an order other than that illustrated or described herein.

[0039] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by those skilled in the art to which this application pertains. The terms used herein are for the purpose of describing the embodiments of this application only and are not intended to limit this application.

[0040] Before further describing the embodiments of the present application in detail, the nouns and terms involved in the embodiments of the present application are explained. The nouns and terms involved in the embodiments of the present application are subject to the following interpretations.

[0041] 1) Transaction: refers to the exchange of value between the transacting parties using currency and / or services as the medium. For example, a transaction can refer to a transaction in an online shopping scenario, such as in a transaction, the buyer's account pays a certain amount of currency to the seller's account, and the seller's account provides specific services (such as specific goods) to the buyer's account. A transaction can also refer to a transaction in an online transfer scenario, such as in a transaction, the paying account pays a specific amount of currency to the receiving account. In the embodiments of the present application, the transaction can be an online transaction or an offline transaction. In addition, a transaction can involve multiple transacting parties, that is, there is no limit on the number of transacting parties.

[0042] 2) Transaction Party: refers to the party participating in the transaction. The embodiment of the present application does not limit the type of transaction party. For example, the transaction party can be an individual, a company or an organization.

[0043] 3) Category: refers to the category set for the transaction party. The embodiment of the present application does not limit the type of category. For example, the category can include two types: trusted category and untrusted category, and can also include three types: low-risk category, medium-risk category and high-risk category.

[0044] 4) Transaction Sample: Corresponding to the transacting party, this refers to a sample constructed from the transacting party's relevant data. Transaction samples may include at least one of the following: the transacting party's data, such as its name, and data on transactions in which the transacting party participated. Data on transactions in which the transacting party participated may include the total transaction amount, the number of transactions (transactions), and the unit price of each transaction.

[0045] 5) Transaction Risk Control Model: A classification model built based on machine learning principles, which can be a binary classification model or a multi-classification model, depending on the set categories. The present embodiment does not limit the type of transaction risk control model; for example, it can be an eXtreme Gradient Boosting (XGBoost) model, a Light Gradient Boosting Machine (LightGBM) model, or a CatBoost model.

[0046] 6) Prediction Category: This refers to the category to which the transaction party is predicted to belong. Furthermore, the confidence level of the prediction category indicates the degree of trustworthiness of the prediction category.

[0047] 7) Backpropagation: Models built on machine learning principles often involve two processes: forward propagation (also known as forward propagation) and backpropagation. For example, in a neural network model consisting of an input layer, hidden layers, and an output layer, forward propagation involves performing a series of calculations on the input data in the order input layer, hidden layer, and output layer, ultimately obtaining a predicted output result at the output layer. Backpropagation involves propagating the difference between the predicted result and the actual result to each layer in the order output layer, hidden layer, and input layer, updating the weight parameters of each layer along the gradient direction (e.g., gradient descent).

[0048] 8) Database: A collection of data stored together in a certain manner, shared by multiple users, with minimal redundancy, and independent of the application. Users can perform operations such as adding, querying, updating, and deleting data in the database. In the embodiments of this application, the database can be used to store transactions involving transacting parties or transaction samples corresponding to transacting parties.

[0049] The present invention provides an artificial intelligence-based transaction processing method, apparatus, electronic device, and computer-readable storage medium, which can improve transaction processing efficiency and resource utilization. The following describes exemplary applications of the electronic device provided in the present invention. The electronic device provided in the present invention can be implemented as various types of terminal devices or as a server.

[0050] See also Figure 1 , Figure 1 This is an architectural diagram of an artificial intelligence-based transaction processing system 100 provided in an embodiment of the present application. The terminal device 400 is connected to the server 200 through the network 300, and the server 200 is connected to the database 500. The network 300 can be a wide area network or a local area network, or a combination of the two.

[0051] In some embodiments, taking the electronic device as a terminal device as an example, the artificial intelligence-based transaction processing method provided in the embodiments of the present application can be implemented by the terminal device. For example, the terminal device 400 runs a client 410 (such as a client of a trading platform), and the client 410 obtains transaction samples corresponding to transaction parties of the marked category and transaction samples corresponding to transaction parties of the unmarked category, and trains the transaction risk control model based on the transaction samples corresponding to the transaction parties of the marked category, wherein the transaction samples can be pre-stored locally on the client 410 or obtained by the client 410 from the outside (such as a database 500). The client 410 performs category prediction processing on the transaction samples corresponding to the transaction parties of the unmarked category based on the trained transaction risk control model, obtains the confidence of the predicted category of the transaction parties of the unmarked category, and then labels the transaction parties corresponding to the confidence greater than the first confidence threshold and the transaction parties corresponding to the confidence less than the second confidence threshold as categories. After completing the category labeling, the client 410 retrains the transaction risk control model based on the transaction samples corresponding to all the marked categories of the transaction parties and stores the retrained transaction risk control model locally. In this way, when the client 410 receives a transaction request with the transaction party to be tested, it calls the retrained transaction risk control model stored locally, performs category prediction processing on the transaction samples to be tested corresponding to the transaction party to be tested, and performs risk control processing on the transactions in which the transaction party to be tested participates (i.e., transactions conducted with the transaction party to be tested) based on the obtained prediction category.

[0052] As an example, in Figure 1The figure shows a transfer interface between a transaction party 411 in a logged-in state on client 410 and a transaction party to be tested (which can be any transaction party other than transaction party 411). Upon receiving a request (i.e., a transaction request) from transaction party 411 to transfer xx yuan to the transaction party to be tested, client 410 invokes the retrained transaction risk control model to determine whether the transaction party to be tested belongs to the trusted or untrusted category. If the transaction party to be tested is determined to be untrusted, client 410 intercepts the transaction to transfer xx yuan, i.e., prohibits the execution of the transaction, and simultaneously outputs a prompt message 412 as a warning. This can prevent transaction party 411 from suffering financial losses. Of course, the risk control processing method is not limited to this.

[0053] In some embodiments, taking the electronic device as a server as an example, the artificial intelligence-based transaction processing method provided in the embodiments of the present application can also be implemented by the server. For example, the server 200 can be the backend server of the trading platform. The server 200 obtains transaction samples corresponding to the transaction parties of the marked category and the transaction samples corresponding to the transaction parties of the unmarked category from the database 500. After a series of processing, the transaction risk control model is retrained and the retrained transaction risk control model is stored locally, such as in the database 500 or the distributed file system of the server 200. Then, the server 200 can use the transaction parties of the unmarked category as the transaction parties to be tested, call the retrained transaction risk control model to perform category prediction processing on the transaction samples to be tested corresponding to the transaction parties to be tested, and perform risk control processing on the transactions in which the transaction parties to be tested participate based on the obtained predicted categories.

[0054] In some embodiments, the artificial intelligence-based transaction processing method provided in the embodiments of the present application can also be implemented collaboratively by a server and a terminal device. For example, after receiving a transaction request sent by the client 410, the server 200 determines the transaction party to be tested corresponding to the transaction request, calls the retrained transaction risk control model to perform category prediction processing on the transaction samples to be tested corresponding to the transaction party to be tested, and performs risk control processing on the transactions in which the transaction party to be tested participates based on the obtained prediction category. Among them, the transaction party to be tested corresponding to the transaction request can be the transaction party that initiates the transaction request, such as the transaction party 411 of the client 410, or it can be the transaction party corresponding to the transaction request other than the transaction party that initiates the transaction request. In addition, the server 200 can perform risk control processing on the transaction locally, or send the relevant risk control strategy to the client 410 so that the client 410 performs risk control processing locally according to the risk control strategy.

[0055] In some embodiments, the terminal device 400 or the server 200 can implement the artificial intelligence-based transaction processing method provided in the embodiments of the present application by running a computer program. For example, the computer program can be a native program or software module in the operating system; it can be a local (Native) application (APP, Application), that is, a program that needs to be installed in the operating system to run, such as an application for trading (corresponding to the client 410 above), such as an online shopping application or an online transfer application, etc.; it can also be a small program, that is, a program that can be run only after being downloaded to a browser environment; it can also be a small program that can be embedded in any APP, such as a small program component embedded in an application for trading, wherein the small program component can be controlled to run or closed by the user. In short, the above-mentioned computer program can be an application, module or plug-in in any form.

[0056] In some embodiments, the server 200 can be an independent physical server, or a server cluster or distributed system composed of multiple physical servers. It can also be a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communications, middleware services, domain name services, security services, CDN, and big data and artificial intelligence platforms. Among them, the cloud service can be a transaction processing service for the terminal device 400 to call. The terminal device 400 can be a smart phone, tablet computer, laptop computer, desktop computer, smart TV, smart watch, etc., but is not limited to this. The terminal device and the server can be directly or indirectly connected through wired or wireless communication, which is not limited in the embodiments of the present application.

[0057] It is worth noting that cloud technology refers to a hosting technology that unifies hardware, software, network, and other resources within a wide area network or local area network to enable data computing, storage, processing, and sharing. Cloud security is an important application of cloud technology, referring to the security software, hardware, users, organizations, and security cloud platforms used in cloud computing business models. In the embodiments of this application, transactions can be processed in the cloud (e.g., server 200), thereby providing cloud security services.

[0058] Taking the electronic device provided in the embodiment of the present application as an example, it can be understood that in the case where the electronic device is a server, Figure 2 Some of the structures shown in FIG (such as the user interface, presentation module, and input processing module) may be omitted. Figure 2 , Figure 2 is a schematic diagram of the structure of a terminal device 400 provided in an embodiment of the present application. Figure 2The terminal device 400 shown includes: at least one processor 410, a memory 450, at least one network interface 420, and a user interface 430. The various components in the terminal device 400 are coupled together via a bus system 440. It is understood that the bus system 440 is used to achieve connection and communication between these components. In addition to including a data bus, the bus system 440 also includes a power bus, a control bus, and a status signal bus. However, for the sake of clarity, the bus system 440 is not shown in FIG. Figure 2 Various buses are labeled as bus system 440 .

[0059] The processor 410 can be an integrated circuit chip with signal processing capabilities, such as a general-purpose processor, a digital signal processor (DSP), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc., where the general-purpose processor can be a microprocessor or any conventional processor, etc.

[0060] The user interface 430 includes one or more output devices 431 that enable presentation of media content, including one or more speakers and / or one or more visual display screens. The user interface 430 also includes one or more input devices 432, including user interface components that facilitate user input, such as a keyboard, mouse, microphone, touch screen display, camera, other input buttons and controls.

[0061] The memory 450 may be removable, non-removable, or a combination thereof. Exemplary hardware devices include solid-state memory, hard drives, optical drives, etc. The memory 450 may optionally include one or more storage devices that are physically remote from the processor 410.

[0062] The memory 450 includes volatile memory or non-volatile memory, or may include both volatile and non-volatile memory. The non-volatile memory may be a read-only memory (ROM), and the volatile memory may be a random access memory (RAM). The memory 450 described in the embodiments of the present application is intended to include any suitable type of memory.

[0063] In some embodiments, the memory 450 can store data to support various operations, examples of which include programs, modules, and data structures, or a subset or superset thereof, as exemplified below.

[0064] Operating system 451, including system programs for processing various basic system services and performing hardware-related tasks, such as the framework layer, core library layer, and driver layer, which are used to implement various basic services and process hardware-based tasks;

[0065] A network communication module 452 for reaching other computing devices via one or more (wired or wireless) network interfaces 420 , exemplary network interfaces 420 including Bluetooth, WiFi, and USB;

[0066] a presentation module 453 for enabling presentation of information via one or more output devices 431 (e.g., a display screen, a speaker, etc.) associated with the user interface 430 (e.g., a user interface for operating peripheral devices and displaying content and information);

[0067] The input processing module 454 is configured to detect one or more user inputs or interactions from one of the one or more input devices 432 and to translate the detected inputs or interactions.

[0068] In some embodiments, the apparatus provided in the embodiments of the present application may be implemented in software. Figure 2 The diagram shows an AI-based transaction processing device 455 stored in memory 450. This device can be software in the form of a program or plug-in, and includes the following software modules: a training module 4551, a category prediction module 4552, a labeling module 4553, and a risk control module 4554. These modules are logical and can be arbitrarily combined or further separated based on the functions they implement. The functions of each module are described below.

[0069] The artificial intelligence-based transaction processing method provided in the embodiments of the present application will be explained in combination with the exemplary application and implementation of the electronic device provided in the embodiments of the present application.

[0070] See also Figure 3A , Figure 3A This is a flow chart of the transaction processing method based on artificial intelligence provided by the embodiment of the present application, which will be combined with Figure 3A The steps shown are explained.

[0071] In step 101, a transaction risk control model is trained based on transaction samples corresponding to transaction parties of labeled categories.

[0072] Here, transaction parties with marked categories are identified from among multiple transaction parties. For example, the marked categories can be identified from all registered transaction parties on a transaction platform (such as a payment platform or an online shopping platform). For each transaction party with a marked category, a corresponding transaction sample is constructed based on the relevant data of the transaction party. The transaction sample may include the transaction party's own data, such as its name, registration date, and registered capital; it may also include data on transactions that the transaction party has historically participated in (i.e., completed transactions), such as the total transaction amount, number of transactions, and the unit price of each transaction. In this way, the characteristics of the transaction party can be fully reflected from two aspects.

[0073] Next, the initial transaction risk control model is trained based on transaction samples corresponding to multiple transaction parties with labeled categories. This means that the model learns how to predict categories. The category prediction task performed by the transaction risk control model can be adapted to the type of set categories. For example, if the set categories include two, the category prediction task performed by the transaction risk control model can be a binary classification task.

[0074] In some embodiments, before step 101, it also includes: performing any of the following processing for any transaction party: notifying the manual marking party according to any transaction party, and obtaining the category marked by the manual marking party for any transaction party; determining the category marked for any transaction party according to the complaint record and appeal result of any transaction party.

[0075] Here, two ways of marking categories are provided. For the sake of convenience, any transaction party is taken as an example. The first way is that the transaction party notifies the manual marking party so that the manual marking party marks the category for the transaction party, wherein the manual marking party can be an expert of the transaction platform or other relevant personnel. The embodiment of the present application does not limit the method of notification, for example, it can be notified through front-end display, text message or email, etc.; the method of manually marking the category for the transaction party by the manual marking party is also not limited. In some embodiments, the manual marking party can also mark the category for the transaction party without notification, and the electronic device can directly obtain the category marked by the manual marking party.

[0076] The second method is to determine the category of a trading party based on its complaint records and appeal results. Complaint records for a trading party can be submitted by other trading parties on the trading platform; appeal results for a trading party can be manually determined, such as by a manual tagger processing a complaint request submitted by the trading party, or by AI-based processing of the complaint request. Complaint records and appeal results for trading parties can be obtained from the trading platform's database, but these methods are not limited to these.

[0077] To illustrate the second approach, a corresponding record quantity interval can be set for each category. For example, if the categories include low-risk, medium-risk, and high-risk categories, the record quantity interval corresponding to the low-risk category can be set to [0, 20), the record quantity interval corresponding to the medium-risk category to [20, 50), and the record quantity interval corresponding to the high-risk category to [50, +∞). That is, the number of records within the record quantity interval continuously increases. When the number of records in a transaction party's complaint record successfully matches a certain record quantity interval (i.e., falls within a certain record quantity interval), and the appeal result is the first appeal result, it proves that the complaint record has a high reference value. Therefore, the category corresponding to the successfully matched record quantity interval is used as the category marked for the transaction party, where the first appeal result is empty or the appeal fails. When the transaction party's complaint record is not empty (i.e., the number of records is at least one), and the appeal result is the second appeal result, it proves that the appeal result has a high reference value. Therefore, the category marked for the transaction party is determined to be the lowest risk category, i.e., the low-risk category, where the second appeal result is successful.

[0078] To illustrate the second approach again, if the set categories include trusted and untrusted categories, the categories can be marked primarily based on the appeal result. For example, if the transaction party's complaint record is not empty and the appeal result is the first appeal result, the category marked for the transaction party is determined to be untrusted; if the complaint record is not empty and the appeal result is the second appeal result, the category marked for the transaction party is determined to be trusted.

[0079] The above two methods can improve the flexibility of marking categories. It is worth noting that the above two methods can be used either or both. For example, you can first use the second method to automatically mark categories for some transaction parties, and then use the first method to mark categories for other transaction parties, thereby saving manual marking costs.

[0080] In some embodiments, the above-mentioned training of the transaction risk control model based on transaction samples corresponding to transaction parties of labeled categories can be achieved in the following manner: category prediction processing is performed on the transaction samples corresponding to transaction parties of labeled categories according to the transaction risk control model to obtain the predicted category to be compared; based on the difference between the labeled category corresponding to the transaction sample and the predicted category to be compared, back propagation is performed in the transaction risk control model, and the weight parameters of the transaction risk control model are updated during the back propagation process.

[0081] Here, an example method for training a transaction risk control model is provided. For the sake of convenience, a transaction sample corresponding to a transaction party of any labeled category is used as an example. First, the transaction sample is subjected to probability mapping processing according to the transaction risk control model to obtain prediction probabilities corresponding to the set multiple categories (also called multiple categories). Then, a category corresponding to the predicted probability with the largest value can be used as the predicted category, or a category corresponding to the predicted probability greater than the category probability threshold can be used as the predicted category. The category probability threshold can be set according to the actual application scenario. For example, in order to ensure that there is only one predicted probability greater than the category probability threshold, the category probability threshold can be set to 0.5. In order to facilitate distinction, the predicted category obtained here is named the predicted category to be compared.

[0082] Since the transaction samples correspond to labeled categories, the difference between the labeled categories corresponding to the transaction samples and the predicted categories to be compared can be determined. This difference is the loss value, which can be calculated based on the loss function of the transaction risk control model. The embodiment of the present application does not limit the type of loss function, for example, it can be a cross-entropy loss function. Then, based on the obtained difference, backpropagation is performed in the transaction risk control model, and during the backpropagation process, the weight parameters of the transaction risk control model are updated along the gradient direction (such as the gradient descent direction) to achieve effective training of the transaction risk control model.

[0083] In step 102, category prediction processing is performed on transaction samples corresponding to transaction parties of the unlabeled category based on the trained transaction risk control model to obtain the confidence level of the predicted category of the transaction parties of the unlabeled category.

[0084] In real-world scenarios, due to limited labeling capabilities, in addition to labeled transaction parties, there are often a large number of unlabeled transaction parties. For each unlabeled transaction party, a corresponding transaction sample can also be constructed based on the relevant data of the transaction party.

[0085] After the transaction risk control model is trained based on the transaction samples corresponding to all labeled transaction parties, the transaction samples corresponding to unlabeled transaction parties can be subjected to category prediction processing based on the trained transaction risk control model to obtain the predicted category of the unlabeled transaction party and the confidence of the predicted category, where the confidence indicates the degree of credibility of the predicted category.

[0086] In some embodiments, the above-mentioned category prediction processing of transaction samples corresponding to the unlabeled category of transaction parties based on the trained transaction risk control model can be achieved in the following way: probability mapping processing is performed on the transaction samples corresponding to the unlabeled category of transaction parties based on the trained transaction risk control model to obtain prediction probabilities corresponding to multiple categories respectively; the category corresponding to the predicted probability with the largest numerical value is used as the predicted category of the unlabeled category of transaction parties, and the predicted probability with the largest numerical value is used as the confidence of the predicted category.

[0087] Here, we take the transaction sample corresponding to any unlabeled category of the transaction party as an example for explanation. First, the transaction sample is probability mapped according to the trained transaction risk control model to obtain the predicted probabilities corresponding to the set multiple categories. Then, the category corresponding to the predicted probability with the largest value can be used as the predicted category, and the predicted probability with the largest value can be used as the confidence of the predicted category; alternatively, the category corresponding to the predicted probability greater than the category probability threshold can be used as the predicted category, and the predicted probability greater than the category probability threshold can be used as the confidence of the predicted category. In the above manner, the accuracy of the determined predicted category and confidence can be improved, ensuring that the predicted category is the category predicted by the transaction risk control model and the transaction party is most likely to belong to.

[0088] In step 103, transaction parties corresponding to a confidence level greater than a first confidence threshold and transaction parties corresponding to a confidence level less than a second confidence threshold are marked with categories, respectively; wherein the first confidence threshold is greater than the second confidence threshold.

[0089] After obtaining the confidence level of the predicted category for each unlabeled trading partner, we can filter out trading partners with confidence levels greater than a first confidence threshold. For these trading partners, the trading risk control model has the lowest uncertainty about their predicted category. We can also filter out trading partners with confidence levels less than a second confidence threshold. For these trading partners, the trading risk control model has the highest uncertainty about their predicted category. The first confidence threshold is greater than the second confidence threshold and can be set based on the actual application scenario. For example, the first confidence threshold can be set to 0.99 and the second confidence threshold to 0.6.

[0090] The selected transaction parties have a higher marking value than other transaction parties. Therefore, in the embodiment of the present application, the selected transaction parties are marked as categories. The category marking method for transaction parties corresponding to a confidence level greater than a first confidence threshold and transaction parties corresponding to a confidence level less than a second confidence threshold can be different.

[0091] In some embodiments, the above-mentioned marking categories of the transaction parties corresponding to the confidence level greater than the first confidence threshold and the transaction parties corresponding to the confidence level less than the second confidence threshold can be achieved in the following manner: the predicted category of the transaction party corresponding to the confidence level greater than the first confidence threshold is used as the category marked for the transaction party corresponding to the confidence level greater than the first confidence threshold; the manual marking party is notified based on the transaction party corresponding to the confidence level less than the second confidence threshold, and the category marked by the manual marking party for the transaction party corresponding to the confidence level less than the second confidence threshold is obtained.

[0092] Here, for transaction parties with confidence levels greater than the first confidence threshold, the uncertainty of their predicted categories is the lowest, and therefore, the predicted categories can be directly used as the labeled categories. For transaction parties with confidence levels less than the second confidence threshold, the uncertainty of their predicted categories is the highest. To ensure the accuracy of the labeled categories, the manual labeler is notified based on the transaction party, and the labeled categories are obtained. By performing differentiated processing on the different types of transaction parties selected in this way, the accuracy of the labeled categories can be improved.

[0093] In step 104, the transaction risk control model is retrained based on the transaction samples corresponding to the transaction parties of the marked categories.

[0094] Step 103 yields a number of newly labeled transaction parties. Here, the transaction risk control model can be retrained based on the transaction samples corresponding to all of the labeled transaction parties. It is worth noting that retraining in this embodiment of the application refers to retraining the initial transaction risk control model. Compared to step 101, this step includes a greater number of labeled transaction parties and a higher degree of accuracy in the labeled categories, thereby improving the training effect of the transaction risk control model.

[0095] In step 105, based on the retrained transaction risk control model, category prediction processing is performed on the transaction samples to be tested corresponding to the transaction party to be tested, and risk control processing is performed on the transactions in which the transaction party to be tested participates based on the obtained predicted category.

[0096] Here, parties whose categories remain unlabeled can be considered test parties, as can parties involved in ongoing transactions (i.e., uncompleted transactions). For each identified test party, corresponding transaction samples are constructed and, for ease of distinction, are named test transaction samples. Then, using the retrained transaction risk control model, a category prediction is performed on the test transaction samples to obtain a predicted category for the test party. This predicted category can then be used to perform risk control on transactions involving the test party (e.g., ongoing transactions).

[0097] In the embodiment of the present application, a corresponding risk control method can be set for each category. For example, when the predicted category of the transaction party to be tested is a low-risk category, the risk control method is to allow the execution of the transaction in which the transaction party to be tested participates; when the predicted category of the transaction party to be tested is a medium-risk category, the risk control method is to prohibit the execution of the transaction in which the transaction party to be tested participates, and send an alarm message (prompt message) to the other transaction parties involved in the transaction; when the predicted category of the transaction party to be tested is a high-risk category, the risk control method is to prohibit the execution of the transaction in which the transaction party to be tested participates, and send an alarm message to the police and other transaction parties involved in the transaction.

[0098] In some embodiments, the above-mentioned category prediction processing of the transaction samples to be tested corresponding to the transaction party to be tested based on the retrained transaction risk control model can be achieved in the following way: when the retrained transaction risk control model meets the training stop condition, category prediction processing is performed on the transaction samples to be tested corresponding to the transaction party to be tested based on the retrained transaction risk control model.

[0099] In the embodiment of the present application, a training stop condition can be set for the transaction risk control model to determine whether the transaction risk control model meets the expected requirements. For example, the training stop condition can be a threshold of the number of training times (such as 100 times) or an indicator condition of the model indicator.

[0100] When the retrained trading risk control model meets the training termination criteria, it is proven that the retrained trading risk control model is ready for use. The retrained trading risk control model can then be used to perform category prediction on the trading samples corresponding to the tested trading parties. This approach ensures the prediction accuracy of the trading risk control model that is ultimately put into use (e.g., on an online trading platform), effectively protecting transactions.

[0101] In some embodiments, after step 104, it also includes: when the retrained transaction risk control model does not meet the training stop condition, marking the category of the transaction party corresponding to the new confidence level less than the second confidence threshold, and retraining the transaction risk control model based on the transaction samples corresponding to the transaction parties of the marked category; wherein the new confidence level is obtained by performing category prediction processing on the transaction samples corresponding to the transaction parties of the unmarked category based on the retrained transaction risk control model.

[0102] Another scenario is when the retrained trading risk control model fails to meet the training termination criteria. In this case, the retrained trading risk control model is used to perform category prediction on transaction samples corresponding to unlabeled trading parties. This results in a predicted category and a confidence level for the predicted category. For ease of distinction, this confidence level is referred to as the new confidence level.

[0103] Then, the transaction parties corresponding to the new confidence level greater than the first confidence threshold and the transaction parties corresponding to the new confidence level less than the second confidence threshold can be labeled with categories respectively, and the transaction risk control model can be retrained based on the transaction samples corresponding to all the labeled transaction parties, that is, entering the next round of iteration until the retrained transaction risk control model in a certain round of iteration meets the training stop condition.

[0104] In multiple rounds of iterations, if the transaction parties corresponding to the new confidence level greater than the first confidence threshold are always labeled as categories, the accuracy of the labeled categories may decrease, especially when the predicted category is directly used as the labeled category. Therefore, after obtaining the new confidence level, it is also possible to only label the transaction parties corresponding to the new confidence level less than the second confidence threshold (for example, relying on manual labeling to ensure the absolute accuracy of the labeled categories), and retrain the transaction risk control model based on the transaction samples corresponding to all the labeled transaction parties, that is, enter the next round of iteration until the retrained transaction risk control model in a certain round of iteration meets the training stop condition. In this way, the training effect of the transaction risk control model in multiple rounds of iteration can be further improved.

[0105] In some embodiments, before step 105, it also includes: taking the transaction parties that participated in the transactions in the previous cycle as the transaction parties to be tested; the above-mentioned risk control processing of the transactions in which the transaction parties to be tested participate according to the obtained prediction categories can be achieved in this way: when the transaction parties to be tested participate in the transactions in the latest cycle, risk control processing is performed on the transactions in which the transaction parties to be tested participate in the latest cycle.

[0106] In an embodiment of the present application, all transaction parties that participated in historical transactions (completed transactions) can be used as transaction parties to be tested, and the predicted category of each transaction party to be tested can be determined based on the retrained transaction risk control model and stored. In this way, risk control processing can be performed on transactions in which the transaction parties to be tested participate based on the stored predicted categories. This method has a large amount of calculation and is prone to waste of computing resources and storage resources. For example, some transaction parties participated in transactions a long time ago and there is no transaction demand at present. If the predicted categories of these transaction parties are determined and stored, the consumed resources cannot be effectively utilized.

[0107] Therefore, in an embodiment of the present application, risk control processing can also be performed based on a cycle, wherein the cycle can be set according to the actual application scenario, such as being set to one day. For example, all transaction parties participating in transactions in the previous cycle are treated as transaction parties to be tested, and the predicted category of each transaction party to be tested is determined based on the retrained transaction risk control model and stored. When it is detected that a transaction party to be tested has participated in transactions in the latest cycle, risk control processing is performed on the transactions in which the transaction party to be tested participated in the latest cycle based on the stored predicted category of the transaction party to be tested. In this way, the real-time performance of the risk control processing can be improved, and the consumed computing resources and storage resources can also be effectively utilized.

[0108] It's worth noting that as time passes and cycles change, stored prediction categories can be accumulated, thereby improving the comprehensiveness of risk control. For example, if the prediction category for Party A is stored on January 1st and the prediction category for Party B on January 2nd, then if a transaction involving Party A and Party B occurs on January 3rd, risk control can be applied to both transactions.

[0109] like Figure 3A As shown, the embodiments of the present application mark categories for transaction parties corresponding to a confidence level greater than a first confidence threshold and for transaction parties corresponding to a confidence level less than a second confidence threshold, respectively. This can increase the number of transaction parties in marked categories, thereby improving the training effect of the transaction risk control model, further realizing automated and intelligent risk control processing, reducing manual intervention, and also improving the actual utilization rate of consumed resources (such as manual resources and computing resources, etc.).

[0110] In some embodiments, see Figure 3B , Figure 3B This is a flow chart of a transaction processing method based on artificial intelligence provided by an embodiment of the present application. Figure 3A Step 103 shown can be implemented through steps 201 to 203 , which will be described in conjunction with each step.

[0111] In step 201 , a category is marked for a transaction party corresponding to a confidence level greater than a first confidence threshold.

[0112] In an embodiment of the present application, the process of marking categories for transaction parties corresponding to confidence levels greater than a first confidence threshold and for transaction parties corresponding to confidence levels less than a second confidence threshold can be performed simultaneously or sequentially. The latter case is explained here.

[0113] After obtaining the confidence level of the predicted category for each unlabeled transaction partner in step 102, transaction partners with confidence levels greater than a first confidence threshold may be screened out and labeled. For example, the predicted category of the selected transaction partners may be directly used as the labeled category.

[0114] In step 202, the transaction risk control model is retrained based on transaction samples corresponding to transaction parties of the labeled categories.

[0115] Through step 201, some new labeled categories of transaction parties can be obtained. Then, the transaction risk control model can be retrained based on the transaction samples corresponding to all the labeled categories of transaction parties.

[0116] In step 203, based on the retrained transaction risk control model, category prediction processing is performed on the transaction samples corresponding to the transaction parties of the unlabeled category, and the transaction parties corresponding to the obtained confidence levels less than the second confidence threshold are labeled with categories.

[0117] Here, based on the retrained transaction risk control model, transaction samples corresponding to unclassified transaction parties are subjected to category prediction to obtain predicted categories and confidence levels for the predicted categories. Then, among the confidence levels obtained in this step, transaction parties corresponding to confidence levels below a second confidence threshold are labeled with categories, for example, using manual labeling.

[0118] Since the performance of the transaction risk control model is improved after retraining in step 202, the confidence level obtained in step 203 is more accurate and can accurately determine the marking value of the transaction party in the unmarked category, thereby effectively utilizing the resources (such as manual resources) consumed in marking the category.

[0119] In some embodiments, the above-mentioned labeling categories of transaction parties corresponding to confidence levels greater than a first confidence threshold and transaction parties corresponding to confidence levels less than a second confidence threshold can also be achieved in the following ways: labeling categories of transaction parties corresponding to confidence levels less than the second confidence threshold; retraining the transaction risk control model based on transaction samples corresponding to the labeled categories of transaction parties; performing category prediction processing on transaction samples corresponding to the unlabeled categories of transaction parties based on the retrained transaction risk control model, and labeling categories of transaction parties corresponding to the obtained confidence levels greater than the first confidence threshold.

[0120] This embodiment of the present application also provides another method for sequentially labeling categories, which is suitable for scenarios where resources (such as human resources) are relatively abundant for labeling categories for transaction parties corresponding to confidence levels below a second confidence threshold. First, the transaction parties corresponding to confidence levels below the second confidence threshold are labeled, and the transaction risk control model is retrained based on transaction samples corresponding to all labeled transaction parties.

[0121] Then, based on the retrained transaction risk control model, the transaction samples corresponding to the unlabeled transaction parties are subjected to category prediction processing to obtain the predicted category and the confidence level of the predicted category. Among the obtained confidence levels, the confidence level greater than the first confidence threshold is determined, and the transaction party corresponding to the determined confidence level is labeled with a category. This approach improves the flexibility of category labeling while maintaining labeling accuracy.

[0122] like Figure 3B As shown, for different types of transaction parties, the embodiment of the present application adopts a specific order to mark the categories, so that the marking value of the transaction party can be determined more accurately, and the actual utilization rate of the resources consumed for marking the categories can also be improved.

[0123] In some embodiments, see Figure 3C , Figure 3C This is a flow chart of a transaction processing method based on artificial intelligence provided by an embodiment of the present application. Figure 3A The illustrated step 101 can be implemented through steps 301 to 302 , which will be described in conjunction with each step.

[0124] In step 301, transaction samples corresponding to transaction parties of multiple labeled categories are obtained, and the multiple transaction samples are divided into a training sample set and a verification sample set.

[0125] Here, after obtaining the transaction samples corresponding to all labeled categories of transaction parties, all transaction samples can be divided into a training sample set and a verification sample set. The ratio between the number of transaction samples included in the training sample set and the number of transaction samples included in the verification sample set can be set according to the actual application scenario, such as being set to 7:3.

[0126] For the sake of convenience, the present application provides the following examples: Figure 4 The schematic diagram shown, Figure 4 Transaction samples 1 to N are shown in the figure, where each transaction sample corresponds to a labeled category of transaction parties, and N is an integer greater than 1. When performing partitioning, transaction samples 1 to K are divided into the training sample set, and transaction samples K+1 to N are divided into the validation sample set, where K is an integer greater than 1 and less than N.

[0127] In step 302, the transaction risk control model is trained based on the training sample set.

[0128] For example, see Figure 4 , based on transaction samples 1 to K in the training sample set, the transaction risk control model is trained for the first time.

[0129] Then, based on the trained transaction risk control model, category prediction processing can be performed on the transaction samples corresponding to the unlabeled transaction parties. Figure 4 Transaction samples N+1 to N+M are shown in FIG, where each transaction sample corresponds to an unlabeled category of transaction parties, and M is an integer greater than 1. For each unlabeled category of transaction parties, a predicted category and a confidence level for the predicted category can be obtained. Then, transaction parties corresponding to confidence levels greater than a first confidence threshold and those corresponding to confidence levels less than a second confidence threshold can be labeled with categories, respectively.

[0130] exist Figure 3C middle, Figure 3A The illustrated step 104 can be implemented through steps 303 to 304 , which will be described in conjunction with each step.

[0131] In step 303, transaction samples corresponding to transaction parties of the newly labeled category are added to the training sample set.

[0132] To facilitate differentiation, transaction parties with confidence levels greater than the first confidence threshold and those with confidence levels less than the second confidence threshold are designated as newly labeled transaction parties. In this step, transaction samples corresponding to all newly labeled transaction parties are added to the training sample set to update the training sample set.

[0133] In step 304, the transaction risk control model is retrained based on the updated training sample set.

[0134] For example, see Figure 4 , based on the updated training sample set, the initial transaction risk control model is trained for the second time.

[0135] exist Figure 3C middle, Figure 3A The illustrated step 105 can be implemented through steps 305 to 306 , which will be described in conjunction with each step.

[0136] In step 305 , the model indicators of the retrained transaction risk control model are determined based on the verification sample set.

[0137] For example, see Figure 4Based on transaction samples K+1 to N in the validation sample set, the model metrics of the transaction risk control model after the second training are determined. This embodiment of the application does not limit the type of model metrics; for example, they can be precision, recall, or F1 score. Model metrics can directly reflect the performance of the transaction risk control model in performing category prediction tasks.

[0138] In step 306, when the model indicators meet the training stop conditions, the category prediction processing is performed on the transaction samples to be tested corresponding to the transaction party to be tested based on the retrained transaction risk control model, and risk control processing is performed on the transactions in which the transaction party to be tested participates based on the obtained predicted category.

[0139] Here, the training termination condition can be an indicator condition, such as an indicator threshold. For example, if the model indicator is positively correlated with the performance of the transaction risk control model, when the model indicator is greater than the indicator threshold, the transaction risk control model trained on the second round of training is used to predict the category of the transaction samples corresponding to the test party. Risk control is then applied to the transactions involving the test party based on the predicted category. When the model indicator is less than or equal to the indicator threshold, the next iteration is initiated until the model indicator exceeds the indicator threshold.

[0140] like Figure 3C As shown, the embodiment of the present application trains the transaction risk control model through a training sample set, and verifies the model indicators of the trained transaction risk control model through a verification sample set, which can accurately determine whether the trained transaction risk control model meets the expected requirements, thereby improving the accuracy of risk control processing.

[0141] In some embodiments, see Figure 3D , Figure 3D This is a flow chart of a transaction processing method based on artificial intelligence provided by an embodiment of the present application. Figure 3A Step 105 shown can be implemented through steps 401 to 403 , which will be described in conjunction with each step.

[0142] In step 401, based on the retrained transaction risk control model, category prediction processing is performed on the transaction samples to be tested corresponding to the transaction parties to be tested to obtain a predicted category.

[0143] Here, we take the example of setting categories including trusted categories and untrusted categories to illustrate an example of risk control processing. First, based on the retrained transaction risk control model, the category of the transaction sample corresponding to the tested transaction party is predicted to obtain the predicted category.

[0144] In step 402 , when the predicted category of the transaction party to be tested is a credible category, the transaction in which the transaction party to be tested participates is allowed to be executed.

[0145] Here, when the predicted category of the transaction party to be tested is a credible category, the transaction in which the transaction party to be tested participates is released, that is, the transaction is allowed to be executed.

[0146] It is worth noting that since a transaction involves multiple parties, the transaction is only allowed to be executed when the prediction categories of all parties are credible.

[0147] In step 403, when the predicted category of the transaction party to be tested is the untrustworthy category, at least one of intercepting the transaction in which the transaction party to be tested participates and closing the transaction participation authority of the transaction party to be tested is performed.

[0148] Here, when the predicted category of the transaction party to be tested is a non-trusted category, at least one of the following processes can be executed: 1) interception processing of transactions in which the transaction party to be tested participates; 2) closing processing of the transaction participation rights of the transaction party to be tested, such as closing the payment collection rights of the transaction party to be tested; 3) alarm processing of transactions in which the transaction party to be tested participates, that is, outputting alarm information to prompt relevant personnel to manually process the transaction (such as manual withdrawal, manual interception, etc.), where the relevant personnel are such as the back-end personnel of the transaction platform, and such as the transaction parties who participate in the transaction and belong to the trusted category.

[0149] Of course, the above risk control processing methods are only examples and can be specifically set according to actual application scenarios.

[0150] like Figure 3D As shown, the embodiment of the present application performs different processing according to different prediction categories, which can effectively protect the transaction security of the transaction parties and improve security.

[0151] Below, an exemplary application of an embodiment of the present application in an actual application scenario will be described. For ease of understanding, an example will be given of a situation where the transaction party is a merchant in a payment platform. Among a large number of merchants on the payment platform, some credible merchants (corresponding to the transaction parties belonging to the credible category above) and some malicious merchants (corresponding to the transaction parties belonging to the non-credible category above) can be marked, for example, by an expert (corresponding to the manual marking party above), or by combining the merchant's complaint record and appeal results for marking. When predicting whether a merchant is credible, the prediction is made for all merchants in the payment platform. Therefore, when using labeled samples (i.e., transaction samples corresponding to merchants of the labeled category) for model training, the overall situation of the merchant cannot be accurately reflected.

[0152] The situation where there are a large number of unlabeled samples (i.e., transaction samples corresponding to merchants of unlabeled categories) belongs to the semi-supervised learning problem in machine learning. In the embodiment of the present application, learning method 1) is used to perform semi-supervised learning. As an example, the following is provided: Figure 5The flow chart of the learning method 1) shown will be explained in conjunction with each step.

[0153] Step 1) Train an initial machine learning model (corresponding to the transaction risk control model described above) based on all labeled samples. The labeled samples here include transaction samples corresponding to labeled trustworthy merchants and transaction samples corresponding to labeled malicious merchants. It is worth noting that the embodiment of the present application does not limit the type of machine learning model. For example, it can be an XGBoost model, a LightGBM model, or a CatBoost model. The machine learning model in the embodiment of the present application can apply a binary classification mode.

[0154] Step 2) For each unlabeled sample, the unlabeled sample is subjected to category prediction processing according to the trained machine learning model to obtain the predicted category to which the merchant corresponding to the unlabeled sample is most likely to belong, as well as the confidence of the predicted category. When the confidence corresponding to an unlabeled sample is greater than the first confidence threshold, the predicted category is used as the category marked for the unlabeled sample, wherein the first confidence threshold can be set according to the actual application scenario, such as being set to 0.99. It is worth noting that marking the unlabeled sample is equivalent to marking the merchant corresponding to the unlabeled sample. In order to facilitate distinction, the samples (unlabeled samples) that are category-marked in this step are named new labeled samples.

[0155] Step 3) Retrain the initial machine learning model based on all the newly labeled samples and all the labeled samples in step 1).

[0156] In learning method 1), the number of unlabeled samples corresponding to the confidence level greater than the first confidence threshold may only account for 20% to 30% of all unlabeled samples, which cannot cover most of the unlabeled samples. Therefore, in the embodiment of the present application, learning method 2) is also used for semi-supervised learning. As an example, the following is provided: Figure 6 The flow chart of the learning method 2) shown will be explained in the form of steps.

[0157] Step 1) All labeled samples are divided into a training sample set and a validation sample set. The ratio between the number of labeled samples included in the training sample set and the number of labeled samples included in the validation sample set can be set according to the actual application scenario, such as 7:3.

[0158] Step 2) Train the initial machine learning model based on the training sample set. For easy distinction, the trained machine learning model obtained here is named the initial version model.

[0159] Step 3) Use the initial version model to predict the category of each unlabeled sample in the unlabeled sample set, and obtain the predicted category and confidence level for each unlabeled sample. Here, the value of labeling unlabeled samples can be measured using an uncertainty strategy. For example, the closer the confidence level is to 0.5, the higher the uncertainty of the initial version model for the unlabeled sample corresponding to that confidence level. In other words, the greater the information content of the unlabeled sample, the higher the value of labeling the unlabeled sample.

[0160] Step 4) Experts label the batch of unlabeled samples with the greatest uncertainty (i.e., mark the category). Here, the batch of unlabeled samples corresponding to a confidence level less than a second confidence threshold can be considered the batch of unlabeled samples with the greatest uncertainty. The second confidence threshold can be set based on the actual application scenario, such as 0.6.

[0161] Step 5) The samples marked by the experts are used as new marked samples, the new marked samples are added to the training sample set, and the initial machine learning model is retrained based on the updated training sample set.

[0162] Step 6) Based on the retrained machine learning model, perform category prediction and verification on the validation sample set to obtain model metrics for the retrained machine learning model. If the model metrics meet the criteria, the iteration ends. If the model metrics do not meet the criteria, the iteration continues, using the retrained machine learning model as the initial version and executing steps 3) through 6).

[0163] It is worth noting that other training stop conditions can also be set for the machine learning model. For example, when an instruction to stop labeling samples is received, the iteration ends; when an instruction to stop labeling samples is not received, the iteration continues.

[0164] In learning method 2), expert experience can be used to label a batch of unlabeled samples with the greatest uncertainty and the most difficult to predict in the unlabeled sample set, which can improve the training effect of the machine learning model and achieve comprehensive training.

[0165] In the embodiment of the present application, the above-mentioned learning method 1) and learning method 2) can be applied in combination. As an example, the following is provided: Figure 7The training diagram of the machine learning model shown. For example, labeled samples can be constructed based on the relevant data of trusted merchants to serve as positive samples for model training, and the label value of the positive sample (the label value is used to represent the category corresponding to the sample) is marked as 1, wherein trusted merchants can include merchants that have been judged to be trusted through manual investigation in history (i.e., merchants marked as trusted by experts) and merchants who have successfully submitted appeals after being attacked (i.e., merchants whose complaint records are non-empty and whose appeal results are successful appeals); labeled samples can be constructed based on the relevant data of malicious merchants to serve as negative samples for model training, and the label value of the negative sample is marked as 0, wherein malicious merchants can include merchants that have been judged to be malicious through manual investigation in history and merchants complained by users as malicious (i.e., merchants whose complaint records are non-empty and whose appeal results are empty or whose appeals fail). In all labeled samples, the ratio of the number of positive samples to negative samples can be 1:2, and of course the ratio can be adjusted according to the actual application scenario. The machine learning model to be trained can be an XGBoost model that applies a binary classification mode, but this does not constitute a limitation on the embodiments of the present application. Will combine Figure 7 , explaining the training process in steps.

[0166] Step 1) Divide all labeled samples into a training sample set and a validation sample set.

[0167] Step 2) Train the initial machine learning model based on the training sample set to obtain a preliminary version of the model.

[0168] Step 3) For each unlabeled sample in the unlabeled sample set, perform category prediction processing on the unlabeled sample according to the initial version model to obtain the predicted probability that the label value of the unlabeled sample is 0 (that is, the unlabeled sample belongs to a negative sample, and the corresponding merchant belongs to a malicious merchant), and the predicted probability that the label value is 1 (that is, the unlabeled sample belongs to a positive sample, and the corresponding merchant belongs to a trusted merchant). The label value corresponding to the predicted probability with the largest value is used as the predicted label value, that is, the category corresponding to the predicted probability with the largest value is used as the predicted category, and the predicted probability with the largest value is used as the confidence of the predicted category.

[0169] Step 4) When the confidence level of the predicted category for the unlabeled sample is greater than a first confidence threshold (e.g., 0.99), the predicted label value of the unlabeled sample is used as the labeled label value. This means the predicted category is used as the labeled category, resulting in a newly labeled sample. The newly labeled sample is then added to the training sample set.

[0170] Step 5) Retrain the initial machine learning model based on the updated training sample set to obtain a new version of the model.

[0171] Step 6) Perform category prediction processing on each unlabeled sample in the unlabeled sample set according to the new version of the model to obtain the predicted category corresponding to each unlabeled sample and the confidence level corresponding to the predicted category.

[0172] Step 7) The batch of unlabeled samples corresponding to the confidence level less than the second confidence threshold is regarded as the batch of unlabeled samples with the largest uncertainty (i.e., the batch of unlabeled samples with the largest amount of information), and the experts are notified to label this batch of unlabeled samples, i.e., manually mark the label values.

[0173] Step 8) The samples marked by the experts are used as new marked samples, the new marked samples are added to the training sample set, and the initial machine learning model is retrained based on the updated training sample set.

[0174] Step 9) Based on the retrained machine learning model obtained in step 8), perform category prediction and verification processing on the validation sample set to obtain model indicators for the retrained machine learning model. When the model indicators meet the indicator conditions, the iteration ends, i.e., the training of the machine learning model is completed; when the model indicators do not meet the indicator conditions, the iteration continues, i.e., the retrained machine learning model obtained in step 8) is used as the new version of the model, and steps 6) to 9) are executed.

[0175] The combined use of learning methods 1) and 2) not only yields a large number of valid new labeled samples, thereby improving the training effect of the machine learning model, but also saves the cost of expert labeling and improves the training efficiency of the machine learning model.

[0176] In the embodiment of the present application, it is possible to periodically predict whether merchants in the payment platform are trustworthy. For example, the trained machine learning model can be deployed to a server that has Python and its data science dependent components installed, and a field named "merchant trust probability" can be created in the key-value database in advance. After the deployment of the machine learning model and the creation of the field are completed, Figure 8 As shown, the prediction process is as follows.

[0177] Step 1) At 1:00 AM each day, the payment platform's commercial payment log data for the previous day (i.e., the previous cycle) is retrieved to identify merchants that participated in transactions the previous day. Merchant feature data for these merchants is then processed to obtain transaction samples corresponding to the merchants. These transaction samples include, but are not limited to, the merchant's transaction amount, number of users, number of transactions, unit price per transaction, and merchant name.

[0178] Step 2) Call the trained machine learning model to perform category prediction on the transaction samples to obtain the prediction probability corresponding to the credible category.

[0179] Step 3) The predicted probability corresponding to the trustworthy category is sent to an online Key-Value database through a data interface for storage, for example, in the form of predicted probability corresponding to merchant-trustworthy category.

[0180] Step 4) Set judgment rules in the real-time transaction decision engine. For example, when a new transaction is detected, the "Merchant Trust Probability" field in the Key-Value Database is called to determine whether the merchant participating in the new transaction is trustworthy. For example, a category probability threshold (such as 0.5) can be set. When the predicted probability corresponding to the merchant's trustworthy category is greater than the category probability threshold, the merchant is determined to be trustworthy; otherwise, the merchant is determined to be malicious.

[0181] Step 5) When the merchant participating in the new transaction is a trusted merchant, the new transaction is allowed to be executed.

[0182] Step 6) When a merchant participating in a new transaction is a malicious merchant, the new transaction or the malicious merchant can be cracked down on, such as intercepting the new transaction or disabling the malicious merchant's transaction participation rights (such as payment collection rights). Other risk control strategies can also be adopted to achieve effective risk control.

[0183] Of course, the above online processing flow is only an example and does not constitute a limitation on the embodiments of the present application. For example, in step 2), the predicted category can be directly determined, and in step 3), the predicted category can be sent to the Key-Value database for storage, for example, in the form of merchant-prediction category.

[0184] Experimental verification has shown that, while the number of malicious merchants on the payment platform has not increased, applying the solution of the embodiments of this application can reduce the number of merchant attacks and transaction attacks by approximately 20%, while ensuring a normal merchant collection experience and user payment experience. It also significantly reduces the number of merchant complaints, thereby saving the manpower cost required to handle merchant complaints.

[0185] The following continues to describe the exemplary structure of the artificial intelligence-based transaction processing device 455 provided in the embodiment of the present application implemented as a software module. In some embodiments, such as Figure 2As shown, the software modules stored in the artificial intelligence-based transaction processing device 455 of the memory 450 may include: a training module 4551, which is used to train the transaction risk control model based on the transaction samples corresponding to the labeled transaction parties; a category prediction module 4552, which is used to perform category prediction processing on the transaction samples corresponding to the unlabeled transaction parties according to the trained transaction risk control model, and obtain the confidence of the predicted category of the unlabeled transaction parties; a marking module 4553, which is used to mark categories for the transaction parties corresponding to the confidence greater than the first confidence threshold and the transaction parties corresponding to the confidence less than the second confidence threshold, respectively; wherein the first confidence threshold is greater than the second confidence threshold; the training module 4551 is also used to retrain the transaction risk control model based on the transaction samples corresponding to the labeled transaction parties; a risk control module 4554, which is used to perform category prediction processing on the transaction samples to be tested corresponding to the transaction parties to be tested according to the retrained transaction risk control model, and perform risk control processing on the transactions in which the transaction parties to be tested participate according to the obtained predicted categories.

[0186] In some embodiments, the marking module 4553 is further used to: mark categories for transaction parties corresponding to confidence levels greater than a first confidence threshold; retrain the transaction risk control model based on transaction samples corresponding to transaction parties of the marked categories; perform category prediction processing on transaction samples corresponding to transaction parties of the unmarked categories based on the retrained transaction risk control model, and mark categories for transaction parties corresponding to confidence levels less than a second confidence threshold.

[0187] In some embodiments, the marking module 4553 is further used to: use the predicted category of the transaction party corresponding to the confidence level greater than the first confidence threshold as the category marked for the transaction party corresponding to the confidence level greater than the first confidence threshold; notify the manual marking party based on the transaction party corresponding to the confidence level less than the second confidence threshold, and obtain the category marked by the manual marking party for the transaction party corresponding to the confidence level less than the second confidence threshold.

[0188] In some embodiments, the risk control module 4554 is further used to: when the retrained transaction risk control model meets the training stop condition, perform category prediction processing on the transaction samples to be tested corresponding to the transaction party to be tested based on the retrained transaction risk control model.

[0189] In some embodiments, the training module 4551 is further used to: when the retrained transaction risk control model does not meet the training stop condition, label the transaction party corresponding to the new confidence level that is less than the second confidence threshold, and retrain the transaction risk control model based on the transaction samples corresponding to the transaction parties of the labeled category; wherein the new confidence level is obtained by performing category prediction processing on the transaction samples corresponding to the transaction parties of the unlabeled category based on the retrained transaction risk control model.

[0190] In some embodiments, the training module 4551 is further used to: obtain transaction samples corresponding to multiple labeled categories of transaction parties, and divide the multiple transaction samples into a training sample set and a verification sample set; train the transaction risk control model according to the training sample set; the risk control module 4554 is further used to: determine the model indicators of the retrained transaction risk control model according to the verification sample set; when the model indicators meet the training stop conditions, perform category prediction processing on the transaction samples to be tested corresponding to the transaction parties to be tested according to the retrained transaction risk control model.

[0191] In some embodiments, the training module 4551 is further used to: add transaction samples corresponding to transaction parties of the newly marked category to the training sample set; and retrain the transaction risk control model based on the updated training sample set.

[0192] In some embodiments, the marking module 4553 is also used to perform any of the following processing on any transaction party: according to any transaction party notifying the manual marking party, and obtaining the category marked by the manual marking party for any transaction party; according to the complaint record and appeal result of any transaction party, determining the category marked for any transaction party.

[0193] In some embodiments, the categories include trusted categories and untrusted categories; the marking module 4553 is also used to: when the complaint record is not empty and the appeal result is the first appeal result, determine that the category marked for any transaction party is an untrusted category; when the complaint record is not empty and the appeal result is the second appeal result, determine that the category marked for any transaction party is a trusted category; wherein the first appeal result is empty or the appeal fails; the second appeal result is a successful appeal.

[0194] In some embodiments, the category prediction module 4552 is further used to: perform probability mapping processing on transaction samples corresponding to the unlabeled category of transaction parties according to the trained transaction risk control model to obtain prediction probabilities corresponding to multiple categories; use the category corresponding to the predicted probability with the largest numerical value as the predicted category of the unlabeled category of transaction parties, and use the predicted probability with the largest numerical value as the confidence of the predicted category.

[0195] In some embodiments, the training module 4551 is further used to: perform category prediction processing on transaction samples corresponding to the labeled categories of the transaction parties according to the transaction risk control model to obtain the predicted categories to be compared; perform backpropagation in the transaction risk control model according to the difference between the labeled categories corresponding to the transaction samples and the predicted categories to be compared, and update the weight parameters of the transaction risk control model during the backpropagation process.

[0196] In some embodiments, the risk control module 4554 is also used to: when the predicted category of the transaction party to be tested is a trusted category, allow the execution of transactions in which the transaction party to be tested participates; when the predicted category of the transaction party to be tested is an untrusted category, perform at least one of interception processing of transactions in which the transaction party to be tested participates, and closing processing of the transaction participation rights of the transaction party to be tested.

[0197] In some embodiments, the risk control module 4554 is further used to: use the transaction parties that participated in the transactions in the previous cycle as the transaction parties to be tested; when the transaction parties to be tested participate in the transactions in the latest cycle, perform risk control processing on the transactions in which the transaction parties to be tested participate in the latest cycle.

[0198] The present invention provides a computer program product or computer program, which includes computer instructions stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium and executes the computer instructions, causing the computer device to perform the artificial intelligence-based transaction processing method described above in the present invention.

[0199] The embodiment of the present application provides a computer-readable storage medium storing executable instructions, wherein the executable instructions are stored. When the executable instructions are executed by a processor, the processor will execute the method provided by the embodiment of the present application, for example, Figure 3A 、 Figure 3B 、 Figure 3C and Figure 3D The transaction processing method based on artificial intelligence is shown.

[0200] In some embodiments, the computer-readable storage medium may be a memory such as FRAM, ROM, PROM, EPROM, EEPROM, flash memory, magnetic surface storage, optical disk, or CD-ROM; or various devices including one or any combination of the above memories.

[0201] In some embodiments, executable instructions may be in the form of a program, software, software module, script, or code, written in any form of programming language (including compiled or interpreted languages, or declarative or procedural languages), and may be deployed in any form, including as a stand-alone program or as a module, component, subroutine, or other unit suitable for use in a computing environment.

[0202] As an example, executable instructions may, but need not, correspond to a file in a file system, may be stored as part of a file that stores other programs or data, such as in one or more scripts in a HyperText Markup Language (HTML) document, in a single file dedicated to the program in question, or in multiple coordinating files (e.g., files storing one or more modules, subroutines, or code portions).

[0203] By way of example, executable instructions may be deployed to be executed on one computing device, or on multiple computing devices at one site, or on multiple computing devices distributed across multiple sites and interconnected by a communication network.

[0204] The above are merely examples of the present application and are not intended to limit the scope of protection of the present application. Any modifications, equivalent replacements, and improvements made within the spirit and scope of the present application are included in the scope of protection of the present application.

Claims

1. A transaction processing method based on artificial intelligence, characterized in that: The method comprises: Train the transaction risk control model based on transaction samples corresponding to the labeled categories of transaction parties; Performing category prediction processing on transaction samples corresponding to unlabeled transaction parties according to the trained transaction risk control model to obtain confidence scores for the predicted categories of the unlabeled transaction parties; Labeling categories for transaction parties corresponding to confidence levels greater than a first confidence threshold; retraining the transaction risk control model based on transaction samples corresponding to transaction parties of the labeled categories; performing category prediction processing on transaction samples corresponding to transaction parties of unlabeled categories based on the retrained transaction risk control model, and labeling categories for transaction parties corresponding to confidence levels obtained that are less than a second confidence threshold, wherein the first confidence threshold is greater than the second confidence threshold; Retraining the transaction risk control model based on transaction samples corresponding to the labeled transaction parties; When the retrained transaction risk control model meets the training stop condition, category prediction processing is performed on the transaction samples to be tested corresponding to the transaction party to be tested based on the retrained transaction risk control model, wherein the training stop condition includes a training number threshold or an indicator condition of a model indicator, and the categories include a trusted category and an untrusted category; When the predicted category of the transaction party to be tested is the credible category, allowing execution of the transaction in which the transaction party to be tested participates; When the predicted category of the transaction party to be tested is the untrusted category, at least one of intercepting the transaction in which the transaction party to be tested participates and closing the transaction participation authority of the transaction party to be tested is performed.

2. The method according to claim 1, characterized in that The transaction party marking category corresponding to the confidence level greater than the first confidence threshold includes: using the predicted category of the transaction party corresponding to the confidence level greater than the first confidence threshold as the category marked for the transaction party corresponding to the confidence level greater than the first confidence threshold; The transaction party marking category corresponding to the obtained confidence level less than the second confidence threshold includes: The manual marking party is notified according to the transaction party corresponding to the confidence level less than the second confidence threshold, and the category marked by the manual marking party for the transaction party corresponding to the confidence level less than the second confidence threshold is obtained.

3. The method according to claim 1, characterized in that After retraining the transaction risk control model based on transaction samples corresponding to transaction parties of labeled categories, the method further includes: When the retrained transaction risk control model does not meet the training stop condition, labeling the transaction parties corresponding to the new confidence level less than the second confidence threshold with a category, and retraining the transaction risk control model based on transaction samples corresponding to the transaction parties of the labeled category; The new confidence level is obtained by performing category prediction processing on transaction samples corresponding to transaction parties of unlabeled categories based on the retrained transaction risk control model.

4. The method according to claim 1, wherein The training of the transaction risk control model based on transaction samples corresponding to the transaction parties of the labeled categories includes: Acquire transaction samples corresponding to a plurality of labeled categories of transaction parties, and divide the plurality of transaction samples into a training sample set and a verification sample set; Training the transaction risk control model according to the training sample set; The step of performing category prediction processing on the transaction samples to be tested corresponding to the transaction parties to be tested based on the retrained transaction risk control model includes: Determining model indicators of the retrained transaction risk control model based on the verification sample set; When the model indicators meet the training stop conditions, category prediction processing is performed on the transaction samples to be tested corresponding to the transaction parties to be tested based on the retrained transaction risk control model.

5. The method according to claim 4, characterized in that The retraining of the transaction risk control model based on transaction samples corresponding to the transaction parties of the marked categories includes: Adding transaction samples corresponding to transaction parties of the newly marked category to the training sample set; Retrain the transaction risk control model based on the updated training sample set.

6. The method according to any one of claims 1 to 5, characterized in that Before training the transaction risk control model based on transaction samples corresponding to transaction parties of labeled categories, the method further includes: For any transaction party, perform any of the following processing: Notifying the manual marking party according to any one of the transaction parties and obtaining the category marked by the manual marking party for the any one of the transaction parties; Determine a category marked for any transaction party based on the complaint record and complaint result of any transaction party.

7. The method according to claim 6, characterized in that The categories include trusted categories and untrusted categories; the categories marked for any transaction party based on the complaint records and appeal results of any transaction party include: When the complained record is not empty and the appeal result is the first appeal result, determining that the category marked for any one of the transaction parties is the untrustworthy category; When the complained record is not empty and the appeal result is the second appeal result, determining that the category marked for any one of the transaction parties is the trusted category; Among them, the first appeal result is empty or the appeal fails; the second appeal result is successful.

8. The method according to any one of claims 1 to 5, characterized in that The processing of performing category prediction on transaction samples corresponding to unlabeled transaction parties according to the trained transaction risk control model to obtain the confidence level of the predicted category of the unlabeled transaction parties includes: Performing probability mapping processing on transaction samples corresponding to the unlabeled categories of transaction parties according to the trained transaction risk control model to obtain predicted probabilities corresponding to multiple categories; The category corresponding to the predicted probability with the largest value is used as the predicted category of the transaction party of the unlabeled category, and the predicted probability with the largest value is used as the confidence of the predicted category.

9. The method according to any one of claims 1 to 5, characterized in that The training of the transaction risk control model based on transaction samples corresponding to the transaction parties of the labeled categories includes: Performing category prediction processing on transaction samples corresponding to the transaction parties of the marked categories according to the transaction risk control model to obtain predicted categories to be compared; Based on the difference between the marked category corresponding to the transaction sample and the predicted category to be compared, back propagation is performed in the transaction risk control model, and the weight parameters of the transaction risk control model are updated during the back propagation process.

10. The method according to any one of claims 1 to 5, characterized in that The method further comprises: The trading parties that participated in the transactions in the previous cycle are used as the trading parties to be tested; The performing risk control processing on the transaction in which the tested transaction party participates according to the obtained prediction category includes: When the transaction party to be tested participates in transactions in the latest cycle, risk control processing is performed on the transactions participated in by the transaction party to be tested in the latest cycle.

11. A transaction processing device based on artificial intelligence, characterized in that: The device comprises: A training module is used to train the transaction risk control model based on transaction samples corresponding to the labeled transaction parties; A category prediction module, configured to perform category prediction processing on transaction samples corresponding to transaction parties of unlabeled categories based on the trained transaction risk control model, and obtain a confidence level of the predicted category of the transaction parties of the unlabeled categories; a labeling module configured to label categories for transaction parties corresponding to confidence levels greater than a first confidence threshold; retrain the transaction risk control model based on transaction samples corresponding to transaction parties of the labeled categories; perform category prediction processing on transaction samples corresponding to transaction parties of unlabeled categories based on the retrained transaction risk control model, and label categories for transaction parties corresponding to confidence levels obtained that are less than a second confidence threshold, wherein the first confidence threshold is greater than the second confidence threshold; The training module is further configured to retrain the transaction risk control model based on transaction samples corresponding to transaction parties of the marked categories; A risk control module is used to perform category prediction processing on the transaction samples to be tested corresponding to the transaction party to be tested according to the retrained transaction risk control model when the retrained transaction risk control model meets the training stop condition, wherein the training stop condition includes a threshold of the number of times trained or an indicator condition of a model indicator, and the categories include trusted categories and untrusted categories; when the predicted category of the transaction party to be tested is the trusted category, allowing the execution of transactions in which the transaction party to be tested participates; when the predicted category of the transaction party to be tested is the untrusted category, performing at least one of intercepting the transactions in which the transaction party to be tested participates and closing the transaction participation rights of the transaction party to be tested.

12. The device according to claim 11, characterized in that The device further comprises: The marking module is further used to use the predicted category of the transaction party corresponding to the confidence level greater than the first confidence threshold as the category marked for the transaction party corresponding to the confidence level greater than the first confidence threshold; notify the manual marking party based on the transaction party corresponding to the confidence level less than the second confidence threshold, and obtain the category marked by the manual marking party for the transaction party corresponding to the confidence level less than the second confidence threshold.

13. The device according to claim 11, characterized in that The device further comprises: The risk control module is also used to perform category prediction processing on the transaction samples to be tested corresponding to the transaction party to be tested according to the retrained transaction risk control model when the retrained transaction risk control model meets the training stop condition.

14. An electronic device, characterized in that: include: a memory for storing executable instructions; A processor, configured to implement the artificial intelligence-based transaction processing method according to any one of claims 1 to 10 when executing the executable instructions stored in the memory.

15. A computer-readable storage medium, characterized in that Executable instructions are stored for implementing the artificial intelligence-based transaction processing method according to any one of claims 1 to 10 when executed by a processor.

16. A computer program product comprising executable instructions, characterized in that When the executable instructions are executed by the processor, the artificial intelligence-based transaction processing method according to any one of claims 1 to 10 is implemented.

Images