Financial risk perception-oriented distributed artificial intelligence system and method

Abstract

The invention provides a financial risk perception-oriented distributed artificial intelligence system and method, and the method comprises the steps: carrying out the dynamic compensation of missing data in a financial Internet through the static features of static business data and the time-space correlation features of dynamic time series data in the financial Internet, and obtaining the reconstruction data in the financial Internet; deducing a risk evolution state of the financial internet by using a risk conduction mechanism in the financial risk perception model; when the risk conduction intensity of the risk evolution state meets a preset event driving condition, carrying out combined driving constraint on the risk propagation constraint in the risk evolution state based on the information gain entropy of a local node and the topology collaboration of a global system, and generating a distributed training constraint containing an adaptive gain parameter and a calculation communication resource ratio; and completing distributed training and risk optimization of the financial risk perception model according to the distributed training constraint. Based on the scheme, distributed collaborative modeling of the risk perception model in the Internet financial environment can be realized.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of Internet consumer finance, more specifically, the present application relates to a distributed artificial intelligence system and method for financial risk perception. BACKGROUND

[0002] Financial Internet is an ecological network of deep integration of financial services and Internet technology. Financial Internet integrates payment, lending, investment, insurance and other businesses through online platforms, connecting financial institutions, enterprises and individual users. Its core features are business data, service scene and participation of the public, which improve efficiency and experience, but also bring new challenges such as data security and accelerated risk transmission.

[0003] The traditional distributed training in the financial field adopts a fixed resource allocation strategy, i.e. uniform allocation of computing resources and periodic synchronization, which cannot be flexibly adjusted according to the importance of real-time tasks and system state. In the high-concurrency and dynamic financial environment, the fixed mode leads to low training efficiency and difficulty in prioritizing key risk events, which seriously affects the risk warning and decision-making efficiency of financial institutions and restricts their application effect in complex financial scenarios. Therefore, a more intelligent and flexible resource allocation mechanism is needed to optimize the training process and improve the ability to respond to risks. Therefore, how to realize the distributed collaborative modeling of the risk perception model in the Internet financial environment has become a difficult problem in the industry. SUMMARY

[0004] The present application provides a distributed artificial intelligence system and method for financial risk perception, which can realize distributed collaborative modeling of the risk perception model in the Internet financial environment.

[0005] In the first aspect, the present application provides a financial risk perception model distributed training method based on artificial intelligence, comprising: Obtain multi-source heterogeneous data in financial Internet, the multi-source heterogeneous data includes static business data and dynamic time series data, and identify missing data in the multi-source heterogeneous data; Compensate the missing data by the static features in the static business data and the spatio-temporal correlation features in the dynamic time series data to obtain reconstructed data in the financial Internet, construct a financial risk perception model based on the multi-dimensional risk factors of each data node in the reconstructed data, and then use the risk transmission mechanism in the financial risk perception model to deduce the risk evolution state of the financial Internet; When the risk transmission intensity of the risk evolution state meets the preset event-driven condition, jointly drive and constrain the risk propagation constraint in the risk evolution state based on the information gain entropy of the local node and the topological synergy of the global system, generate a distributed training constraint containing adaptive gain parameters and computing communication resource ratio; Based on the distributed training constraints, the adaptive gain mechanism is invoked to update the parameters of the financial risk perception model and dynamically allocate the computing and communication resources in the financial risk perception model, thereby completing the distributed training and risk optimization of the financial risk perception model.

[0006] In some embodiments, identifying missing data in the multi-source heterogeneous data specifically includes: Perform an integrity scan on the multi-source heterogeneous data and mark randomly missing data blocks; Continuity detection is performed on the multi-source heterogeneous data to identify non-random missing data blocks; Missing data is determined by the randomly missing data blocks and the non-random missing data blocks.

[0007] In some embodiments, the missing data is dynamically compensated by using static features in the static business data and spatiotemporal correlation features in the dynamic time-series data to obtain reconstructed data in the financial internet, specifically including: Extract static features from the static business data and spatiotemporal correlation features from the dynamic time-series data; The static business data in the missing data is compensated using a graph diffusion model to obtain static compensation fragments; Anisotropic diffusion partial differential equations are used to compensate for the dynamic time-series data in the missing data to obtain dynamic compensation segments. The static compensation fragment and the dynamic compensation fragment are merged and aligned into the static business data to obtain the reconstructed data in the financial internet.

[0008] In some embodiments, constructing a financial risk perception model based on the multidimensional risk factors of each data node in the reconstructed data specifically includes: Multidimensional risk factors for each data node are extracted from the reconstructed data, thereby determining the risk transmission relationship between the data nodes; Construct a financial risk graph with data nodes as vertices and risk transmission relationships as edges; A graph neural network is initialized by using each multidimensional risk factor as a vertex feature and the risk transmission relationship as the edge weight. A financial risk perception model is generated based on the financial risk map and the graph neural network.

[0009] In some embodiments, using the risk transmission mechanism in the financial risk perception model to deduce the risk evolution state of the financial internet specifically includes: The risk transmission is iteratively simulated through the message passing mechanism of the graph neural network in the financial risk perception model, thereby iteratively updating the state vector of each data node in the financial Internet. The risk evolution state of the financial internet is determined based on all state vectors.

[0010] In some embodiments, based on the information gain entropy of local nodes and the topological synergy of the global system, a joint driving constraint is applied to the risk propagation constraint in the risk evolution state to generate distributed training constraints containing adaptive gain parameters and computational communication resource allocation. Specifically, this includes: Calculate the information gain entropy of local nodes and evaluate the topological cooperativeness of the global system; The joint driving value of the risk evolution state is calculated using the information gain entropy and the topological coordination. The joint driving value is used to map and generate adaptive gain parameters and computational communication resource allocation, which together constitute distributed training constraints.

[0011] In some embodiments, according to the distributed training constraints, invoking the adaptive gain mechanism to update the parameters of the financial risk perception model and dynamically allocating the computing and communication resources in the financial risk perception model specifically includes: Distributed stochastic gradient descent is performed on each data node participating in the training based on the adaptive gain parameter in the distributed training constraints to update the local model parameters of the financial risk perception model. Based on the computational and communication resource allocation ratio in the distributed training constraints, the total system computational resources are allocated to each data node for computation according to the computational resource allocation ratio, and the total communication bandwidth is allocated between nodes according to the communication resource allocation ratio to complete the synchronization in the financial risk perception model.

[0012] Secondly, this application provides a distributed artificial intelligence system for financial risk perception, including a model training unit, wherein the model training unit includes: The acquisition module is used to acquire multi-source heterogeneous data in the financial internet, including static business data and dynamic time-series data, and to identify missing data in the multi-source heterogeneous data. The processing module is used to dynamically compensate for the missing data by using the static features in the static business data and the spatiotemporal correlation features in the dynamic time series data to obtain reconstructed data in the financial internet. Based on the multidimensional risk factors of each data node in the reconstructed data, a financial risk perception model is constructed, and then the risk transmission mechanism in the financial risk perception model is used to deduce the risk evolution state of the financial internet. The processing module is also used to jointly drive the risk propagation constraints in the risk evolution state based on the information gain entropy of local nodes and the topological synergy of the global system when the risk propagation intensity of the risk evolution state meets the preset event-driven conditions, and generate distributed training constraints including adaptive gain parameters and computing and communication resource allocation. The execution module is used to update the parameters of the financial risk perception model by calling the adaptive gain mechanism according to the distributed training constraints, and to dynamically allocate the computing and communication resources in the financial risk perception model, thereby completing the distributed training and risk optimization of the financial risk perception model.

[0013] Thirdly, this application provides a computer device, which includes a memory and a processor. The memory is used to store a computer program, and the processor is used to call and run the computer program from the memory, so that the computer device executes the above-described distributed training method for a financial risk perception model based on artificial intelligence.

[0014] Fourthly, this application provides a computer-readable storage medium storing instructions or code that, when executed on a computer, cause the computer to implement the aforementioned distributed training method for an artificial intelligence-based financial risk perception model.

[0015] The technical solutions provided by the embodiments disclosed in this application have the following beneficial effects: This application provides a distributed artificial intelligence system and method for financial risk perception. It acquires multi-source heterogeneous data from the financial internet, including static business data and dynamic time-series data, and identifies missing data within this data. The missing data is dynamically compensated using static features in the static business data and spatiotemporal correlation features in the dynamic time-series data to obtain reconstructed data from the financial internet. Based on the multidimensional risk factors of each data node in the reconstructed data, a financial risk perception model is constructed. The risk transmission mechanism within this model is then used to deduce the risk evolution state of the financial internet. When the risk transmission intensity of the risk evolution state meets preset event-driven conditions, joint driving constraints are applied to the risk propagation constraints in the risk evolution state based on the information gain entropy of local nodes and the topological synergy of the global system, generating distributed training constraints that include adaptive gain parameters and computational and communication resource allocation. According to these distributed training constraints, the adaptive gain mechanism is invoked to update the parameters of the financial risk perception model, and computational and communication resources within the model are dynamically allocated, completing the distributed training and risk optimization of the financial risk perception model.

[0016] Therefore, in this application, based on the distributed training constraints, an adaptive gain mechanism is invoked to update the parameters of the financial risk perception model and dynamically allocate the computational and communication resources in the financial risk perception model, thus completing the distributed training and risk optimization of the financial risk perception model. First, determining the risk evolution state yields a comprehensive situation map that quantitatively reflects the real-time dynamics and distribution of financial risk across the entire network, providing a unified and objective decision-making benchmark for distributed collaborative modeling. The risk evolution state not only identifies the current risk hotspots but also reveals the path and intensity of risk transmission, enabling collaborative training of various local models distributed throughout the system. This ensures that all local models have a consistent understanding of the global risk situation, avoiding decision conflicts or blind spots caused by different perspectives. Then, determining the distributed training constraints yields a dynamic resource scheduling and model optimization strategy that matches the real-time risk situation. This approach achieves a synergistic balance between training efficiency and the accuracy of risk response. By introducing an event-driven mechanism, when the risk evolution exceeds a threshold, a constraint generation process is automatically triggered. Based on the information gain entropy of local nodes (identifying which nodes' model updates are most valuable globally) and global topological synergy (evaluating network communication efficiency), training constraints containing adaptive gain parameters and computational communication resource allocation are dynamically generated. This essentially installs an intelligent speed controller in the distributed training system. When risks escalate, resources are prioritized for key model updates on nodes with high information value and good network reachability. This ensures that limited resources are used for the most urgent and effective model optimization, enabling the entire distributed system to concentrate its efforts on risk management and respond to risk changes. Ultimately, this improves the overall efficiency and robustness of collaborative modeling. In summary, based on the above scheme, distributed collaborative modeling of risk perception models in the Internet finance environment can be achieved. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is an exemplary flowchart of a distributed training method for an artificial intelligence-based financial risk perception model, as shown in some embodiments of this application. Figure 2 This is a flowchart illustrating the process of determining distributed training constraints according to some embodiments of this application; Figure 3 This is a schematic diagram of the structure of a model training unit according to some embodiments of this application; Figure 4 This is a schematic diagram of the structure of a computer device that implements a distributed training method for an artificial intelligence-based financial risk perception model, according to some embodiments of this application. Detailed Implementation

[0019] To better understand the technical solution of this application, the technical solution of this application will be described in detail below with reference to the accompanying drawings and specific embodiments.

[0020] refer to Figure 1 The figure is an exemplary flowchart of a distributed training method for an artificial intelligence-based financial risk perception model, according to some embodiments of this application. The distributed training method for the artificial intelligence-based financial risk perception model mainly includes the following steps: In step 101, multi-source heterogeneous data from the financial internet is acquired. The multi-source heterogeneous data includes static business data and dynamic time-series data, and missing data in the multi-source heterogeneous data is identified.

[0021] It should be noted that, in this application, multi-source heterogeneous data refers to a collection of financial-related data with wide sources and different structures and formats; static business data refers to financial entity attribute data that is relatively stable and has a low frequency of change within a certain period; and dynamic time series data refers to financial process and behavior data that changes continuously over time and has obvious order and timeliness.

[0022] In practice, data is collected from various data sources in the financial internet by configuring data interfaces and web crawlers. For static business data stored in databases, such as corporate balance sheets and customer basic information tables, batch extraction is performed using a structured query language interface. For dynamic time-series data, such as real-time transaction records, market data sequences, and social media sentiment streams, continuous subscription is performed through application programming interfaces or real-time collection is performed through a streaming processing framework. The combination of static business data and dynamic time-series data is considered as multi-source heterogeneous data in the financial internet.

[0023] In some embodiments, identifying missing data in the multi-source heterogeneous data can be achieved by the following steps: Perform an integrity scan on the multi-source heterogeneous data and mark randomly missing data blocks; Continuity detection is performed on the multi-source heterogeneous data to identify non-random missing data blocks; Missing data is determined by the randomly missing data blocks and the non-random missing data blocks.

[0024] It should be noted that, in this application, missing data is a general term for all data gaps that have been identified and marked; random missing data blocks are data gaps that are unrelated to the observed values ​​themselves and are caused entirely by random factors; non-random missing data blocks are data gaps that are systematically related to the unobserved data values ​​themselves.

[0025] In practical implementation, firstly, by traversing each field and record in the multi-source heterogeneous data, comparisons are performed according to pre-defined integrity rules. For static business data, key fields (e.g., enterprise identifier, timestamp, numerical indicators) are checked for null values, placeholders, or outliers that clearly exceed reasonable ranges. For dynamic time-series data, the timestamp sequence of data points is checked for expected intervals. When missing data is found and statistical analysis confirms that the missing pattern is not significantly related to the data value itself, this part of the data is marked as a randomly missing data block, and its location and range are recorded. Then, for dynamic time-series data, the continuity of the time series in the dynamic time-series data is analyzed. This involves calculating the time interval between adjacent data points to identify discontinuities exceeding the normal data collection frequency threshold. Simultaneously, it analyzes the business context in which the data was generated. For example, it checks whether consecutive data gaps occur after events within a preset range (e.g., system failures, market trading halts). When data gaps are determined to be due to a systemic reason (e.g., data not recorded due to excessively high / low values), this portion of data is marked as a non-random missing data block, and its missing context information is recorded. Finally, all randomly and non-randomly missing data blocks are aggregated and integrated, i.e., matched and deduplicated using data identifiers (such as primary keys and timestamps) to identify the missing data.

[0026] In step 102, the missing data is dynamically compensated by the static features in the static business data and the spatiotemporal correlation features in the dynamic time series data to obtain reconstructed data in the financial internet. Based on the multidimensional risk factors of each data node in the reconstructed data, a financial risk perception model is constructed, and then the risk transmission mechanism in the financial risk perception model is used to deduce the risk evolution state of the financial internet.

[0027] In some embodiments, the missing data can be dynamically compensated by using static features in the static business data and spatiotemporal correlation features in the dynamic time-series data to obtain reconstructed data in the financial internet. This can be achieved through the following steps: Extract static features from the static business data and spatiotemporal correlation features from the dynamic time-series data; The static business data in the missing data is compensated using a graph diffusion model to obtain static compensation fragments; Anisotropic diffusion partial differential equations are used to compensate for the dynamic time-series data in the missing data to obtain dynamic compensation segments. The static compensation fragment and the dynamic compensation fragment are merged and aligned into the static business data to obtain the reconstructed data in the financial internet.

[0028] It should be noted that in this application, the reconstructed data is a complete dataset that can be used for subsequent modeling and analysis; static features are statistics and category labels used to describe the inherent attributes of financial entities; spatiotemporal correlation features are features that can simultaneously reflect the data's changing patterns over time and its spatial or topological correlation; static compensation fragments are repaired data blocks used to fill in missing parts of static business data; and dynamic compensation fragments are repaired time series data segments used to fill in missing parts of dynamic time series data.

[0029] In specific implementation, firstly, extracting static features from the static business data and spatiotemporal correlation features from the dynamic time-series data can be achieved in the following way: For static business data, calculate the statistics (e.g., mean, variance, quantiles) of numerical fields in the static business data and encode their categorical fields, using the set of statistics for all numerical fields as static features; for dynamic time-series data, use the sliding window technique to calculate the statistical characteristics, trends, and periodic indicators of the time series within each window, and calculate the correlation strength between nodes, for example, using a graph structure to calculate the correlation strength between data nodes, thereby using the set of statistical characteristics, trends, periodic indicators, and all correlation strengths as spatiotemporal features. The association features; secondly, the static business data in the missing data is compensated using a graph diffusion model. The static compensation fragment can be obtained in the following way: the entities (e.g., enterprises, individuals) in the static business data are regarded as graph nodes. Based on the known relationships between entities (e.g., equity associations, transactions), an adjacency matrix of the graph is constructed. Then, the known and complete data in the static business data is used as the information source. In the graph diffusion model, the process of information propagation from known data nodes to missing nodes in multiple iterations is simulated. In each iteration, the information of each data node is updated to a weighted average of its own information and the information of its neighboring nodes. After a sufficient number of iterations, the information reaches a stable state in the entire graph. The information value converged by the missing node is used as the compensation value, and the set of compensation values ​​for all missing nodes is taken as the static compensation segment. Then, the dynamic time-series data in the missing data is compensated using anisotropic diffusion partial differential equations. The dynamic compensation segment can be obtained by the following method: the dynamic time-series data is regarded as a two-dimensional data field (time dimension and index dimension) that changes with time. The data of the known region in the data field is used as the boundary condition and initial condition. The anisotropic diffusion partial differential equation is numerically solved. The core of the anisotropic diffusion partial differential equation is that its diffusion coefficient is a function of the local gradient of the data. Strong diffusion is performed in flat regions with small gradients to smooth noise and fill in missing data, and strong diffusion is performed in regions with large gradients to smooth noise and fill in missing data. Weak diffusion is applied to edge regions to protect the abrupt changes in data. Through iterative solving, the data values ​​in the missing regions are gradually filled and smoothed out from the known regions. When the solution process converges, the missing time series segments are repaired, and the set of data with all repaired missing time series segments is taken as a dynamic compensation fragment. Finally, the static compensation fragment and the dynamic compensation fragment are fused and aligned into the static business data to obtain the reconstructed data in the financial internet. This can be achieved in the following way: matching the static compensation fragment to the corresponding missing position in the static business data based on a unique identifier (e.g., enterprise code); matching the dynamic compensation fragment to the corresponding missing position in the dynamic time series data based on the same identifier and timestamp.Simultaneously, consistency checks are performed, such as verifying whether the attributes of the same entity after static compensation and its behavior after dynamic compensation are logically reasonable. Then, the fully compensated static business data and dynamic time-series data are integrated according to their relationships to form a complete dataset containing all entities, all time points, and all indicators—this is the final reconstructed data required for the financial internet.

[0030] In some embodiments, constructing a financial risk perception model based on the multidimensional risk factors of each data node in the reconstructed data can be achieved through the following steps: Multidimensional risk factors for each data node are extracted from the reconstructed data, thereby determining the risk transmission relationship between the data nodes; Construct a financial risk graph with data nodes as vertices and risk transmission relationships as edges; A graph neural network is initialized by using each multidimensional risk factor as a vertex feature and the risk transmission relationship as the edge weight. A financial risk perception model is generated based on the financial risk map and the graph neural network.

[0031] It should be noted that, in this application, the financial risk perception model is a graph neural network model capable of learning from the financial risk graph and outputting risk perception results; the multidimensional risk factor is an indicator that quantifies the risk level of data nodes from different dimensions; the risk transmission relationship is a substantial association between data nodes that can lead to the transmission of risk from one party to another; the financial risk graph is a data model that uses a graph structure to describe the distribution and transmission path of financial risks, where vertices represent risk carriers and edges represent the probability of risk transmission; the graph neural network is a deep learning model used to process graph structure data, which can learn the representation of vertices and the graph by aggregating information from neighboring nodes.

[0032] In specific implementation, firstly, multidimensional risk factors for each data node are extracted from the reconstructed data. The risk transmission relationship between data nodes can be determined as follows: For each data node (e.g., an enterprise), a set of preset risk indicators is statistically analyzed from the reconstructed data. These indicators include financial dimensions (e.g., debt-to-equity ratio, cash flow ratio), market dimensions (e.g., stock price volatility, price-to-earnings ratio), and correlation dimensions (e.g., the number of guarantee circles). The set of all risk indicators serves as the multidimensional risk factor for each data node. Based on the correlation information in the reconstructed data (e.g., equity investment relationships, major transactions, supply chain dependencies), a judgment is made according to preset rules (e.g., a risk transmission path is identified if the shareholding ratio exceeds a specified threshold). This determines whether a risk transmission relationship exists between any two data nodes and clarifies the direction of the transmission relationship. The risk transmission relationship between each data node can then be obtained. Secondly, a financial data structure is constructed with data nodes as vertices and risk transmission relationships as edges. A financial risk graph can be implemented as follows: Each data node is mapped to a vertex in the graph, and each established risk transmission relationship is mapped to a directed edge connecting two vertices. Ultimately, all vertices and edges together constitute a complete financial risk graph that can visually represent the potential transmission paths of financial risks in the network. Then, using multidimensional risk factors as vertex features and risk transmission relationships as edge weights, a graph neural network can be initialized as follows: Each vertex in the financial risk graph is assigned a feature vector, the value of which is the multidimensional risk factor of the corresponding data node. Simultaneously, each directed edge in the financial risk graph is assigned a weight value, which can be pre-calculated based on the strength of the transmission relationship (e.g., transaction amount percentage, shareholding ratio). A basic graph neural network architecture (e.g., graph convolutional network or graph attention network) is selected, and the financial risk graph with assigned vertex features and edge weights is input into this architecture to complete the parameter initialization. The parameter-initialized graph neural network architecture is then used as the graph neural network.Finally, the financial risk perception model generated based on the financial risk graph and the graph neural network can be implemented in the following way: The initialized graph neural network is combined with the financial risk graph as input data. Through a specified graph neural network forward propagation algorithm (e.g., message passing paradigm of graph convolutional networks, weighted aggregation mechanism of graph attention networks, or neighborhood sampling strategy of graph sampling and aggregation), the model can aggregate and transmit risk information along the edge structure of the graph, thereby learning the deep risk representation of each vertex (data node) in its network context. This trainable model, capable of processing graph data, fusing multi-dimensional risk factors, and considering risk transmission effects, can then serve as the final required financial risk perception model.

[0033] In some embodiments, the risk transmission mechanism in the financial risk perception model can be used to deduce the risk evolution state of the financial internet through the following steps: The risk transmission is iteratively simulated through the message passing mechanism of the graph neural network in the financial risk perception model, thereby iteratively updating the state vector of each data node in the financial Internet. The risk evolution state of the financial internet is determined based on all state vectors.

[0034] It should be noted that in this application, risk evolution state; message passing mechanism: refers to the core algorithm process in graph neural networks that updates the node state by passing and aggregating information (i.e., messages) of adjacent nodes along the graph edges; state vector is a hidden feature vector used to characterize the risk status of the corresponding data node at the current moment.

[0035] In specific implementation, firstly, the risk transmission is iteratively simulated through the message passing mechanism of the graph neural network in the financial risk perception model. This iterative update of the state vector of each data node in the financial internet can be achieved in the following way: The simulation process of the graph neural network in the financial risk perception model is initiated. In each iteration, each data node (corresponding to a vertex in the graph) receives state vector information from its neighboring nodes (nodes directly connected by edges) with which it has a risk transmission relationship. Each data node aggregates all the received neighbor information with its own current state vector (e.g., by averaging or weighted summation), and then passes it through a learnable nonlinear transformation function to generate the updated state vector of the corresponding data node in this iteration. Through multiple rounds of such iterations, risk information is propagated from local to global in the graph structure of the financial internet, thus simulating the dynamic transmission process of risk in the network and enabling each data node to... The state vector of a node contains its own risk information and that of its local network, thus enabling iterative updates of the state vector of each data node in the financial internet. The risk evolution state of the financial internet can then be determined based on all state vectors in the following manner: After the message passing process of the graph neural network reaches a preset number of iterations or converges, each data node possesses a state vector updated to the final iteration. This state vector encodes its risk status after being affected by the global risk transmission within the network. A global pooling function (e.g., calculating the mean of all node state vectors or selecting their maximum norm) is used to aggregate the final state vectors of all data nodes. The result of this aggregation is a comprehensive vector or scalar, integrating the risk information of all network nodes. Therefore, this aggregation result can be used as the risk evolution state of the financial internet, reflecting the overall risk level and distribution characteristics of the system.

[0036] In step 103, when the risk propagation intensity of the risk evolution state meets the preset event-driven conditions, the risk propagation constraints in the risk evolution state are jointly driven by the information gain entropy of local nodes and the topological synergy of the global system, generating distributed training constraints that include adaptive gain parameters and computing and communication resource allocation.

[0037] It should be noted that in this application, when the risk transmission intensity of the risk evolution state meets the preset event-driven conditions, it means that the overall risk level of the financial internet has reached a critical threshold, indicating that the risk is rapidly accumulating or spreading in the network. At this time, if a fixed, predefined distributed training strategy is continued, it may not be able to effectively cope with the rapidly changing risk situation, and may even lead to a decrease in model training efficiency or waste of resources. The fixed training strategy cannot adaptively balance the urgency of model updates (reflected by the information gain entropy of local risk nodes) with the efficiency of distributed system collaboration (reflected by the global network topology collaboration), and cannot take the physical transmission law of financial risk itself as a constraint condition. By executing joint driving constraints, an optimal training strategy that is highly matched with the current risk evolution state can be dynamically generated, thereby ensuring that the distributed artificial intelligence system accurately invests its limited computing and communication resources in the most critical model update task, and achieves accurate and efficient optimization of the financial risk perception model to cope with the current high-risk situation.

[0038] In some embodiments, based on the information gain entropy of local nodes and the topological synergy of the global system, joint driving constraints are applied to the risk propagation constraints in the risk evolution state, generating distributed training constraints that include adaptive gain parameters and computational communication resource allocation, referencing... Figure 2 The diagram is a flowchart illustrating the process of determining distributed training constraints in some embodiments of this application. In this embodiment, determining distributed training constraints can be achieved using the following steps: In step 1031, the information gain entropy of local nodes is calculated, and the topological cooperativeness of the global system is evaluated; In step 1032, the joint driving value of the risk evolution state is calculated using the information gain entropy and the topological coordination. In step 1033, the joint driving value is used to map and generate adaptive gain parameters and computational communication resource allocation, which together constitute distributed training constraints.

[0039] It should be noted that, in this application, distributed training constraints are key parameter settings used to guide and limit the distributed training process; adaptive gain parameters are adjustable parameters used to control the step size of model parameter updates; computational and communication resource allocation is the proportional planning of system resources allocated to model computation tasks and inter-node communication tasks in distributed training; information gain entropy is an indicator used to quantify the value of information contained in the local model update of a single node; topological synergy is an indicator used to measure the network connection structure and communication efficiency of the entire distributed system; and joint driving value is a scalar value used to make decisions on adjusting training strategies.

[0040] In specific implementation, firstly, calculating the information gain entropy of local nodes and evaluating the topological coherence of the global system can be achieved in the following way: For each local node participating in the calculation, collect the changes in its local model parameters relative to the last global synchronization, calculate the probability distribution of these changes, and use the information entropy formula to calculate the result, which is the information gain entropy of that node; obtain a topological graph describing the connection relationships between all computing nodes, calculate graph theory indices such as algebraic connectivity of the topological graph, and use this index value as a measure of the topological coherence of the global system to evaluate the overall communication efficiency of the network structure. In other embodiments, the topological coherence of the global system can also be obtained directly from the central control console of the Internet finance system, which is not limited here; then, calculating the joint driving value of the risk evolution state through the information gain entropy and the topological coherence can be achieved in the following way: aggregate the information gain entropy of all local nodes (e.g., take the mean or minimum value), and compare it with the evaluated global topological coherence. The combined result is weighted and summed with a scalarized risk intensity index (e.g., the norm of the state vector) extracted from the risk evolution state. A series of preset weight coefficients used for the weighted summation reflect different emphases on the three dimensions of information value, network efficiency, and risk level (which can be preset through historical experience). The result of the weighted summation is used as the joint driving value. Finally, the adaptive gain parameters and the ratio of computing and communication resources are mapped and generated through the joint driving value to form the distributed training constraints. This can be achieved by querying a predefined mapping table based on the joint driving value. This mapping table defines the optimal adaptive gain parameters (e.g., the specific value of the learning rate) corresponding to different ranges of joint driving values ​​and the resource allocation ratio between computing and communication tasks. Based on the current joint driving value falling within the range of this mapping table, the corresponding specific parameter value is retrieved or calculated. The automatically generated adaptive gain parameters and the set of computing and communication resource ratio settings are used as the distributed training constraints.

[0041] In step 104, based on the distributed training constraints, the adaptive gain mechanism is invoked to update the parameters of the financial risk perception model, and the computing and communication resources in the financial risk perception model are dynamically allocated to complete the distributed training and risk optimization of the financial risk perception model.

[0042] In some embodiments, updating the parameters of the financial risk perception model by invoking the adaptive gain mechanism according to the distributed training constraints, and dynamically allocating the computing and communication resources in the financial risk perception model, can be achieved through the following steps: Distributed stochastic gradient descent is performed on each data node participating in the training based on the adaptive gain parameter in the distributed training constraints to update the local model parameters of the financial risk perception model. Based on the computational and communication resource allocation ratio in the distributed training constraints, the total system computational resources are allocated to each data node for computation according to the computational resource allocation ratio, and the total communication bandwidth is allocated between nodes according to the communication resource allocation ratio to complete the synchronization in the financial risk perception model.

[0043] It should be noted that, in this application, based on stochastic optimization theory and resource allocation theory, the distributed training constraints are decoupled into two parallel subtasks: local update and global synchronization, to achieve collaborative optimization of the financial risk perception model under resource-constrained conditions. The specific process is as follows: First, each data node participating in the training independently calculates the gradient of the loss function using its local data and updates the parameters of its held copy of the financial risk perception model, based on the adaptive gain parameter (i.e., learning rate) specified in the distributed training constraints. Simultaneously, a central resource scheduler, according to the computational and communication resource allocation ratio explicitly stated in the distributed training constraints, proportionally divides the total computational resources of the cluster (e.g., when using GPUs) into computational and communication shares. The computational share is allocated to each data node to execute the aforementioned parameter update task, while the communication share is used to support each data node in synchronizing the updated local model parameters to the global model through a global aggregation operation (e.g., weighted average) at a specified synchronization point. By iteratively executing the above loop of local computation and global synchronization until the financial risk perception model converges, the distributed training and risk optimization of the financial risk perception model are finally completed.

[0044] Furthermore, in another aspect of this application, in some embodiments, this application provides a distributed artificial intelligence system for financial risk perception, which includes a model training unit, referencing... Figure 3 The figure is a schematic diagram of the structure of a model training unit according to some embodiments of this application. The model training unit includes: an acquisition module 201, a processing module 202, and an execution module 203, which are described below: The acquisition module 201 in this application is mainly used to acquire multi-source heterogeneous data in the financial Internet, the multi-source heterogeneous data including static business data and dynamic time series data, and to identify missing data in the multi-source heterogeneous data. Processing module 202, in this application, is used to dynamically compensate for the missing data by using the static features in the static business data and the spatiotemporal correlation features in the dynamic time series data to obtain reconstructed data in the financial internet. Based on the multidimensional risk factors of each data node in the reconstructed data, a financial risk perception model is constructed, and then the risk transmission mechanism in the financial risk perception model is used to deduce the risk evolution state of the financial internet. It should be noted that the processing module 202 is also used to jointly drive the risk propagation constraints in the risk evolution state based on the information gain entropy of local nodes and the topological synergy of the global system when the risk propagation intensity of the risk evolution state meets the preset event-driven conditions, and generate distributed training constraints including adaptive gain parameters and computing and communication resource allocation. The execution module 203 in this application is mainly used to update the parameters of the financial risk perception model by calling the adaptive gain mechanism according to the distributed training constraints, and to dynamically allocate the computing and communication resources in the financial risk perception model, so as to complete the distributed training and risk optimization of the financial risk perception model.

[0045] The foregoing has detailed examples of a distributed artificial intelligence system and method for financial risk perception provided in the embodiments of this application. It is understood that the corresponding apparatus, in order to achieve the above functions, includes hardware structures and / or software modules corresponding to the execution of each function. Those skilled in the art should readily recognize that, based on the units and algorithm steps of the examples described in conjunction with the embodiments disclosed herein, this application can be implemented in hardware or a combination of hardware and computer software. Whether a function is executed in hardware or by computer software driving hardware depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0046] In some embodiments, this application also provides a computer device, the computer device including a memory and a processor, the memory for storing a computer program, and the processor for calling and running the computer program from the memory, so that the computer device executes the above-described distributed training method for an artificial intelligence-based financial risk perception model.

[0047] In some embodiments, reference Figure 4 The dashed lines in the figure indicate that the unit or module is optional. This figure is a schematic diagram of the structure of a computer device implementing a distributed training method for an artificial intelligence-based financial risk perception model according to an embodiment of this application. The distributed training method for an artificial intelligence-based financial risk perception model described in the above embodiments can be achieved through… Figure 4 The computer device shown is used to implement this, and the computer device includes at least one processor 301, a memory 302 and at least one communication unit 305. The computer device may be a terminal device, a server or a chip.

[0048] Processor 301 can be a general-purpose processor or a special-purpose processor. For example, processor 301 can be a central processing unit (CPU), which can be used to control computer devices, execute software programs, and process data from software programs. The computer device may also include a communication unit 305 for inputting (receiving) and outputting (transmitting) signals.

[0049] For example, the computer device may be a chip, and the communication unit 305 may be the input and / or output circuit of the chip, or the communication unit 305 may be the communication interface of the chip, which may be a component of a terminal device, network device or other device.

[0050] For example, the computer device may be a terminal device or a server, and the communication unit 305 may be a transceiver of the terminal device or the server, or the communication unit 305 may be a transceiver circuit of the terminal device or the server.

[0051] The computer device may include one or more memories 302 storing a program 304. The program 304 can be executed by a processor 301 to generate instructions 303, causing the processor 301 to execute the method described in the above method embodiments according to the instructions 303. Optionally, the memory 302 may also store data (such as a target audit model). Optionally, the processor 301 may also read data stored in the memory 302, which may be stored at the same storage address as the program 304, or it may be stored at a different storage address than the program 304.

[0052] The processor 301 and memory 302 can be configured separately or integrated together, for example, integrated on the system on chip (SOC) of the terminal device.

[0053] It should be understood that each step of the above method embodiment can be completed by hardware logic circuits or software instructions in the processor 301. The processor 301 can be a CPU, a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or other programmable logic devices, such as discrete gate, transistor logic devices, or discrete hardware components.

[0054] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0055] For example, in some embodiments, this application also provides a computer-readable storage medium storing instructions or code that, when executed on a computer, cause the computer to implement the above-described distributed training method for an artificial intelligence-based financial risk perception model.

[0056] Although preferred embodiments of this application have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of this application.

[0057] Obviously, those skilled in the art can make various modifications and variations to this application without departing from the spirit and scope of this application. Therefore, if such modifications and variations fall within the scope of the claims of this application and their equivalents, this application also intends to include such modifications and variations.

Claims

1. A distributed training method for a financial risk perception model based on artificial intelligence, used for distributed training of the financial risk perception model in a distributed artificial intelligence system for financial risk perception, characterized in that, The method includes the following steps: Acquire multi-source heterogeneous data from the financial internet, the multi-source heterogeneous data including static business data and dynamic time-series data, and identify missing data in the multi-source heterogeneous data; By dynamically compensating for the missing data through the static features in the static business data and the spatiotemporal correlation features in the dynamic time series data, reconstructed data in the financial internet is obtained. Based on the multidimensional risk factors of each data node in the reconstructed data, a financial risk perception model is constructed, and then the risk transmission mechanism in the financial risk perception model is used to deduce the risk evolution state of the financial internet. When the risk propagation intensity of the risk evolution state meets the preset event-driven conditions, based on the information gain entropy of local nodes and the topological synergy of the global system, the risk propagation constraints in the risk evolution state are jointly driven to generate distributed training constraints that include adaptive gain parameters and computing and communication resource allocation. Based on the distributed training constraints, the adaptive gain mechanism is invoked to update the parameters of the financial risk perception model and dynamically allocate the computing and communication resources in the financial risk perception model, thereby completing the distributed training and risk optimization of the financial risk perception model.

2. The method as described in claim 1, characterized in that, Identifying missing data in the multi-source heterogeneous data specifically includes: Perform an integrity scan on the multi-source heterogeneous data and mark randomly missing data blocks; Continuity detection is performed on the multi-source heterogeneous data to identify non-random missing data blocks; Missing data is determined by the randomly missing data blocks and the non-random missing data blocks.

3. The method as described in claim 1, characterized in that, By dynamically compensating for the missing data using static features in the static business data and spatiotemporal correlation features in the dynamic time-series data, the reconstructed data in the financial internet specifically includes: Extract static features from the static business data and spatiotemporal correlation features from the dynamic time-series data; The static business data in the missing data is compensated using a graph diffusion model to obtain static compensation fragments; Anisotropic diffusion partial differential equations are used to compensate for the dynamic time-series data in the missing data to obtain dynamic compensation segments. The static compensation fragment and the dynamic compensation fragment are merged and aligned into the static business data to obtain the reconstructed data in the financial internet.

4. The method as described in claim 1, characterized in that, Based on the multidimensional risk factors of each data node in the reconstructed data, the construction of the financial risk perception model specifically includes: Multidimensional risk factors for each data node are extracted from the reconstructed data, thereby determining the risk transmission relationship between the data nodes; Construct a financial risk graph with data nodes as vertices and risk transmission relationships as edges; A graph neural network is initialized by using each multidimensional risk factor as a vertex feature and the risk transmission relationship as the edge weight. A financial risk perception model is generated based on the financial risk map and the graph neural network.

5. The method as described in claim 1, characterized in that, Using the risk transmission mechanism in the aforementioned financial risk perception model to deduce the risk evolution state of the financial internet specifically includes: The risk transmission is iteratively simulated through the message passing mechanism of the graph neural network in the financial risk perception model, thereby iteratively updating the state vector of each data node in the financial Internet. The risk evolution state of the financial internet is determined based on all state vectors.

6. The method as described in claim 1, characterized in that, Based on the information gain entropy of local nodes and the topological synergy of the global system, a joint driving constraint is applied to the risk propagation constraint in the risk evolution state, generating distributed training constraints that include adaptive gain parameters and computational communication resource allocation. Specifically, this includes: Calculate the information gain entropy of local nodes and evaluate the topological cooperativeness of the global system; The joint driving value of the risk evolution state is calculated using the information gain entropy and the topological coordination. The joint driving value is used to map and generate adaptive gain parameters and computational communication resource allocation, which together constitute distributed training constraints.

7. The method as described in claim 1, characterized in that, Based on the aforementioned distributed training constraints, the adaptive gain mechanism is invoked to update the parameters of the financial risk perception model, and the computational and communication resources in the financial risk perception model are dynamically allocated, specifically including: Distributed stochastic gradient descent is performed on each data node participating in the training based on the adaptive gain parameter in the distributed training constraints to update the local model parameters of the financial risk perception model. Based on the computational and communication resource allocation ratio in the distributed training constraints, the total system computational resources are allocated to each data node for computation according to the computational resource allocation ratio, and the total communication bandwidth is allocated between nodes according to the communication resource allocation ratio to complete the synchronization in the financial risk perception model.

8. A distributed artificial intelligence system for financial risk perception, the distributed artificial intelligence system for financial risk perception including a model training unit, characterized in that, The model training unit includes: The acquisition module is used to acquire multi-source heterogeneous data in the financial internet, including static business data and dynamic time-series data, and to identify missing data in the multi-source heterogeneous data. The processing module is used to dynamically compensate for the missing data by using the static features in the static business data and the spatiotemporal correlation features in the dynamic time series data to obtain reconstructed data in the financial internet. Based on the multidimensional risk factors of each data node in the reconstructed data, a financial risk perception model is constructed, and then the risk transmission mechanism in the financial risk perception model is used to deduce the risk evolution state of the financial internet. The processing module is also used to jointly drive the risk propagation constraints in the risk evolution state based on the information gain entropy of local nodes and the topological synergy of the global system when the risk propagation intensity of the risk evolution state meets the preset event-driven conditions, and generate distributed training constraints including adaptive gain parameters and computing and communication resource allocation. The execution module is used to update the parameters of the financial risk perception model by calling the adaptive gain mechanism according to the distributed training constraints, and to dynamically allocate the computing and communication resources in the financial risk perception model, thereby completing the distributed training and risk optimization of the financial risk perception model.

9. A computer device, characterized in that, The computer device includes a memory and a processor. The memory is used to store computer programs, and the processor is used to call and run the computer programs from the memory, so that the computer device executes the distributed training method for the financial risk perception model based on artificial intelligence as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores instructions or code that, when executed on a computer, cause the computer to implement the distributed training method for the artificial intelligence-based financial risk perception model as described in any one of claims 1 to 7.

Images