Financial data processing method and system

Abstract

The invention discloses a financial data processing method and system. The method comprises the steps that a center node sends a training instruction to a first edge node and a second edge node; the first and second edge nodes respectively perform local training on the model to obtain first and second model parameters; the first edge node and the second edge node respectively send the first model parameter and the second model parameter to the center node through instant messaging connection, so that the center node performs model parameter aggregation to obtain an aggregated model parameter; the first edge node receives aggregation model parameters; and the first edge node updates the model according to the aggregation model parameters to obtain an aggregation model so as to execute the calculation task. Thus, in the model training process, only the process that the model parameters are transmitted between the center node and the edge nodes exists, the financial training data, used for model training, of the edge nodes deployed in the branch mechanisms are not transmitted to other nodes, and the problem that transmission of the financial training data of the nodes is limited is effectively solved.

Description

Technical Field

[0001] The present application relates to the field of computers, and in particular to a financial data processing method and system. Background Art

[0002] At present, with the continuous development of computer and artificial intelligence (AI) technology, more and more industries or fields have begun to use AI models to analyze, model or apply relevant data to improve the efficiency and quality of the data processing process. For example, in the financial industry, financial institutions can use AI models to analyze, model or apply the large amount of business data generated in different fields such as credit, anti-money laundering, internal control and compliance, thereby effectively improving the effect of business data processing.

[0003] Generally speaking, the quality of AI models will have an impact on the effectiveness of data processing. Usually, the larger the data scale of AI model training and the more comprehensive the data coverage, the better the quality of the AI ​​model, and the more versatile and robust the data processing process, the more accurate the processing results. For example, with the development of economic globalization, financial institutions can integrate the business data generated by domestic and overseas branches, and use the business data with a larger scale of integrated data to train AI models, thereby effectively improving the versatility, robustness and accuracy of the data analysis process, and avoiding the limitations of using part of the domestic data or part of the overseas data to independently train AI models.

[0004] However, when financial institutions integrate business data generated by domestic and overseas branches for AI model training, some countries or regions where the branches are located may restrict the transmission of business data due to factors such as data sovereignty, privacy protection or regulatory constraints. For example, the GDPR (General Data Protection Regulation) passed in 2016 strictly regulates the behavior of all organizations involved in processing personal data of EU citizens, thereby greatly restricting the transmission process of business data.

[0005] In actual application, financial institutions can homomorphically encrypt and transmit the business data of branches to cope with restrictions on data transmission in the country or region to which the branch belongs. Financial institutions can directly perform specific computing operations without decrypting the homomorphically encrypted business data, and the results of the computing operations are consistent with the results of the original business data after the same computing operations are performed and then encrypted. However, when financial institutions use homomorphically encrypted business data to train AI models, the consumption of computing resources will increase significantly, the computing efficiency will also be significantly reduced, and considering that the computing operations that can be performed on homomorphically encrypted business data are limited, the complexity of AI model training is also relatively high. Not only that, the data occupancy of homomorphically encrypted business data is larger, and the communication overhead of the business data transmission process will also increase accordingly. Summary of the invention

[0006] The embodiment of the present application provides a financial data processing method to avoid the transmission restriction problem of the financial training data of the branch during the model training process. In addition, the embodiment of the present application also provides a corresponding financial data processing system.

[0007] In a first aspect, an embodiment of the present application provides a financial data processing method, which is applied to a financial data processing system, wherein the financial data processing system includes: a central node deployed in a financial institution, a first edge node deployed in a first branch, and a second edge node deployed in a second branch, wherein the first branch and the second branch are overseas branches of the financial institution, an instant communication connection is established between the central node and the first edge node and the second edge node, and the financial training data of the first branch and the financial training data of the second branch are restricted from being transmitted to the financial institution or other branches of the financial institution; the method includes: the central node sends a training instruction to the first edge node and the second edge node, wherein the training instruction is used to instruct the first edge node and the second edge node to perform local training on the model; the first edge The edge node locally trains the model according to the financial training data of the first branch to obtain first model parameters, and the second edge node locally trains the model according to the financial training data of the second branch to obtain second model parameters; the first edge node sends the first model parameters to the central node through the instant communication connection, and the second edge node sends the second model parameters to the central node through the instant communication connection; the central node aggregates model parameters according to the first model parameters and the second model parameters to obtain aggregated model parameters; the first edge node receives the aggregated model parameters sent by the central node; the first edge node updates the model according to the aggregated model parameters to obtain an aggregated model; the first edge node executes a calculation task for the financial task data of the first branch according to the aggregated model.

[0008] In a possible implementation, the central node performs model parameter aggregation according to the first model parameter and the second model parameter to obtain aggregated model parameters, including: the central node determines a weight value of the first model parameter and a weight value of the second model parameter, the weight value of the first model parameter is positively correlated with the data volume of the financial training data of the first branch, and the weight value of the second model parameter is positively correlated with the data volume of the financial training data of the second branch; the central node performs model parameter aggregation according to the first model parameter, the second model parameter, the weight value of the first model parameter, and the weight value of the second model parameter to obtain the aggregated model parameters.

[0009] In a possible implementation, the first edge node performs a computing task for the financial task data of the first branch according to the aggregation model, including: the first edge node locally trains the aggregation model according to the financial training data of the first branch to obtain a third model parameter; the first edge node updates the aggregation model according to the third model parameter to obtain a personalized model; the first edge node performs a computing task for the financial task data of the first branch according to the personalized model.

[0010] In a possible implementation, the first edge node performs a computing task for the financial task data of the first branch according to the aggregation model, including: the first edge node determines a convergence state of the aggregation model; when the convergence state meets a convergence condition, the first edge node performs a computing task for the financial task data of the first branch according to the aggregation model.

[0011] In a possible implementation, before the central node aggregates model parameters according to the first model parameters and the second model parameters, the method further includes: the central node locally trains the model according to the financial training data of the financial institution to obtain fourth model parameters, and the financial training data of the financial institution is restricted from being transmitted to other branches of the financial institution; the central node aggregates model parameters according to the first model parameters and the second model parameters to obtain aggregated model parameters, including: the central node aggregates model parameters according to the first model parameters, the second model parameters and the fourth model parameters to obtain the aggregated model parameters.

[0012] In a possible implementation, the central node sends a training instruction to the first edge node and the second edge node, including: the central node determines whether the first edge node and the second edge node meet training conditions, the training conditions including that the data volume of the financial training data of the branch to which the edge node belongs meets the data volume range; when the first edge node and the second edge node meet the training conditions, the central node sends the training instruction to the first edge node and the second edge node.

[0013] In a possible implementation, the method further includes: the central node updating the model according to the aggregation model parameters to obtain the aggregation model; and the central node executing a calculation task for the financial task data of the financial institution according to the aggregation model.

[0014] In a possible implementation, the financial task data of the first branch is the internal control and compliance data of the first branch, and the first edge node performs a computing task on the financial task data of the first branch according to the aggregation model, including: the first edge node obtains the internal control and compliance data of the first branch according to a database deployed in the first branch; the first edge node uses the aggregation model to parse the internal control and compliance data of the first branch to obtain multiple internal control and compliance labels of the first branch.

[0015] In a possible implementation, the method also includes: the first edge node sends the multiple internal control compliance labels to the central node; based on the multiple internal control compliance labels, the central node generates a hot word cloud map, and the hot word cloud map is used to indicate the number of different types of internal control compliance labels in the multiple internal control compliance labels.

[0016] In a second aspect, an embodiment of the present application provides a financial data processing system, including: a central node deployed in a financial institution, a first edge node deployed in a first branch, and a second edge node deployed in a second branch, the first branch and the second branch are overseas branches of the financial institution, an instant communication connection is established between the central node and the first edge node and the second edge node, and the financial training data of the first branch and the financial training data of the second branch are restricted from being transmitted to the financial institution or other branches of the financial institution; the central node is used to send a training instruction to the first edge node and the second edge node, the training instruction is used to instruct the first edge node and the second edge node to perform local training on the model; the central node is also used to perform modeling according to the first model parameters and the second model parameters. The first edge node is used to locally train the model according to the financial training data of the first branch to obtain the first model parameters; the first edge node is also used to send the first model parameters to the central node through the instant communication connection; the first edge node is also used to receive the aggregated model parameters sent by the central node; the first edge node is also used to update the model according to the aggregated model parameters to obtain the aggregated model; the first edge node is also used to perform a calculation task for the financial task data of the first branch according to the aggregated model; the second edge node is used to locally train the model according to the financial training data of the second branch to obtain the second model parameters; the second edge node is also used to send the second model parameters to the central node through the instant communication connection.

[0017] In a possible implementation, the central node is specifically used to: determine a weight value of the first model parameter and a weight value of the second model parameter, the weight value of the first model parameter is positively correlated with the data volume of the financial training data of the first branch, and the weight value of the second model parameter is positively correlated with the data volume of the financial training data of the second branch; perform model parameter aggregation according to the first model parameter, the second model parameter, the weight value of the first model parameter and the weight value of the second model parameter to obtain the aggregated model parameter.

[0018] In a possible implementation, the first edge node is specifically used to: locally train the aggregation model according to the financial training data of the first branch to obtain third model parameters; update the aggregation model according to the third model parameters to obtain a personalized model; and execute a calculation task for the financial task data of the first branch according to the personalized model.

[0019] In a possible implementation, the first edge node is specifically used to: determine a convergence state of the aggregation model; and when the convergence state meets a convergence condition, execute a calculation task for the financial task data of the first branch according to the aggregation model.

[0020] In a possible implementation, the central node is also used to: locally train the model according to the financial training data of the financial institution to obtain a fourth model parameter, and the financial training data of the financial institution is restricted from being transmitted to other branches of the financial institution; the central node is specifically used to: aggregate model parameters according to the first model parameters, the second model parameters and the fourth model parameters to obtain the aggregated model parameters.

[0021] In one possible implementation, the central node is specifically used to: determine whether the first edge node and the second edge node meet training conditions, the training conditions including that the amount of financial training data of the branch to which the edge node belongs meets a data amount range; when the first edge node and the second edge node meet the training conditions, send the training instruction to the first edge node and the second edge node.

[0022] In a possible implementation, the central node is further used to: update the model according to the aggregation model parameters to obtain the aggregation model; and execute a calculation task for the financial task data of the financial institution according to the aggregation model.

[0023] In a possible implementation, the financial task data of the first branch is the internal control and compliance data of the first branch, and the first edge node is specifically used to: obtain the internal control and compliance data of the first branch according to a database deployed in the first branch; and parse the internal control and compliance data of the first branch using the aggregation model to obtain multiple internal control and compliance labels of the first branch.

[0024] In a possible implementation, the first edge node is also used for: the first edge node sends the multiple internal control compliance labels to the central node; the central node is also used for generating a hot word cloud map based on the multiple internal control compliance labels, and the hot word cloud map is used to indicate the number of different types of internal control compliance labels in the multiple internal control compliance labels.

[0025] In the above implementation method of the embodiment of the present application, the central node sends a training instruction to the first edge node and the second edge node, and the training instruction is used to instruct the first edge node and the second edge node to perform local training on the model; the first edge node locally trains the model according to the financial training data of the first branch to obtain first model parameters, and the second edge node locally trains the model according to the financial training data of the second branch to obtain second model parameters; the first edge node sends the first model parameters to the central node through an instant communication connection, and the second edge node sends the second model parameters to the central node through an instant communication connection; the central node aggregates model parameters according to the first model parameters and the second model parameters to obtain aggregated model parameters; the first edge node receives the aggregated model parameters sent by the central node; the first edge node updates the model according to the aggregated model parameters to obtain an aggregated model; the first edge node executes the calculation task for the financial task data of the first branch according to the aggregated model.

[0026] In this way, during the model training process, there is only the process of transmitting model parameters between the central node and the edge node. The financial training data used for model training at the edge node deployed in the branch will not be transmitted to other nodes, effectively avoiding the transmission restriction problem of the financial training data of the node. At the same time, the data size of the financial training data is usually smaller than the size of the model parameters, which reduces the communication overhead of data transmission during the model training process. In addition, the financial training data used for training has not undergone complicated processing such as encryption. During the model training process, the complexity of model training is reduced, the overall efficiency of model training and computing task execution is improved, and the adaptability to different models and different computing tasks is improved. Moreover, the central node aggregates the model parameters obtained by local training of the models by different edge nodes. Since the data scale involved in the model training process is larger and the data coverage is more comprehensive, the quality of the obtained aggregated model parameters is often better. Then, the edge nodes deployed in the branches update the models according to the aggregated model parameters, and use the updated aggregated models to perform calculation tasks for the financial task data of the branches. As a result, the versatility of the calculation task execution is often stronger, the robustness is often higher, and the execution results are often more accurate, which effectively avoids the limitations of the calculation task execution results when the edge nodes independently train the models. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in the present application. For ordinary technicians in this field, other drawings can also be obtained based on these drawings.

[0028] Figure 1 A schematic diagram of a financial data processing scenario in an embodiment of the present application;

[0029] Figure 2 A flowchart of a financial data processing method in an embodiment of the present application;

[0030] Figure 3 It is a structural diagram of a financial data processing system in an embodiment of the present application. DETAILED DESCRIPTION

[0031] See also Figure 1 , is a schematic diagram of a financial data processing scenario provided by this application. Figure 1 As shown, the financial data processing scenario 10 includes at least three computing nodes and at least three memories. Figure 1The example includes computing node 101, computing node 102, computing node 103, memory 201, memory 202 and memory 203. In actual application scenarios, the financial data processing scenario 10 may include a larger number of computing units and memories.

[0032] The computing node 101 and the memory 201, the computing node 102 and the memory 202, and the computing node 103 and the memory 203 may be connected via a bus, which may be, for example, a peripheral component interconnect express (PCIE) bus, etc. The computing nodes 101 and 102, and the computing nodes 101 and 103 may also be connected via a network, which may be, for example, a local area network (LAN), a wide area network (WAN), etc. in terms of coverage, and may be, for example, a wired network or a wireless network, etc. in terms of connection mode.

[0033] Among them, computing node 101 (computing node 102, computing node 103 are similar) refers to a device with data processing and control capabilities. For example, computing node 101 can be a graphics processing unit (GPU), a central processing unit (CPU) or a combination of multiple processors, etc., and there is no limitation on this.

[0034] The memory 201 (similar to the memory 202 and the memory 203) refers to a device with file storage capability, such as a semiconductor memory, a magnetic memory, an optical memory, etc., which is not limited. In addition, the memory 201 may store a database for data management and data query, which is not limited.

[0035] In the embodiment of the present application, computing node 101 is a central node deployed in a financial institution, computing node 102 and computing node 103 are edge nodes deployed in a first branch and a second branch respectively, for example, computing node 102 may be a first edge node, and computing node 103 may be a second edge node. The first branch and the second branch are overseas branches of the financial institution. In actual application, the financial institution may have more branches, and computing nodes may also be deployed in other branches.

[0036] In actual application scenarios, different fields of the financial industry (such as credit, anti-money laundering, internal control and compliance, or other financial fields) may generate large amounts of business data. Financial institutions can use AI models to analyze, model or apply business data, thereby effectively improving the effectiveness of business data processing.

[0037] Generally speaking, the quality of AI models is closely related to the effect of data processing. Usually, as the data scale increases and the data coverage increases, the performance of AI models tends to be better, which makes the data processing process more versatile and robust, and the processing results more accurate.

[0038] To adapt to the development of economic globalization, financial institutions often set up branches in overseas countries or regions. Financial institutions can integrate the business data of overseas branches and train AI models by building larger-scale and more comprehensive data sets, thereby overcoming the limitations of data training in a single overseas branch and improving the breadth, stability and accuracy of data analysis.

[0039] However, due to strict supervision by regulatory authorities in foreign countries or regions, the overseas data generated by branches of financial institutions are often restricted due to factors such as data sovereignty, privacy protection or regulatory constraints, making it difficult to centralize the financial data used for training AI models. For example, the GDPR regulation adopted by the European Union in 2016 aims to restrict the behavior of organizations involved in processing personal data of EU citizens, including the behavior of overseas branches of financial institutions transmitting financial data to financial institutions or other branches.

[0040] In the embodiment of the present application, the financial training data of the first branch and the financial training data of the second branch are both restricted from being transmitted to the financial institution or other branches of the financial institution.

[0041] In actual application, financial institutions can use homomorphic encryption technology to solve the problem of limited cross-border data transmission. Through homomorphic encryption, financial institutions can directly perform calculations without decrypting financial data to achieve the purpose of model training, and the calculation results are consistent with the results of the original data after encryption.

[0042] However, this approach creates other problems.

[0043] First, this implementation method will greatly increase computing resource consumption and reduce computing efficiency. For example, homomorphic encryption technology often involves complex mathematical operations, which means that some computing resources have to be used in the data encryption stage, taking away computing resources that can be used in the model training and model reasoning stages to truly solve actual business problems. In addition, the computational complexity of this method is usually higher than the complexity of computing on the original data, resulting in a decline in computing efficiency.

[0044] Secondly, this implementation method has poor adaptability and flexibility. The introduction of homomorphic encryption technology often requires targeted modifications to model-related algorithms, such as adding encoding conversion processing operations, which increases the complexity of model implementation and reduces feasibility.

[0045] Thirdly, this implementation method will significantly increase the communication overhead of data transmission, because the encrypted data often occupies more space than the original data before encryption.

[0046] Based on this, the central node sends a training instruction to the first edge node and the second edge node, and the training instruction is used to instruct the first edge node and the second edge node to perform local training on the model; the first edge node locally trains the model according to the financial training data of the first branch to obtain first model parameters, and the second edge node locally trains the model according to the financial training data of the second branch to obtain second model parameters; the first edge node sends the first model parameters to the central node through an instant communication connection, and the second edge node sends the second model parameters to the central node through an instant communication connection; the central node aggregates model parameters according to the first model parameters and the second model parameters to obtain aggregated model parameters; the first edge node receives the aggregated model parameters sent by the central node; the first edge node updates the model according to the aggregated model parameters to obtain an aggregated model; the first edge node executes the calculation task for the financial task data of the first branch according to the aggregated model.

[0047] In this way, during the model training process, there is only the process of transmitting model parameters between the central node and the edge node. The financial training data used for model training at the edge node deployed in the branch will not be transmitted to other nodes, effectively avoiding the transmission restriction problem of the financial training data of the node. At the same time, the data size of the financial training data is usually smaller than the size of the model parameters, which reduces the communication overhead of data transmission during the model training process. In addition, the financial training data used for training has not undergone complicated processing such as encryption. During the model training process, the complexity of model training is reduced, the overall efficiency of model training and computing task execution is improved, and the adaptability to different models and different computing tasks is improved. Moreover, the central node aggregates the model parameters obtained by local training of the models by different edge nodes. Since the data scale involved in the model training process is larger and the data coverage is more comprehensive, the quality of the obtained aggregated model parameters is often better. Then, the edge nodes deployed in the branches update the models according to the aggregated model parameters, and use the updated aggregated models to perform calculation tasks for the financial task data of the branches. As a result, the versatility of the calculation task execution is often stronger, the robustness is often higher, and the execution results are often more accurate, which effectively avoids the limitations of the calculation task execution results when the edge nodes independently train the models.

[0048] In order to make the above-mentioned purposes, features and advantages of the present application more obvious and easy to understand, various non-limiting implementation methods in the embodiments of the present application are exemplarily described below in conjunction with the accompanying drawings. Obviously, the described embodiments are part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in the field without making creative work are within the scope of protection of this application.

[0049] See also Figure 2 , Figure 2 A schematic diagram of a financial data processing method in an embodiment of the present application is shown. The method can be applied to Figure 1 The financial data processing scenario 10 shown in FIG. 10 may be applied to other applicable financial data processing scenarios. Figure 1 The financial data processing scenario 10 shown in the figure is used as an example to illustrate the financial data processing. Figure 2 As shown, the method may specifically include the following steps.

[0050] S201: The central node sends a training instruction to the first edge node and the second edge node, where the training instruction is used to instruct the first edge node and the second edge node to perform local training on the model.

[0051] The central node deployed in the financial institution can send a training instruction to the edge node deployed in the branch, so that each edge node starts to train the model locally. The training instruction sent by the central node to each edge node can be a control message of the instant messaging connection established between the central node and the edge node.

[0052] In actual application, optionally, before the central node sends a training instruction to the edge node, the central node can determine a list of edge nodes that can be used for training, so that the central node can select at least two edge nodes that meet the training conditions from the list for training based on the requirements of model training. Exemplarily, after the edge node deployed in the branch is awakened, it can send a wake-up confirmation message to the central node, and then, after the central node receives the wake-up confirmation message, the central node can send an instruction message to the edge node to instruct the edge node to initialize. Subsequently, each edge node performs initialization before model training based on the received instruction message. For example, each edge node can load the model and the financial training data deployed in the memory of the branch, and after the initialization is completed, the initialized edge node can send a training standby message to the central node. Finally, the central node adds the edge node that is already in the training standby state to the list of edge nodes that can be used for training based on the training standby message received from the edge node. For example, the central node can add the ID of the edge node to the list.

[0053] Furthermore, to meet the computational requirements of model training, the number of edge nodes used for training needs to reach the number of computational requirements required by the model. When the number of edge nodes in the list meets the computational requirements of model training, the central node can select all edge nodes in the list to train the model.

[0054] In addition, since the amount of financial training data of the branch to which the edge node belongs is relatively low, the model parameters obtained when the edge node trains the model have little impact on the global situation. Based on this, the central node can select edge nodes in the list that meet the training conditions for training, where the training conditions include that the amount of financial training data of the branch to which the edge node belongs meets the data amount range.

[0055] In the embodiment of the present application, the training of the model is described by taking the first edge node deployed in the first branch and the second edge node deployed in the second branch as an example. In actual application, the central node can also determine more edge nodes for training the model. When the first edge node and the second edge node meet the above training conditions, the central node sends a training instruction to the first edge node and the second edge node.

[0056] It should be noted that the training instructions sent by the central node to the first edge node and the second edge node may be implicit instructions. For example, the central node may send training parameters to the edge node to instruct the edge node to start executing model training.

[0057] S202: The first edge node locally trains the model according to the financial training data of the first branch to obtain first model parameters, and the second edge node locally trains the model according to the financial training data of the second branch to obtain second model parameters.

[0058] After receiving the training instruction sent by the central node, the first edge node may perform local training on the model according to the financial training data of the first branch, wherein the financial training data of the first branch may be stored in the memory 202. Similarly, the second edge node may also perform local training on the model according to the financial training data of the second branch.

[0059] In actual application, the first edge node (same for the second edge node) can perform multiple iterations in the process of model training, wherein the number of iterations can be controlled by the central node by sending control information to control the first edge node to iterate a preset number of iterations in the process of model training. Alternatively, after each iteration, the first edge node can determine whether the model after iteration converges based on the financial training data of the first branch, and when the convergence state of the model after iteration meets the convergence condition, the first edge node stops the model training process accordingly.

[0060] After the training is completed, the first edge node can obtain the first model parameters, and the second edge node can obtain the second model parameters.

[0061] In this way, the financial training data used for model training at the edge nodes deployed in the branch offices is not transmitted to other nodes, effectively avoiding the transmission restriction problem of the financial training data of the nodes. In addition, the financial training data used for training is not encrypted or processed in a complicated way. In the process of model training, the complexity of model training is reduced, the overall efficiency of model training and computing task execution is improved, and the adaptability to different models and different computing tasks is improved.

[0062] S203: The first edge node sends the first model parameter to the central node through the instant communication connection, and the second edge node sends the second model parameter to the central node through the instant communication connection.

[0063] After the first edge node and the second edge node have finished training the model, the first model parameters and the second model parameters obtained from the training can be sent to the central node, so that the central node can perform the aggregation process of the model parameters. In actual application, the first edge node (the second edge node in the same way) can also send parameter transmission instruction information to the central node along with the first model parameters, so that the central node can accept the first model parameters when receiving the parameter transmission instruction information. In addition, the first edge node (the second edge node in the same way) and the central node transmit data through an instant messaging connection, which improves the efficiency of the entire process of model training and aggregation.

[0064] In this way, since the data size of financial training data is usually smaller than the size of model parameters, the communication overhead of data transmission during model training is reduced.

[0065] S204: The central node aggregates model parameters according to the first model parameters and the second model parameters to obtain aggregated model parameters.

[0066] After receiving the first model parameter and the second model parameter, the central node may aggregate the model parameters to obtain an aggregated model parameter.

[0067] In this embodiment, the following examples of implementing model parameter aggregation by the central node are provided.

[0068] In the first implementation example, the central node may average the first model parameter and the second model parameter to obtain the aggregate model parameter. For example, the first model parameter and the second model parameter are respectively a multidimensional vector, and the central node may average the one-to-one corresponding elements in the two multidimensional vectors to obtain a new multidimensional vector, that is, the aggregate model parameter.

[0069] In the second implementation example, the central node may first determine the weight value of the first model parameter and the weight value of the second model parameter, wherein the weight value of the first model parameter is positively correlated with the data volume of the financial training data of the first branch, and the weight value of the second model parameter is positively correlated with the data volume of the financial training data of the second branch, and then, the central node may calculate the product between the first model parameter and the weight value of the first model parameter, and the product between the second model parameter and the weight value of the second model parameter, and determine the average value between the two products, that is, perform weighted average aggregation on the model parameters. In this way, when there are differences in the data volume of the financial training data of each branch, the dominant influence of the model parameters corresponding to the financial training data with a smaller data volume on the aggregated model parameters and the mismatch of the data volume is avoided. In addition, the weight value of the model parameter can also be dynamically adjusted according to the location of the branch or other factors, which is not limited.

[0070] In actual application, the central node can also aggregate model parameters based on other implementation examples, which will not be elaborated here.

[0071] Furthermore, before the central node aggregates the model parameters, the central node may also perform local training on the model according to the financial training data of the financial institution to obtain the fourth model parameters. Similarly, the financial training data of the financial institution is also restricted from being transmitted to other branches of the financial institution. Furthermore, the central node may aggregate the model parameters according to the first model parameters, the second model parameters and the fourth model parameters to obtain the aggregated model parameters.

[0072] Alternatively, the central node may aggregate the model parameters according to the first model parameters and the second model parameters to obtain the first aggregate model parameters. Further, the central node updates the model based on the first aggregate model parameters, and performs local training on the updated model according to the financial training data of the financial institution to obtain the second aggregate model parameters.

[0073] Alternatively, the central node may also introduce the influence of financial training data of financial institutions on the aggregation model parameters based on other means.

[0074] S205: The first edge node receives the aggregation model parameters sent by the central node.

[0075] After the central node aggregates the model parameters, the obtained aggregated model parameters may be sent to the first edge node, so that the first edge node updates the model and performs the computing task. Alternatively, the central node may also send the obtained aggregated model parameters to the second edge node, so that the second edge node updates the model and performs the computing task. Alternatively, the central node may also send the obtained aggregated model parameters to the third edge node, so that the third edge node updates the model and performs the computing task, wherein the third edge node is an edge node deployed in the third branch.

[0076] S206: The first edge node updates the model according to the aggregation model parameters to obtain an aggregation model.

[0077] S207: The first edge node executes a computing task for the financial task data of the first branch according to the aggregation model.

[0078] The first edge node can obtain an aggregation model by updating the model according to the aggregation model parameters. Then, the first edge node can perform a calculation task for the financial task data of the first branch according to the aggregation model.

[0079] In this way, the central node aggregates the model parameters obtained by local training of the model by different edge nodes. Since the data scale involved in the model training process is larger and the data coverage is more comprehensive, the quality of the obtained aggregated model parameters is often better. Then, the edge nodes deployed in the branch offices update the model according to the aggregated model parameters, and use the updated aggregated model to perform calculation tasks for the financial task data of the branch offices. As a result, the versatility of the calculation task execution is often stronger, the robustness is often higher, and the execution results are often more accurate, which effectively avoids the limitations of the calculation task execution results when the edge nodes independently train the model.

[0080] In addition, the central node, the second edge node, and the third edge node may also update the model according to the aggregation model parameters to obtain the aggregation model, and execute the calculation task for the financial task data of the affiliated branch according to the aggregation model.

[0081] Alternatively, the financial data processing method provided in this embodiment can be combined with the application system of a financial institution, that is, the application system can periodically send model training tasks to the central node, and after the central node completes the aggregation of the model parameters, the obtained aggregated model parameters can be sent to the application system.

[0082] Optionally, the central node may first update the model according to the aggregation model parameters to obtain the aggregation model, and determine whether the aggregation model after aggregation converges. The first edge node may accept the convergence state of the aggregation model sent by the central node, and when the convergence state meets the convergence condition, the first edge node will perform the calculation task for the financial task data of the first branch according to the aggregation model. When the convergence state does not meet the convergence condition, the first edge node may refer to steps S202 to S207 to repeat the local training and the central node may aggregate the model parameters.

[0083] Furthermore, the first edge node can locally train the aggregation model according to the financial training data of the first branch to obtain the third model parameter. Subsequently, the first edge node updates the aggregation model according to the third model parameter to obtain the personalized model. Furthermore, the first edge node performs the calculation task for the financial task data of the first branch according to the personalized model.

[0084] In this way, the first edge node retrains the aggregation model, and then introduces the impact of the financial data of the first branch that is different from other branches on the model during the repeated training process to meet the personalized requirements of the branch's edge node for the model. For example, the country or region where the first branch is located may produce more personalized data that fits itself due to geographical issues, avoiding the problem of reduced applicability of the aggregation model with global commonality.

[0085] It should be noted that, in the process of re-training the aggregation model by the first edge node, the financial training data used for training can be adjusted to training data with personalized characteristics of the first branch, so that the final personalized model has both global commonalities and personalized characteristics.

[0086] Exemplarily, the financial task data of the first branch is the internal control compliance data of the first branch, and the first edge node performs a computing task on the financial task data of the first branch. In actual application, the financial institution can be a head office of a multinational bank, and the branch is a branch of the multinational bank located overseas.

[0087] The internal control and compliance system of the head office of a multinational bank can generate a hot word cloud map of internal control and compliance based on the internal control and compliance tags generated by domestic and overseas branches for viewing by relevant personnel in internal control and compliance work, and to assist in the analysis and decision-making of matters related to internal control and compliance.

[0088] Normally, each branch of a multinational bank located overseas can complete data analysis, statistics and result aggregation based on local internal control and compliance data, and manually declare or transmit the aggregated internal control and compliance labels to the head office of the multinational bank via the Internet, thereby generating a hot word cloud map of internal control and compliance. However, since the quality of the internal control and compliance labels analyzed by each branch varies, the efficiency is low, and the labor cost is high, based on this, an aggregation model can be obtained by referring to the method provided in this embodiment, so that each branch can process the internal control and compliance data based on the aggregation model to obtain internal control and compliance labels with higher quality and unified results.

[0089] Take the first edge node processing the internal control compliance data as an example. The first edge node can obtain the internal control compliance data of the first branch from the memory of the first branch. Among them, the memory of the first branch may include the database of the first branch, and accordingly, the first edge node can obtain the internal control compliance data of the first branch according to the database. Furthermore, the first edge node uses the aggregation model to parse the internal control compliance data of the first branch to obtain multiple internal control compliance labels of the first branch. In actual application, the aggregation model can be a FastText model, and the first edge node can use the aggregation model to parse the internal control compliance data of the first branch at a certain period based on a timer.

[0090] On this basis, the first edge node can send multiple internal control compliance labels obtained by parsing to the central node deployed in the financial institution. Then, based on the multiple internal control compliance labels received, the central node generates a corresponding hot word cloud map, wherein the hot word cloud map is used to indicate the number of different types of internal control compliance labels among the multiple internal control compliance labels, that is, the more internal control compliance labels of the same type, the higher the popularity of the label, and the more obvious the label is presented to relevant personnel of internal control compliance work in the hot word cloud map.

[0091] In addition, the present application also provides a financial data processing system. Figure 3 , Figure 3 A schematic diagram of the structure of a financial data processing system in an embodiment of the present application is shown. Figure 3 The illustrated financial data processing system 300 includes:

[0092] A central node 301 deployed at a financial institution, a first edge node 302 deployed at a first branch, and a second edge node 303 deployed at a second branch, wherein the first branch and the second branch are overseas branches of the financial institution, an instant communication connection is established between the central node 301 and the first edge node 302 and the second edge node 303, and the financial training data of the first branch and the financial training data of the second branch are restricted from being transmitted to the financial institution or other branches of the financial institution.

[0093] The central node 301 is used to send training instructions to the first edge node 302 and the second edge node 303, and the training instructions are used to instruct the first edge node 302 and the second edge node 303 to perform local training on the model; the central node 301 is also used to aggregate model parameters according to the first model parameters and the second model parameters to obtain aggregated model parameters.

[0094] The first edge node 302 is used to locally train the model according to the financial training data of the first branch to obtain first model parameters; the first edge node 302 is also used to send the first model parameters to the central node 301 through the instant communication connection; the first edge node 302 is also used to receive the aggregate model parameters sent by the central node 301; the first edge node 302 is also used to update the model according to the aggregate model parameters to obtain an aggregate model; the first edge node 302 is also used to perform calculation tasks for the financial task data of the first branch according to the aggregate model.

[0095] The second edge node 303 is used to locally train the model according to the financial training data of the second branch to obtain second model parameters; the second edge node 303 is also used to send the second model parameters to the central node 301 through the instant messaging connection.

[0096] In a possible implementation, the central node 301 is specifically used to: determine the weight value of the first model parameter and the weight value of the second model parameter, the weight value of the first model parameter is positively correlated with the data volume of the financial training data of the first branch, and the weight value of the second model parameter is positively correlated with the data volume of the financial training data of the second branch; perform model parameter aggregation according to the first model parameter, the second model parameter, the weight value of the first model parameter and the weight value of the second model parameter to obtain the aggregated model parameter.

[0097] In a possible implementation, the first edge node 302 is specifically used to: locally train the aggregation model according to the financial training data of the first branch to obtain third model parameters; update the aggregation model according to the third model parameters to obtain a personalized model; and execute a calculation task for the financial task data of the first branch according to the personalized model.

[0098] In a possible implementation manner, the first edge node 302 is specifically configured to: determine a convergence state of the aggregation model; and when the convergence state satisfies a convergence condition, execute a calculation task for the financial task data of the first branch according to the aggregation model.

[0099] In a possible implementation, the central node 301 is also used to: locally train the model according to the financial training data of the financial institution to obtain a fourth model parameter, and the financial training data of the financial institution is restricted from being transmitted to other branches of the financial institution; the central node 301 is specifically used to: aggregate model parameters according to the first model parameters, the second model parameters and the fourth model parameters to obtain the aggregated model parameters.

[0100] In a possible implementation, the central node 301 is specifically used to: determine whether the first edge node 302 and the second edge node 303 meet training conditions, the training conditions including that the data volume of the financial training data of the branch to which the edge node belongs meets the data volume range; when the first edge node 302 and the second edge node 303 meet the training conditions, send the training indication to the first edge node 302 and the second edge node 303.

[0101] In a possible implementation, the central node 301 is further used to: update the model according to the aggregation model parameters to obtain the aggregation model; and execute a calculation task for the financial task data of the financial institution according to the aggregation model.

[0102] In a possible implementation, the financial task data of the first branch is the internal control and compliance data of the first branch, and the first edge node 302 is specifically used to: obtain the internal control and compliance data of the first branch according to the database deployed in the first branch; and parse the internal control and compliance data of the first branch using the aggregation model to obtain multiple internal control and compliance labels of the first branch.

[0103] In a possible implementation, the first edge node 302 is further used for: the first edge node 302 sends the multiple internal control compliance labels to the central node 301; the central node 301 is further used for generating a hot word cloud map based on the multiple internal control compliance labels, and the hot word cloud map is used to indicate the number of different types of internal control compliance labels in the multiple internal control compliance labels.

[0104] It should be noted that the information interaction, execution process and other contents between the nodes of the above-mentioned system are based on the same concept as the method embodiment in the embodiments of the present application, and the technical effects they bring are the same as those of the method embodiment in the embodiments of the present application. For specific contents, please refer to the description in the method embodiment shown in the aforementioned embodiments of the present application, and will not be repeated here.

[0105] The "first" in the names of "first edge node" and "first branch" mentioned in the embodiments of the present application is only used as a name identifier and does not mean the first in order. The same rule applies to "second", "third", etc.

[0106] Through the description of the above implementation methods, it can be known that those skilled in the art can clearly understand that all or part of the steps in the above-mentioned embodiment method can be implemented by means of software plus a general hardware platform. Based on such an understanding, the technical solution of the present application can be embodied in the form of a software product, which can be stored in a storage medium, such as a read-only memory (ROM) / RAM, a magnetic disk, an optical disk, etc., including a number of instructions for a computer device (which can be a personal computer, a server, or a network communication device such as a router) to execute the methods described in each embodiment of the present application or some parts of the embodiments.

[0107] Each embodiment in this specification is described in a progressive manner, and the same or similar parts between the embodiments can refer to each other, and each embodiment focuses on the differences from other embodiments. In particular, for the device embodiment, since it is basically similar to the method embodiment, the description is relatively simple, and the relevant parts can refer to the partial description of the method embodiment. The device embodiment described above is only schematic, in which the modules described as separate components may or may not be physically separated, and the components displayed as modules may or may not be physical modules, that is, they may be located in one place, or they may be distributed on multiple network units. Some or all of the modules can be selected according to actual needs to achieve the purpose of the scheme of this embodiment. Ordinary technicians in this field can understand and implement it without paying creative work.

[0108] The above description is merely an exemplary embodiment of the present application and is not intended to limit the protection scope of the present application.

Claims

1. A financial data processing method, characterized in that: The method is applied to a financial data processing system, which includes: a central node deployed in a financial institution, a first edge node deployed in a first branch, and a second edge node deployed in a second branch, the first branch and the second branch are overseas branches of the financial institution, an instant communication connection is established between the central node and the first edge node and the second edge node, and the financial training data of the first branch and the financial training data of the second branch are restricted from being transmitted to the financial institution or other branches of the financial institution; The method comprises: The central node sends a training instruction to the first edge node and the second edge node, where the training instruction is used to instruct the first edge node and the second edge node to perform local training on the model; The first edge node performs local training on the model according to the financial training data of the first branch to obtain first model parameters, and the second edge node performs local training on the model according to the financial training data of the second branch to obtain second model parameters; The first edge node sends the first model parameter to the central node through the instant communication connection, and the second edge node sends the second model parameter to the central node through the instant communication connection; The central node aggregates model parameters according to the first model parameters and the second model parameters to obtain aggregated model parameters; The first edge node receives the aggregation model parameters sent by the central node; The first edge node updates the model according to the aggregation model parameters to obtain an aggregation model; The first edge node performs a computing task for the financial task data of the first branch according to the aggregation model.

2. The method according to claim 1, characterized in that The central node aggregates model parameters according to the first model parameters and the second model parameters to obtain aggregated model parameters, including: The central node determines a weight value of the first model parameter and a weight value of the second model parameter, wherein the weight value of the first model parameter is positively correlated with the data volume of the financial training data of the first branch, and the weight value of the second model parameter is positively correlated with the data volume of the financial training data of the second branch; The central node performs model parameter aggregation according to the first model parameter, the second model parameter, the weight value of the first model parameter, and the weight value of the second model parameter to obtain the aggregated model parameter.

3. The method according to claim 1, characterized in that The first edge node performs a computing task for the financial task data of the first branch according to the aggregation model, including: The first edge node locally trains the aggregation model according to the financial training data of the first branch to obtain third model parameters; The first edge node updates the aggregation model according to the third model parameter to obtain a personalized model; The first edge node performs a computing task for the financial task data of the first branch according to the personalized model.

4. The method according to claim 1, characterized in that: The first edge node performs a computing task for the financial task data of the first branch according to the aggregation model, including: The first edge node determines a convergence state of the aggregation model; When the convergence state satisfies the convergence condition, the first edge node executes the calculation task for the financial task data of the first branch according to the aggregation model.

5. The method according to claim 1, characterized in that Before the central node aggregates model parameters according to the first model parameters and the second model parameters, the method further includes: The central node performs local training on the model according to the financial training data of the financial institution to obtain fourth model parameters, and the financial training data of the financial institution is restricted from being transmitted to other branches of the financial institution; The central node aggregates model parameters according to the first model parameters and the second model parameters to obtain aggregated model parameters, including: The central node performs model parameter aggregation according to the first model parameter, the second model parameter and the fourth model parameter to obtain the aggregated model parameter.

6. The method according to claim 1, characterized in that The central node sends a training instruction to the first edge node and the second edge node, including: The central node determines whether the first edge node and the second edge node meet a training condition, wherein the training condition includes that the amount of financial training data of the branch to which the edge node belongs meets a data amount range; When the first edge node and the second edge node meet the training condition, the central node sends the training instruction to the first edge node and the second edge node.

7. The method according to claim 1, characterized in that The method further comprises: The central node updates the model according to the aggregation model parameters to obtain the aggregation model; The central node executes the computing task for the financial task data of the financial institution according to the aggregation model.

8. The method according to claim 1, characterized in that The financial task data of the first branch is internal control compliance data of the first branch, and the first edge node performs a computing task for the financial task data of the first branch according to the aggregation model, including: The first edge node obtains the internal control compliance data of the first branch according to a database deployed in the first branch; The first edge node parses the internal control compliance data of the first branch using the aggregation model to obtain a plurality of internal control compliance labels of the first branch.

9. The method according to claim 8, characterized in that The method further comprises: The first edge node sends the multiple internal control compliance tags to the central node; According to the multiple internal control and compliance labels, the central node generates a hot word cloud map, and the hot word cloud map is used to indicate the number of different types of internal control and compliance labels in the multiple internal control and compliance labels.

10. A financial data processing system, characterized in that: The system includes: a central node deployed in a financial institution, a first edge node deployed in a first branch, and a second edge node deployed in a second branch, the first branch and the second branch are overseas branches of the financial institution, an instant communication connection is established between the central node and the first edge node and the second edge node, and the financial training data of the first branch and the financial training data of the second branch are restricted from being transmitted to the financial institution or other branches of the financial institution; The central node is used to send a training instruction to the first edge node and the second edge node, wherein the training instruction is used to instruct the first edge node and the second edge node to perform local training on the model; the central node is also used to perform model parameter aggregation according to the first model parameter and the second model parameter to obtain an aggregated model parameter; The first edge node is used to perform local training on the model according to the financial training data of the first branch to obtain first model parameters; the first edge node is also used to send the first model parameters to the central node through the instant messaging connection; the first edge node is also used to receive the aggregate model parameters sent by the central node; the first edge node is also used to update the model according to the aggregate model parameters to obtain an aggregate model; the first edge node is also used to perform a calculation task for the financial task data of the first branch according to the aggregate model; The second edge node is used to locally train the model according to the financial training data of the second branch to obtain second model parameters; the second edge node is also used to send the second model parameters to the central node through the instant messaging connection.

Images