Business risk management method and device, computer program product and electronic equipment

Abstract

The invention discloses a business risk management method and device, a computer program product and electronic equipment. The method relates to the field of artificial intelligence and big data, and comprises the following steps: obtaining business data, and encoding the business data to obtain a business vector; the business vector is input into a target model to obtain a risk prediction level, the target model comprises an input layer, a hidden layer, an attention mechanism layer and an output layer, the input layer is used for extracting a state vector from the business vector, the hidden layer is used for extracting hidden features from the state vector, the attention mechanism layer is used for calculating an attention weight, and the output layer is used for outputting the attention weight; the target feature is determined through the hidden feature and the attention weight, the attention weight is determined by the similarity between the business risk index and the hidden feature and the position code of the hidden feature, and the output layer outputs a risk prediction level based on the target feature; and determining a target risk management strategy through the risk prediction level. Through the method and the device, the problem of low business risk management efficiency in related technologies is solved.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of artificial intelligence and big data, in particular, to a business risk management method and device, a computer program product and an electronic device. BACKGROUND

[0002] With the development of technology, the business risks faced by enterprises have increased significantly. With the continuous improvement of laws and regulations and the continuous strengthening of supervision, enterprises in various industries must strictly follow the corresponding compliance requirements in financial, legal, operational and other business fields to prevent legal sanctions, economic losses and damage to brand reputation. For example, if a financial institution does not comply with transaction compliance, it may face heavy penalties from regulatory agencies or even business restrictions; if a pharmaceutical company violates the sales regulations of drug approval, it may face product recall and litigation compensation crises.

[0003] In related technologies, the business risk management method of an enterprise relies on manual review and fixed rule judgment. Compliance officers collect various documents such as financial statements, contract texts, and operation logs from various departments for detailed review. This process highly depends on the professional skills and concentration of humans, but due to the large amount of data and long review time, the risk assessment is lagging and highly subjective. Enterprises develop detailed compliance rules, such as transaction amount limits and operation approval processes, to detect risks by automatically comparing data and rules, but the static nature of the rules makes it difficult to respond to sudden changes in business models and unknown risks.

[0004] However, the risk management method in related technologies is difficult to keep up with the pace of data growth, resulting in delayed and incomplete risk assessment. Manual judgment is easily affected by individual differences, and different auditors have different standards, reducing the reliability of risk assessment. Rule-based systems can quickly respond to preset scenarios, but they are not effective in the face of new business models or regulatory changes, and cannot capture non-rule-based new risks.

[0005] To address the problem of low efficiency of business risk management in related technologies, no effective solutions have been proposed so far. SUMMARY

[0006] The main purpose of the present application is to provide a business risk management method, device, computer program product and electronic device to solve the problem of low efficiency of business risk management in related technologies.

[0007] To achieve the above object, according to an aspect of the present application, a business risk management method is provided. The method comprises: obtaining business data, and performing encoding processing on the business data to obtain a business vector; inputting the business vector into a target model to obtain a risk prediction level, wherein the target model comprises an input layer, a hidden layer, an attention mechanism layer and an output layer, the input layer is used to extract a state vector from the business vector, the hidden layer is used to extract a hidden feature from the state vector, the attention mechanism layer is used to calculate an attention weight and determine a target feature through the hidden feature and the attention weight, the attention weight is determined by a similarity between a business risk indicator and the hidden feature and a position encoding of the hidden feature, and the output layer outputs the risk prediction level based on the target feature; and determining a target risk management strategy through the risk prediction level, wherein the target risk management strategy comprises at least one of the following: a risk warning notification, a suspension of business operation and supplementary audit materials.

[0008] Optionally, the target model is obtained in the following manner: a plurality of historical business data is extracted from a business system, each historical business data is encoded to obtain a plurality of historical business vectors; a historical risk prediction level pre-labeled for each historical business vector is determined, each historical business vector and the historical risk prediction level of the historical business vector are determined as a training sample to obtain a training sample set; the number of nodes of the input layer is determined based on the number of feature dimensions of all training samples in the training sample set, and the number of nodes of the output layer is determined based on the number of preset risk prediction level types; and the parameters of the input layer, the hidden layer, the attention mechanism layer and the output layer are trained through the training sample set to obtain the target model.

[0009] Optionally, the hidden layer comprises an input gate, a forget gate and an output gate, and the hidden feature is extracted in the following manner: a feature of each time step contained in the state vector is determined, the feature is input into the forget gate in a time step order to obtain a feature vector after the forget feature; the feature in the feature vector after the forget feature is input into the input gate in a time step order to obtain a screened candidate feature vector; and the feature in the screened candidate feature vector is input into the output gate in a time step order to obtain the hidden feature.

[0010] Optionally, the training of the parameters of the input layer, the hidden layer, the attention mechanism layer and the output layer based on the set of training samples to obtain the target model comprises: screening associated features related to the business risk from the features of the business data, constructing a state space based on the associated features, constructing an action space based on a preset risk management strategy, wherein the risk management strategy comprises at least one of the following: risk warning notification, suspension of business operation and supplementary audit material; performing multiple iterations of training of the parameters of the input layer, the hidden layer, the attention mechanism layer and the output layer based on the state space, the action space and the set of training samples to obtain a candidate model after each iteration of training, wherein each candidate model corresponds to the parameters of the input layer, the hidden layer, the attention mechanism layer and the output layer after one iteration of training; determining a target loss function, calculating a loss value of each candidate model based on the target loss function, and determining the candidate model corresponding to the minimum loss value as the target model.

[0011] Optionally, the multiple iterations of training of the parameters of the input layer, the hidden layer, the attention mechanism layer and the output layer based on the state space, the action space and the set of training samples comprises: for each iteration of training, inputting the training sample into the candidate model to obtain a predicted risk prediction level; inputting the predicted risk prediction level into a policy network, the policy network determining a risk management strategy to be executed from the action space, inputting the risk management strategy into a value network to obtain a reward value; updating the parameters of the candidate model based on the reward value, the risk management strategy to be executed and a state vector of the state space to obtain the candidate model after the iteration of training, wherein the parameters of the candidate model comprise the parameters of the input layer, the hidden layer, the attention mechanism layer and the output layer.

[0012] Optionally, the updating of the parameters of the candidate model based on the reward value, the risk management strategy to be executed and the state vector of the state space comprises: calculating a time difference error based on the reward value, the risk management strategy to be executed and the state vector of the state space; inputting the time difference error, the predicted risk prediction level and a historical risk prediction level into a target loss function to obtain a loss value; calculating a gradient of the parameters of the candidate model based on the loss value and the target loss function, calculating a product of the gradient and a preset learning rate, and calculating a difference between the parameters of the candidate model of the last iteration and the product to obtain the parameters of the candidate model of the current iteration round.

[0013] Optionally, the attention weight is calculated in the following manner: determining a preset query vector, wherein the preset query vector is used to represent the business risk indicator; for each hidden feature, calculating a feature similarity between the hidden feature and the preset query vector to obtain a similarity weight; encoding the position of each hidden feature in the time step in the hidden feature vector to obtain a position encoding, and determining a position weight based on the position encoding of each hidden feature; for each hidden feature, calculating a sum of the similarity weight and the position weight to obtain the attention weight of the hidden feature.

[0014] To achieve the above object, according to another aspect of the present application, a business risk management apparatus is provided. The apparatus comprises: an acquisition unit configured to acquire business data and encode the business data to obtain a business vector; an input unit configured to input the business vector into a target model to obtain a risk prediction level, wherein the target model comprises an input layer, a hidden layer, an attention mechanism layer and an output layer, the input layer is configured to extract a state vector from the business vector, the hidden layer is configured to extract a hidden feature from the state vector, the attention mechanism layer is configured to calculate an attention weight and determine a target feature by the hidden feature and the attention weight, the attention weight is determined by a similarity between a business risk indicator and the hidden feature and a position encoding of the hidden feature, and the output layer outputs the risk prediction level based on the target feature; and a first determination unit configured to determine a target risk management strategy by the risk prediction level, wherein the target risk management strategy comprises at least one of the following: a risk warning notification, a suspension of business operation and supplementary audit materials.

[0015] To achieve the above object, according to another aspect of the present application, a computer program product is provided, comprising a computer program which, when executed by a processor, implements the steps of the business risk management method described in various embodiments of the present application.

[0016] By the present application, the following steps are adopted: acquiring business data and encoding the business data to obtain a business vector; inputting the business vector into a target model to obtain a risk prediction level, wherein the target model comprises an input layer, a hidden layer, an attention mechanism layer and an output layer, the input layer is configured to extract a state vector from the business vector, the hidden layer is configured to extract a hidden feature from the state vector, the attention mechanism layer is configured to calculate an attention weight and determine a target feature by the hidden feature and the attention weight, the attention weight is determined by a similarity between a business risk indicator and the hidden feature and a position encoding of the hidden feature, and the output layer outputs the risk prediction level based on the target feature; and determining a target risk management strategy by the risk prediction level, wherein the target risk management strategy comprises at least one of the following: a risk warning notification, a suspension of business operation and supplementary audit materials, thereby solving the problem of low business risk management efficiency in the related art. By focusing on the key features in the business data through the attention mechanism layer of the target model, the sensitivity of the target model to the business risk related information is improved, so that the judgment of the business risk is more accurate, the target risk management strategy is determined by the risk prediction level, the target risk management strategy is automatically and reasonably selected according to the risk prediction level, the automation and intelligentization of the business risk management process are realized, and the effect of improving the business risk management efficiency is achieved. BRIEF DESCRIPTION OF DRAWINGS

[0017] The accompanying drawings, which form a part of the present application, are intended to provide further understanding of the present application by presenting the illustrative embodiments of the present application and are not intended to limit the present application. The illustrative embodiments of the present application, as well as the explanations given, are used to explain the present application and should not be construed as a limitation of the present application. In the drawings:

[0018] Figure 1 is a flowchart of a business risk management method according to an embodiment of the present application;

[0019] Figure 2 is a schematic diagram of a business risk management device according to an embodiment of the present application;

[0020] Figure 3 is a schematic diagram of an electronic device according to an embodiment of the present application. DETAILED DESCRIPTION

[0021] It should be noted that the embodiments and features of the embodiments in the present application can be combined with each other without conflict. The present application will be described in detail below with reference to the accompanying drawings and in conjunction with the embodiments.

[0022] In order to enable persons skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the accompanying drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, but not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by persons skilled in the art without creative labor should fall within the scope of protection of the present application.

[0023] It should be noted that the terms "first", "second", and the like in the specification and claims of the present application and the above-mentioned drawings are used to distinguish similar objects, and do not necessarily indicate a specific order or a chronological sequence. It should be understood that the data thus used can be interchanged under appropriate circumstances, so that the embodiments of the present application described herein can be implemented. In addition, the terms "include" and "have" and any variations thereof are intended to cover non-exclusive inclusion, for example, a process, method, system, product or device including a series of steps or units does not necessarily limit to those steps or units clearly listed, but can include other steps or units not clearly listed or inherent to these processes, methods, products or devices.

[0024] The present application will be described below in conjunction with preferred implementation steps, Figure 1 is a flowchart of a business risk management method according to an embodiment of the present application, as shown in Figure 1 the method comprises the following steps:

[0025] Step S101, obtaining business data and performing encoding processing on the business data to obtain a business vector.

[0026] In step S101, the business data can be based on the built Flink (a framework for processing unbounded and bounded data streams) CDC (Change Data Capture, a framework for capturing database change data in real time) plug-in business system (such as financial system, legal system, operation system, etc.) and the historical business archives collect data samples associated with business risk control, including but not limited to business transaction records, internal and external audit reports, violation event details and other information. The business data is encoded to obtain the vector representation of the data sample, that is, the business vector. Encoding the data sample to obtain the vector representation includes numerical data (such as transaction amount, time) standardization processing and feature integration, text data encoding and pooling operation, risk assessment state data embedding and other operations.

[0027] In step S102, the business vector is input into the target model to obtain the risk prediction level, wherein the target model includes an input layer, a hidden layer, an attention mechanism layer and an output layer, the input layer is used to extract a state vector from the business vector, the hidden layer is used to extract hidden features from the state vector, the attention mechanism layer is used to calculate attention weights, and the target features are determined through the hidden features and the attention weights, the attention weights are determined by the similarity between the business risk indicators and the hidden features and the position encoding of the hidden features, and the output layer outputs the risk prediction level based on the target features.

[0028] In step S102, the model architecture of the target model including the input layer, the multi-layer LSTM (Long Short-Term Memory, long short-term memory network) hidden layer, the attention mechanism layer and the output layer is constructed. The input layer obtains the numerical data after standardization processing, the feature vector after text data pooling, and the embedded risk assessment state data from the state space, and inputs them to the input layer to be converted into a state vector; the hidden layer obtains time sequence features by processing the state vector, and the first layer LSTM hidden layer receives the state vector transmitted from the input layer, extracts hidden features from the state vector by using the input gate, the forgetting gate and the output gate contained therein, and forms a hidden state vector.

[0029] After the last layer of LSTM hidden layer, the attention mechanism layer is configured to take the time sequence features as input, and measure the importance of the hidden states corresponding to different time steps in the time sequence features to the business risk judgment by calculating the attention weights. For example, for the hidden states corresponding to a business operation sequence including multiple time steps, the attention mechanism can automatically focus on the hidden states corresponding to some key operation steps that are more strongly associated with business risk to generate target features. The output layer receives the target features (which carry key information extracted from state space data through each layer of processing) from the attention mechanism layer to generate a risk prediction level.

[0030] Step S103, determine the target risk management strategy through the risk prediction level, wherein the target risk management strategy includes at least one of the following: risk warning notification, suspension of business operation and supplementary audit material.

[0031] In step S103, the risk prediction level reflects the likelihood of the business activity being at different risk levels. Based on these prediction levels, corresponding target risk management strategies are automatically selected, including but not limited to risk warning notifications, suspension of business operations, and supplementary audit materials, to ensure compliance and prevent potential business violation risks.

[0032] When the model predicts that a transaction or operation has some risk, but the risk level has not reached the level that requires immediate suspension of business operation, a risk warning notification can be automatically sent. These notifications can be sent to specific compliance team members, business managers or directly involved employees, reminding them to pay attention and further review the business activity. The warning notification can contain the risk characteristics identified by the model, the prediction level and the recommended review action, which helps the relevant parties to take timely measures to prevent the risk from escalating.

[0033] If the predicted risk level is high and reaches the preset threshold, the strategy of suspending business operation is automatically executed. For example, temporarily blocking or delaying the performance of high-risk transactions or operations until additional compliance review or verification is completed. For example, in financial transactions, if a transaction is predicted to involve risk, the transaction is immediately suspended and a detailed investigation process is initiated to confirm its compliance.

[0034] In some cases, the model may predict that a transaction or operation has moderate risk, but not enough to trigger immediate suspension of operation. It requires the business department or relevant personnel to provide more audit materials or conduct additional verification. This can include requiring customers to provide additional documentation, conducting secondary identity verification or providing transaction background information. The strategy of supplementing audit materials aims to enhance the judgment of transaction compliance by collecting more information, reduce the risk of misjudgment or omission, and balance the needs of risk control and business continuity by avoiding unnecessary business interruption.

[0035] The above strategies are automatically selected according to the risk prediction level and preset rules. For example, different thresholds can be set to distinguish low, medium and high risk levels, and each level corresponds to different levels of management strategies. Low risk can only trigger system-level monitoring records, medium risk triggers warning notifications, and high risk immediately suspends business operations.

[0036] The business risk management method provided in the embodiments of the present application obtains business data, encodes the business data to obtain a business vector, inputs the business vector into a target model to obtain a risk prediction level, wherein the target model comprises an input layer, a hidden layer, an attention mechanism layer and an output layer, the input layer is used to extract a state vector from the business vector, the hidden layer is used to extract a hidden feature from the state vector, the attention mechanism layer is used to calculate an attention weight and determine a target feature through the hidden feature and the attention weight, the attention weight is determined by a similarity between a business risk index and the hidden feature and a position encoding of the hidden feature, and the output layer outputs the risk prediction level based on the target feature; and a target risk management strategy is determined through the risk prediction level, wherein the target risk management strategy comprises at least one of the following: a risk warning notification, suspension of business operation and supplementary audit materials, thereby solving the problem of low business risk management efficiency in the related art. Through the attention mechanism layer of the target model, the key features in the business data are focused, the sensitivity of the target model to the business risk related information is improved, the judgment of the business risk is more accurate, the target risk management strategy is determined through the risk prediction level, the target risk management strategy is automatically and reasonably selected according to the risk prediction level, the automation and intelligentization of the business risk management process are realized, and the effect of improving the business risk management efficiency is achieved.

[0037] In order to determine the business risk in time, the target model needs to be trained. Optionally, in the business risk management method provided in the embodiments of the present application, the target model is obtained in the following manner: a plurality of historical business data are extracted from a business system, each historical business data is encoded to obtain a plurality of historical business vectors; a historical risk prediction level pre-labeled for each historical business vector is determined, each historical business vector and the historical risk prediction level of the historical business vector are determined as a training sample, and a training sample set is obtained; the number of nodes of the input layer is determined based on the number of feature dimensions of all training samples in the training sample set, and the number of nodes of the output layer is determined based on the number of preset risk prediction level types; and the parameters of the input layer, the hidden layer, the attention mechanism layer and the output layer are trained through the training sample set, and the target model is obtained.

[0038] In some embodiments, based on the built Flink CDC plug-in business system and the historical compliance archives, data samples associated with business risk control are collected, including but not limited to business transaction records, internal and external audit reports, and details of violation events. Based on the kafka (a distributed stream processing platform) component, the collected historical business data is preprocessed and sent to the real-time data warehouse in the form of associated data stream. In the real-time data warehouse, the historical business data is managed in a columnar storage manner to improve data storage efficiency and query performance. The historical business data is encoded to obtain a sample encoding vector to obtain a historical business vector, and the risk event is labeled to form a training sample set.

[0039] A model architecture is constructed, which includes an input layer (determined by the number of nodes according to the data feature dimension), multiple LSTM hidden layers (unique gate mechanism to process data dependency and time series features, and decreasing number of neurons to refine features), an attention mechanism layer (calculate attention weights based on content and location combination to focus on key features), and an output layer (determine the number of nodes and activation functions according to the risk prediction level, such as Softmax for risk level classification). The number of nodes of the input layer is determined according to the data feature dimension of the vector representation of the data sample in the training sample set. For example, if the data feature dimension of the vector representation of the data sample is 310, the number of nodes of the input layer is set to 310, which is used to receive the vector representation of the business data as the model input.

[0040] Multiple LSTM hidden layers are set, and the number of neurons in each layer can be gradually decreased (such as 128 neurons in the first layer and 64 neurons in the second layer, etc.), to gradually extract abstract features in the vector representation of the training sample. The LSTM hidden layer includes an input gate, a forget gate, and an output gate. The gate control mechanism can effectively process the dependency in the vector representation of the data sample, thereby obtaining time series features (such as the order of operations, the time series of transactions, etc.). After the last LSTM hidden layer, an attention mechanism layer is configured to take the time series features as input, and calculate attention weights to measure the importance of different time steps in the time series features to the business risk judgment. For example, for the hidden states corresponding to a business operation sequence containing multiple time steps, the attention mechanism can automatically focus on the hidden states corresponding to some key operation steps that are more strongly associated with business risk.

[0041] The subsequent hidden layers in the multi-layer LSTM hidden layer successively receive the hidden state vector output by the previous layer and repeatedly process the gating mechanism. Moreover, the number of neurons in each layer is gradually reduced (e.g., from 128 to 64), which further refines and abstracts the features and extracts deep features that are more representative and more relevant to business risks. Each layer processes the hidden state vector output by the previous layer according to the number of neurons and the gating mechanism, so that the hidden state sequence output by the last LSTM hidden layer can be highly focused on the feature information that is important for business risk judgment, thereby providing key input data for the subsequent attention mechanism layer (if any) or directly for the output layer.

[0042] The input vector of the first hidden layer is directly derived from the state vector transmitted by the input layer, which contains various information (e.g., numerical data, text data features, risk assessment state data, etc.) integrated from the original business data. The input vector of the subsequent hidden layer is the hidden state vector output by the previous hidden layer, which has been processed by the gating mechanism of the previous hidden layer and has been preliminarily or further refined. The feature form and information focus of the data are constantly changing, and the business risk-related feature representation is more focused and abstracted based on the previous layer.

[0043] In addition, the number of neurons in each layer is gradually reduced (e.g., from 128 to 64), which affects the processing process. The number of neurons determines the dimensionality of the learnable parameters such as the weight matrix. As the number of neurons decreases, the dimensionality of the input vector when each layer extracts and transforms features is reduced, which is equivalent to further compressing and refining the features. This makes the information carried by the hidden state vector output by each hidden layer more condensed and representative, and constantly converges towards key features that are most closely related to business risk judgment. Through layer-by-layer progression, deep features related to business risks in complex business data are mined, and a highly focused key feature hidden state sequence is finally output for subsequent model structures.

[0044] In calculating the attention weight, the content-based attention (calculating the weight according to the similarity between the hidden state and the query vector) and the position-based attention (allocating the weight considering the position information of the hidden state in the sequence) are combined to calculate the attention weight of different hidden states in the time series features, and then the hidden states with attention weights are fused to obtain target features that are more focused on key features related to business risks in the data.

[0045] The attention mechanism layer takes the hidden state sequence output by the last LSTM hidden layer as input, and this hidden state sequence carries feature information related to business risk after being processed by the hidden layer (these information originally come from various business and risk related data contained in the state space). The importance of different time step corresponding hidden states in the hidden state sequence to the business risk judgment is measured by calculating the attention weight. Based on the content-based attention and location-based attention, the attention weight corresponding to each hidden state is determined. For example, for the hidden state corresponding to a business operation sequence containing multiple time steps, the attention mechanism can automatically focus on the hidden state corresponding to some key operation steps that are more strongly associated with business risk.

[0046] After calculating the attention weight of each hidden state, the hidden state with attention weight is integrated into a vector representation that focuses more on the key features related to business risk in the data through weighted summation. This vector representation highlights the key information more than the hidden state sequence output by the previous hidden layer, and will become an important basis for the subsequent output layer to make business risk prediction and judgment. It is essentially a key feature condensation obtained by deep mining and refining the rich business and risk data contained in the state space.

[0047] The number of output layer nodes and activation function are determined according to the target of business risk prediction (such as risk level classification or binary classification of risk occurrence or not). For example, for three-class risk level prediction, the output layer is set to 3 nodes, and the Softmax activation function is used to convert the output into a probability distribution corresponding to each category, which is convenient for intuitive judgment of the possibility of risk level. The output layer receives the key vector representation from the attention mechanism layer (the vector carries the key information extracted from the state space data through each layer) to generate the risk prediction level.

[0048] Taking the Softmax activation function as an example, the received target feature representation is nonlinearly transformed to convert the output into a probability distribution corresponding to each category, that is, each node outputs a value between 0 and 1, and the sum of all node output values is 1. These probability values respectively represent the possibility of occurrence of the corresponding risk level, thereby generating the final risk prediction level, such as outputting the probability distribution corresponding to different risk levels to intuitively judge the possibility of risk level, complete the forward model inference process based on the state space input data, and obtain the judgment of business risk under the current business state, including the predicted risk category and risk level.

[0049] For example, the output layer receives the vector representation of the target feature transmitted from the attention mechanism layer, and sets the target feature representation as z = [z1, z2, …, zn], where n is the dimension of the vector representation. The output layer uses the Softmax activation function to convert the input vector into a probability distribution corresponding to each category, that is, each node outputs a value between 0 and 1, and the sum of all node output values is 1. These probability values respectively represent the possibility of occurrence of the corresponding risk level, thereby generating the final risk prediction level, such as outputting the probability distribution corresponding to different risk levels to intuitively judge the possibility of risk level, complete the forward model inference process based on the state space input data, and obtain the judgment of business risk under the current business state, including the predicted risk category and risk level. nIn the vector z, n represents the number of neurons in the output layer, which also corresponds to the number of risk prediction level categories. zi (i = 1, 2, ..., n) are elements in the vector z, containing the business risk-related feature information extracted from the previous layers, but not yet presented in probabilistic form. For each node i in the output layer (corresponding to each risk prediction level category), the probability value p corresponding to that category is calculated using the Softmax activation function as follows. i :

[0050]

[0051] For each element z in vector z i Perform exponentiation exp(z) i This transforms the original differences between elements into positive values, and the original differences between the elements are amplified or reduced by the properties of the exponential function, resulting in exp(z1), exp(z2), ..., exp(z... n These values. Next, calculate the sum of all elements after the exponentiation operation, that is... This sum serves as the denominator for subsequent normalization, ensuring that the sum of all class probabilities equals 1. Dividing the result of the exponentiation of each element by this sum yields the corresponding probability value p. i For example, for the risk prediction level corresponding to the (1st)th node, its probability p1 is... And so on, calculate the probability values ​​p1, p2, ..., p for all n nodes. n .

[0052] Due to the properties of the Softmax function, these probability values ​​satisfy the following two conditions: 0 ≤ p i ≤1 indicates that the probability value corresponding to each risk level category is within the reasonable probability range of 0 to 1. =The sum of the probabilities of all risk level categories is 1, forming a complete probability distribution. After calculation by the Softmax activation function, the probability distribution vector output by the output layer is p = [p1, p2, ..., p...]. n This constitutes the final business risk forecast result. Among them, p i This represents the probability value of the occurrence of the i-th risk prediction level category.

[0053] This embodiment integrates multi-source data and uses machine learning algorithms to efficiently process massive and complex data, fully mining the time series features and deep-seated relationships in the data. It overcomes the difficulty of manual review in dealing with massive amounts of data, greatly improves the efficiency and accuracy of risk identification, and avoids the problem of delayed risk assessment.

[0054] The time sequence features in the business data are extracted through the hidden layer. Optionally, in the business risk management method provided in the embodiments of the present application, the hidden layer comprises an input gate, a forget gate and an output gate, and the hidden features are extracted in the following manner: the features of each time step contained in the state vector are determined, the features are input into the forget gate in the order of time steps for processing, and a feature vector after the features are forgotten is obtained; the features in the feature vector after the features are forgotten are input into the input gate in the order of time steps for processing, and a screened candidate feature vector is obtained; and the features in the screened candidate feature vector are input into the output gate in the order of time steps for processing, and the hidden features are obtained.

[0055] In some embodiments, the input gate is used to determine which new information (features) can enter the cell state at the current time, the new information in the state vector transmitted from the input layer is screened, and the amount and content of the information flow are controlled, so as to help the LSTM selectively learn and update the features related to the business risk, avoid irrelevant or interfering information from entering the subsequent processing flow, and thus more accurately extract the key features and form the hidden state vector.

[0056] Firstly, the input state vector and the previous hidden state vector at the current time (when the first time step is processed, the previous hidden state vector is usually initialized as a full 0 vector or a random vector, and in subsequent time steps, the output hidden state vector at the previous time is used) are spliced. Assuming that the dimension of the state vector is n and the dimension of the previous hidden state vector is m, a vector with a dimension of n+m is obtained after splicing. Then, the spliced vector is multiplied by a set of learnable weight matrices W i (including (n+m) x h, where h is the number of neurons in the hidden layer) and a bias term b i (including h) is added, and finally a nonlinear transformation is performed through an activation function (usually a Sigmoid function) to obtain the activation value i t of the input gate, which has a dimension of h and is expressed by the following formula:

[0057] i t = σ(W i ·[x t ,h t-1 ]+b i );

[0058] Wherein, σ represents a Sigmoid (a kind of activation function), x t represents the state vector input at the current time, and h t-1 represents the hidden state vector at the previous time. The same spliced vector (the vector obtained by splicing the state vector and the previous hidden state vector) is multiplied by another set of learnable weight matrices W c (including (n+m) x h), and a bias term b c(h), to obtain a candidate cell state The formula is:

[0059]

[0060] The activation value t of the input gate is calculated as: i and the candidate cell state are multiplied element by element to obtain a new information part to be updated into the cell state. Then, this part of new information is added element by element with the cell state C t-1 of the previous time (the initial time can be set to all 0 or set according to a specific initialization strategy, and the subsequent time is updated from the previous time) to obtain the updated cell state C t of the current time. The formula is:

[0061]

[0062] where represents the element-by-element multiplication operation. Through the processing of the input gate, only those new information that have passed the activation value screening and are considered important are integrated into the cell state, preparing for the subsequent formation of the hidden state vector.

[0063] The function of the forget gate is to control how much past information is forgotten from the cell state. Since the data related to business risk often has time sequence characteristics, the business situation changes over time, and the information at some time before may no longer be important or outdated at the current time. The forget gate is responsible for dynamically deciding which information can be discarded, so that the cell state can focus on the current more critical information, and thus help to accurately extract the features in the hidden state vector that are most relevant to the current business risk judgment.

[0064] First, the input state vector and the previous hidden state vector at the current time are spliced to obtain a vector with dimension n+m. Then, the spliced vector is multiplied by a set of learnable weight matrix W f with dimension (n+m) x h, and then added with bias term b f with dimension h, and finally nonlinearly transformed by the Sigmoid function to obtain the activation value f t of the forget gate, which has dimension h, and the formula is as follows:

[0065] f t = σ(W f · [x t , h t-1 ] + b f );

[0066] Then, the activation value f t of the forget gate is multiplied element by element with the cell state C t-1The element-wise multiplication is performed to obtain the last time cell state information filtered by the forgetting gate, that is, to determine which past information is to be retained to participate in the construction of the current time cell state, and the formula is:

[0067] C' t-1 = f t ⊙C t-1 ;

[0068] This C' t-1 will be used together with the updated cell state C t after the previous input gate to generate the subsequent hidden state vector, and part of the historical information that is no longer important is reasonably removed through the forgetting gate, so that the cell state can reflect the latest key features of business and risk.

[0069] The output gate is responsible for determining how much information in the current time cell state can be output as a hidden state vector, which is based on the current cell state and the last hidden state and other information to make a comprehensive judgment. The information that is valuable for current business risk judgment and suitable for external transmission after the processing of the previous input gate and forgetting gate is extracted to form the final hidden state vector. This hidden state vector will be passed to the next layer (if there are multiple hidden layers) or the subsequent attention mechanism layer (if it exists) for further processing to further mine business risk related features.

[0070] First, the input state vector and the last hidden state vector at the current time are spliced to obtain a vector with dimensions n+m. Then, the spliced vector is multiplied by a set of learnable weight matrix W o with dimensions (n+m) x h, and then added to the bias term b o with dimensions h, and then nonlinearly transformed by the Sigmoid function to obtain the activation value o t of the output gate with dimensions h, and the formula is as follows:

[0071] o t = σ(W o ·[x t , h t-1 ]+b o );

[0072] Next, the updated cell state C t at the current time is processed by the hyperbolic tangent function to obtain a compressed and transformed cell state representation tanh(C t ) with a value range of [-1, 1]. Finally, the activation value o t of the output gate is element-wise multiplied by tanh(C t ) to obtain the hidden state vector h t at the current time, and the formula is:

[0073] h t = o t ⊙ tanh(C t );

[0074] The hidden state vector contains the abstract features related to business risks extracted from the state vector and screened and processed by the gating mechanism of the LSTM unit, which is passed down as the output of the first-layer LSTM hidden layer for further analysis and utilization by subsequent model structure layers to gradually complete the mining of the entire business risk-related features and the judgment of the model on the risks.

[0075] In this embodiment, the gating mechanism enables the hidden layer to extract time series features highly related to business risk assessment from the state vector, thereby improving the accuracy and real-time performance of risk prediction.

[0076] The target model is trained based on the reinforcement learning method. Optionally, in the business risk management method provided in the present application, the parameters of the input layer, the hidden layer, the attention mechanism layer and the output layer are trained by using a training sample set to obtain the target model, including: screening the associated features related to business risks from the features of the business data, constructing a state space based on the associated features, and constructing an action space based on a preset risk management strategy, wherein the risk management strategy includes at least one of the following: risk warning notification, suspension of business operation and supplementary audit material; based on the state space, the action space and the training sample set, the parameters of the input layer, the hidden layer, the attention mechanism layer and the output layer are iteratively trained multiple times to obtain a candidate model after each iteration, wherein each candidate model corresponds to the parameters of the input layer, the hidden layer, the attention mechanism layer and the output layer after one iteration; a target loss function is determined, the loss value of each candidate model is calculated based on the target loss function, and the candidate model corresponding to the minimum loss value is determined as the target model.

[0077] In some embodiments, the state space is constructed based on the vector representation of the training samples: for numerical data in the business risk control related training samples, such as transaction amount, transaction time (which can be converted into a numerical value after timestamp), etc., these data directly correspond to the business data feature part in the state space. For example, the transaction amount can be directly placed as a numerical value of one dimension in the state space. If the numerical range of the transaction amount is too large, it can be standardized to make it in the same order of magnitude range as other numerical data, so as to avoid the influence of a certain dimension of data on the state space. Taking the transaction time as an example, assuming that the transaction time is in the form of a timestamp in the vector representation of the data sample, when constructing the state space, it can be converted into a relative time (such as the time interval from a certain reference time point) or a time interval index (such as dividing a day into multiple time intervals, determining the index of the transaction time interval), etc. as a dimension placed in the state space, for reflecting the time characteristics of the business operation.

[0078] For text data (such as contract text, text description in operation log, etc.), it has been converted into word vector representation in the data preprocessing stage. When constructing the state space, these word vectors can be processed. For example, through average pooling or maximum pooling operation, the word vector sequence corresponding to the text is converted into a fixed dimension vector, and then each dimension of this vector is taken as a part of the state space. Among them, if the text after word vector encoding is a vector sequence with length n and dimension m, through average pooling, a vector with dimension m can be obtained, and each dimension of this vector can be taken as a part of the state space, for reflecting the features related to the text content.

[0079] For the existing risk assessment state data (risk level, risk type), these data can be directly embedded into the state space according to certain encoding rules. For example, the risk level can be placed as a value of one dimension in the state space with the number 1-3 (representing low, medium and high risk respectively); the risk type can be represented by the number 1-k (k is the number of risk types) to represent different risk types, which is also placed as a dimension in the state space.

[0080] According to the business needs and the focus of model training, each of the above text features and numerical features in the training sample is selected and combined. Not all data features contribute equally to the state space. The principal component analysis method is used to select the features most relevant to business risk, and then these features are combined in the appropriate order and manner to form a state space. For example, through principal component analysis, it is found that transaction amount, credit rating of transaction participants, and certain key terms in the contract (features encoded by text) are highly related to business risk. Therefore, when constructing the state space, these features are combined to form a vector space that can effectively reflect the state of business and risk, which is used as the state space.

[0081] After determining the state space, the corresponding action space is set, that is, the risk management strategy options that the target model can take in different states. These options are all around business risk management. In the training process, the target model can select the corresponding action from the pre-set action space according to the current state (information obtained from the state space). For example, the action space can include discrete action options corresponding to control strategies such as sending risk warning notifications, suspending related business operations, and requiring additional audit materials. These different action options have their corresponding numbers or codes, which facilitate the model to select and record during training, and each action will have different effects on business risk management. The model needs to learn continuously to determine which action is most appropriate in which state.

[0082] Randomly initialize the neural network parameters of the input layer, hidden layer (LSTM), attention mechanism layer, and output layer, including weight matrices and bias vectors. Use the training sample set for multiple iterations. In each iteration: input a state vector from the state space into the model, extract time series features through LSTM, focus on key attributes through the attention mechanism layer, and generate a risk prediction level through the output layer. Then, select an action from the action space based on the current state and prediction level, and calculate the reward value. Finally, adjust the model parameters based on the reward value to optimize the strategy for future decision-making.

[0083] This embodiment trains a target model that can extract key features from the state space and intelligently select risk management strategies based on multi-source data fusion and reinforcement learning. It improves the accuracy of risk prediction and automatically adjusts risk management strategies according to real-time changes in business data, effectively preventing and controlling potential business risks, and providing strong support for the stable operation of enterprises.

[0084] The model parameters can be updated based on a reinforcement learning method. Optionally, in the business risk management method provided in the embodiments of the present application, the multiple iterative training of the parameters of the input layer, the hidden layer, the attention mechanism layer and the output layer based on the state space, the action space and the training sample set comprises: for each iteration training, inputting the training sample into the candidate model to obtain a predicted risk prediction level; inputting the predicted risk prediction level into the policy network, and determining a to-be-executed risk management strategy from the action space by the policy network; inputting the risk management strategy into the value network to obtain a reward value; and updating the parameters of the candidate model based on the reward value, the to-be-executed risk management strategy and a state vector of the state space to obtain the candidate model after the iteration training, wherein the parameters of the candidate model comprise the parameters of the input layer, the hidden layer, the attention mechanism layer and the output layer.

[0085] In some embodiments, the target model selects an action to be executed from the preset action space according to the risk prediction level. For example, if the model determines that the business risk degree is high in the current state (obtained from the comprehensive analysis of the risk level and the related business data features contained in the state vector), the model can select the action of “suspending the related business operation” according to the learned strategy, and the corresponding code (assuming 2) will be recorded, indicating the action option selected by the target model this time.

[0086] After selecting and recording the action, the reward value is determined according to the preset reward mechanism and the actual risk feedback obtained by simulation in the training data set. If the risk is effectively controlled (such as the risk level is reduced, the violation events are reduced, etc. which meet the positive reward setting) after the action is executed, a positive reward is given to the model according to the reward mechanism; otherwise, if the risk is deteriorated (such as the risk level is increased, new violation events occur, etc.), a negative reward is given. Then, the related parameters of the model are updated based on the obtained reward value and the current state, the executed action and other information, including the weight parameters of the neural network part (such as the weights of each layer of the LSTM and the output layer are updated by the back propagation method, so that the neural network part can learn better feature extraction and risk judgment ability), so as to continuously optimize the model's ability to control the business risk and adapt to different business and risk state changes. Through multiple rounds of such iterative training process, the training of the model is finally completed, so that the model can accurately identify the business risk and make appropriate control decisions when facing actual business scenarios.

[0087] The embodiments of the present application can automatically and reasonably select the control strategy according to the risk prediction level by constructing an intelligent action space and model decision linkage mechanism, realize the automation and intelligentization of the risk control process, reduce the uncertainty and low efficiency problems caused by manual intervention, and improve the operation efficiency of the entire business risk management system and the ability to cope with complex situations.

[0088] Optionally, in the business risk management method provided in the embodiments of the present application, updating the parameters of the candidate model based on the reward value, the risk management strategy to be executed, and the state vector of the state space comprises: calculating a time difference error based on the reward value, the risk management strategy to be executed, and the state vector of the state space; inputting the time difference error, the predicted risk prediction level, and the historical risk prediction level into a target loss function to obtain a loss value; calculating the gradient of the parameters of the candidate model based on the loss value and the target loss function, calculating the product of the gradient and a preset learning rate, and calculating the difference between the parameters of the candidate model in the last iteration and the product to obtain the parameters of the candidate model in the current iteration round.

[0089] In some embodiments, a policy network (denoted as π(a|s)) is used to generate a probability distribution of selecting different actions a according to the current state s, where a is from a preset action space A, and the state s is the current business and risk state represented by the state vector constructed before. The action space A = {a1, a2, …, a m For a given state s, the policy network outputs the probability corresponding to each action, that is, π(a1|s), π(a2|s), …, π(a m |s), and satisfies After the target model comprehensively analyzes the current state s according to the business risk prediction level and the business data features in the state space, an action is selected based on the output probability of the policy network using the Softmax sampling algorithm. According to the preset reward mechanism, the reward value r is determined based on the actual risk feedback obtained by simulation in the training data set. Let the reward function be R(s, a), which takes the current state s and the executed action a as input, and outputs the corresponding reward value r.

[0090] If the action a is executed, it is found through simulation that the risk has been effectively controlled (such as the risk level is reduced, the violation events are reduced, etc. which meets the positive reward setting), then the reward value r is positive, that is, r = R(s, a) > 0, and the specific function form can be defined according to the actual business, such as constructing a function positively related to the risk control effect according to the risk level change amount, the number of reduced violation events, etc. to quantify the degree of positive reward. If the risk is deteriorating (the risk level is rising, new violation events are occurring, etc.), the reward value r is negative, that is, r = R(s, a) < 0, and the specific function form can also be determined according to the degree of risk deterioration (such as the number of risk level rises, etc.) to determine the specific value of negative reward, so as to reflect the adverse effect of action selection on risk control.

[0091] The parameter update of the neural network part of the model (including each layer of LSTM and the output layer, etc.) aims to minimize a loss function L(θ) related to reinforcement learning, where θ represents the set of all parameters of the neural network (such as weight matrices and bias vectors of each layer, etc.). Its parameter update formula is:

[0092]

[0093] where θ t is the parameter value at the t-th iteration, θ t+1 is the updated parameter value, α is the learning rate (controls the step size of parameter update), is the gradient of the loss function L(θ) with respect to the parameter θ at θ t . In order to calculate the gradient , the backpropagation algorithm needs to be used: the cross-entropy loss function combined with the classification of the Softmax output (used for business risk prediction results), assuming that the predicted probability distribution of the model output is p = [p1, p2, …, p n ](corresponding to n risk level categories), and the true risk level label (existing in the training data set of supervised learning) is y = [y1, y2, …, y n ](where only one element is 1, representing the true category, and the rest are 0), then the cross-entropy loss function can be expressed as:

[0094]

[0095] Through the backpropagation algorithm, starting from the output layer, the gradients of the loss function with respect to the parameters of each layer are calculated in turn according to the chain rule, and these gradient information is propagated back to each layer of the neural network. Then, according to the above parameter update formula, the weight parameters of the neural network are updated, so that the model can learn better feature extraction and risk judgment ability, and the predicted risk level probability distribution is closer to the true situation, and can make more appropriate action choices according to different states.

[0096] Since the deep learning network is a reinforcement learning method that combines deep learning with Q-learning algorithm, it is designed to handle complex decision-making problems with high-dimensional state space, and is very suitable for intelligent compliance risk control scenarios based on multi-source data fusion and reinforcement learning. It mainly includes two core networks, namely the value network and the target network, whose parameters have an update lag relationship with the value network parameters. The value network is used to estimate the value of performing different actions in a given state, and the target network is used to provide relatively stable target Q values to help stabilize the training process and avoid over-optimistic estimation, etc.

[0097] In DQN (Deep Q-Network), the time-difference error plays a key role in guiding the parameter update, which reflects the difference between the current estimate of state-action value and the estimate based on the subsequent reward and the next state value. Its calculation formula is:

[0098]

[0099] Where r represents the immediate reward value obtained from the environment after executing action a, which is determined according to the preset reward mechanism and the actual risk feedback simulation (as described earlier, if the risk is effectively controlled after executing the action, r is positive, and if the risk is deteriorating, r is negative). γ is the discount factor (usually in the range of 0 to 1), which measures the importance of future rewards in current decision-making, reflecting a discount consideration of long-term earnings, for example, γ = 0.9 means that the reward at the next moment is only equivalent to 90% of its actual value, and the influence weight of future rewards will gradually decrease according to this discount rate as the time step goes back. s' represents the next state to which the system moves after executing action a (this next state can be determined in the training data set according to business logic and simulated business development, for example, in the business risk control scenario, if the "send risk warning notification" action is executed, the next state may be that the relevant personnel have received the notification and conducted a certain review of the business, and the corresponding business data and risk assessment state have changed after the state), represents the maximum value estimate calculated for all possible actions a' in the target network for the next state s', which reflects the expected value of future optimal decision based on the target network, while Q(s, a; θ) is the value network's estimate of the value of executing action a in the current state s.

[0100] Based on gradient descent, the value network parameters are updated based on the calculated TD error, using the gradient descent algorithm to update the value network parameters θ, with the goal of minimizing the TD error, so that the value network's estimate of state-action value can more accurately approach the real optimal value. Usually, the loss function in the form of mean square error (MSE) is used to measure the TD error, i.e. (here δ is the TD error calculated above), and then the gradient of the loss function with respect to the parameter θ is calculated to update the parameters. According to the chain rule, the gradient of with respect to θ is:

[0101]

[0102] Further, the update formula of the value network parameters is:

[0103]

[0104] Where, θ t θ is the parameter value of the value network at the t-th iteration. t+1 This is the updated parameter value, α. Q This refers to the learning rate for updating the parameters of the value network (it controls the step size of each parameter update; a value that is too small may lead to slow training, while a value that is too large may make training difficult to converge or even diverge; it needs to be adjusted reasonably according to the specific business scenario and model training situation. For example, different learning rate values ​​can be tried in the early stages of training to observe the model training effect and select an appropriate value). Through this update method, the temporal difference error is backpropagated to each layer of the value network to adjust the weight parameters of each layer, so that the value network can gradually learn a more accurate state-action value estimate, thereby providing more reasonable action selection suggestions when facing different states in the future.

[0105] Each time a batch of sample data (containing multiple states, actions, rewards, and the corresponding next state) is obtained from the training dataset, the temporal difference error is calculated for each batch of samples, and the gradient is accumulated according to the update formula described above. Then, the parameters of the value network are updated uniformly. This approach leverages the statistical characteristics of batch data to update parameters more stably and efficiently, reducing the randomness and noise impact during the parameter update process. Furthermore, to avoid problems such as gradient vanishing or gradient exploding (e.g., when the network has many layers, the gradient may approach zero or become excessively large during backpropagation), techniques such as gradient pruning and normalization can be employed to ensure the smoothness and effectiveness of parameter updates, guaranteeing that the value network continuously converges towards optimizing value estimation.

[0106] When updating the target network parameters, the target network parameter θ - Instead of updating at every step, a relatively lagging update strategy is employed. The aim is to provide the value network with a relatively stable target Q-value reference, preventing training instability caused by the value network chasing the constantly changing target too quickly. A common update method is to copy the parameters of the value network to the target network every certain number of training steps (let's say C steps, e.g., C=100, the specific value can be adjusted according to the actual training situation and model convergence speed). The target network is updated every C steps, and during this period, it keeps the parameters unchanged to provide relatively fixed target values for the value network to calculate the time difference error and guide the parameter update, so that the value network can gradually adjust its parameters based on relatively stable targets, more stably learn the true value relationship between states and actions, and optimize the decision-making strategy of the entire model in the business risk management scenario, so that it can better adapt to different business states and select appropriate actions to deal with risks.

[0107] In this embodiment, the target network parameters are updated periodically, which alleviates the oscillation and instability in the training process caused by frequent updates of the target, so that the model can more effectively learn the optimal action selection strategy in different business and risk states, improve the model's ability to manage business risks and adapt to dynamic changes in business and risks, and ultimately achieve continuous optimization of reinforcement learning-related parameters to enable the entire target model to accurately identify business risks and make appropriate management decisions when facing actual business scenarios. Through the updating process of the reinforcement learning-related parameters in the deep Q network, the target model can continuously use feedback information (rewards, state transitions, etc.) in the training process to optimize its ability to judge the value of actions in different states, thereby improving the overall risk management level and better meeting the intelligent compliance risk management needs based on multi-source data fusion and reinforcement learning.

[0108] In order to screen out hidden features related to business risks, the attention weight is calculated by similarity weight and position weight. Optionally, in the business risk management method provided in the embodiments of the present application, the attention weight is calculated in the following manner: a preset query vector is determined, wherein the preset query vector is used to represent business risk indicators; for each hidden feature, the feature similarity between the hidden feature and the preset query vector is calculated to obtain a similarity weight; the position of each hidden feature in the time step of the hidden feature vector is encoded to obtain a position encoding, and the position weight of each hidden feature is determined based on the position encoding; for each hidden feature, the sum of the similarity weight and the position weight is calculated to obtain the attention weight of the hidden feature.

[0109] In some embodiments, a query vector q is first defined, which is used to measure the similarity of hidden states and business risk-related content. It can be a fixed vector obtained by encoding prior knowledge related to business risks, or a vector dynamically generated according to the current task target. For example, it can be a vector obtained by encoding business risk key indicators (such as high-risk transaction amount range, high-risk business operation type, etc.). For each hidden state h t , the dot product of h and q is calculated as a similarity score i.e. similarity weight, obtaining a similarity score sequence with length T These scores represent how much each hidden state is associated with the content related to business risk.

[0110] Assign a position encoding p to each time step in the state vector t The position encoding can be a sinusoidal-cosine position encoding. The position encoding can be represented as p t，2i = sin(t / 10000 2i / d ) and p t，2i+1 = cos(t / 10000 2i / d ), where d is the dimension of the hidden state and i is the dimension index. The position-based attention weight can be obtained by operating the position encoding with the hidden state, for example, by concatenating the position encoding with the hidden state and then transforming it through a fully connected layer to obtain the position weight, or directly linearly transforming the position encoding to obtain the position weight. Wherein, the position weight b p is calculated through a linear transformation (weight matrix) W p and a bias term, and the calculation formula is as follows:

[0111]

[0112] Fusing the similarity weight and the position weight, the final attention weight can be obtained through the fusion mode of addition:

[0113]

[0114] In order to make the attention weight constitute a reasonable probability distribution, the fused attention weight s t is normalized. The Softmax function is used for normalization to obtain the final attention weight α t represents the relative importance of the hidden state at time step t in the business risk judgment, and

[0115] The embodiment calculates the attention weight based on the similarity weight and the position weight, which can make the target model automatically focus on the hidden state corresponding to some key operation steps with stronger association with business risk, and improve the extraction ability of the model for business risk related features. With the focus of the attention mechanism layer on the key features, the sensitivity of the target model to the business risk related information is further improved, so that the risk judgment is more accurate.

[0116] It should be noted that the steps shown in the flowchart of the accompanying drawings can be executed in a computer system such as a set of computer executable instructions, and although a logical order is shown in the flowchart, in some cases, the steps shown or described can be executed in an order different from that here.

[0117] The embodiment of the present application also provides a business risk management device. It should be noted that the business risk management device of the embodiment of the present application can be used to execute the business risk management method provided by the embodiment of the present application. The business risk management device provided by the embodiment of the present application is introduced as follows.

[0118] Figure 2 is a schematic diagram of the business risk management device provided by the embodiment of the present application. As shown in the figure, Figure 2 The device comprises:

[0119] The acquisition unit 201 is configured to acquire business data and perform encoding processing on the business data to obtain a business vector.

[0120] The input unit 202 is configured to input the business vector into a target model to obtain a risk prediction level, wherein the target model comprises an input layer, a hidden layer, an attention mechanism layer and an output layer, the input layer is configured to extract a state vector from the business vector, the hidden layer is configured to extract a hidden feature from the state vector, the attention mechanism layer is configured to calculate an attention weight and determine a target feature through the hidden feature and the attention weight, the attention weight is determined by a similarity between a business risk index and the hidden feature and a position encoding of the hidden feature, and the output layer outputs the risk prediction level based on the target feature.

[0121] The first determination unit 203 is configured to determine a target risk management strategy through the risk prediction level, wherein the target risk management strategy comprises at least one of the following: a risk warning notification, a suspension of business operation and supplementary audit materials.

[0122] The business risk management apparatus provided in the embodiments of the present application acquires business data through the acquisition unit 201, and performs encoding processing on the business data to obtain a business vector; the input unit 202 inputs the business vector into a target model to obtain a risk prediction level, wherein the target model comprises an input layer, a hidden layer, an attention mechanism layer and an output layer, the input layer is used to extract a state vector from the business vector, the hidden layer is used to extract a hidden feature from the state vector, the attention mechanism layer is used to calculate an attention weight, and a target feature is determined through the hidden feature and the attention weight, the attention weight is determined by the similarity between the business risk indicator and the hidden feature and the position encoding of the hidden feature, and the output layer outputs the risk prediction level based on the target feature; the first determination unit 203 determines a target risk management strategy through the risk prediction level, wherein the target risk management strategy comprises at least one of the following: risk warning notification, suspension of business operation and supplementary audit material, which solves the problem of low business risk management efficiency in the related art, focuses on the key features in the business data through the attention mechanism layer of the target model, improves the sensitivity of the target model to the business risk related information, makes the judgment of the business risk more accurate, determines the target risk management strategy through the risk prediction level, automatically and reasonably selects the target risk management strategy according to the risk prediction level, realizes the automation and intelligentization of the business risk management and control process, and further achieves the effect of improving the business risk management efficiency.

[0123] Optionally, in the business risk management apparatus provided in the embodiments of the present application, the apparatus further comprises: an extraction unit, configured to extract a plurality of historical business data from a business system, encode each historical business data to obtain a plurality of historical business vectors; a second determination unit, configured to determine a historical risk prediction level pre-labeled for each historical business vector, determine each historical business vector and the historical risk prediction level of the historical business vector as a training sample to obtain a training sample set; a third determination unit, configured to determine the number of nodes of the input layer based on the number of feature dimensions of all training samples in the training sample set, and determine the number of nodes of the output layer based on the number of preset risk prediction level types; and a training unit, configured to train the parameters of the input layer, the hidden layer, the attention mechanism layer and the output layer through the training sample set to obtain the target model.

[0124] Optionally, in the business risk management apparatus provided in the embodiments of the present application, the apparatus further comprises: a fourth determination unit, configured to determine the features of each time step contained in the state vector, input the features into a forget gate in a time step order to obtain a feature vector after the features are forgotten; a first screening unit, configured to input the features in the feature vector after the features are forgotten into an input gate in a time step order to obtain a screened candidate feature vector; and a second screening unit, configured to input the features in the screened candidate feature vector into an output gate in a time step order to obtain the hidden feature.

[0125] Optionally, in the business risk management apparatus provided by the embodiments of the present application, the training unit comprises: a screening module, configured to screen out the associated features related to the business risk from the features of the business data, to construct a state space based on the associated features, and to construct an action space based on a preset risk management strategy, wherein the risk management strategy comprises at least one of the following: risk warning notification, suspension of business operation, and supplementary audit material; a training module, configured to perform multiple iterative training on the parameters of the input layer, the hidden layer, the attention mechanism layer, and the output layer based on the state space, the action space, and a training sample set, to obtain a candidate model after each iteration; a first determination module, configured to determine a target loss function, to calculate the loss value of each candidate model based on the target loss function, and to determine the candidate model corresponding to the minimum loss value as a target model.

[0126] Optionally, in the business risk management apparatus provided by the embodiments of the present application, the training module comprises: a first input submodule, configured to input the training sample into the candidate model to obtain a predicted risk prediction level for each iteration; a second input submodule, configured to input the predicted risk prediction level into a policy network, to determine the risk management strategy to be executed from the action space by the policy network, and to input the risk management strategy into a value network to obtain a reward value; and an update submodule, configured to update the parameters of the candidate model based on the reward value, the risk management strategy to be executed, and the state vector of the state space, to obtain the candidate model after iteration, wherein the parameters of the candidate model comprise the parameters of the input layer, the hidden layer, the attention mechanism layer, and the output layer.

[0127] Optionally, in the business risk management apparatus provided by the embodiments of the present application, the update submodule comprises: a first calculation module, configured to calculate a time difference error based on the reward value, the risk management strategy to be executed, and the state vector of the state space; a loss value determination module, configured to input the time difference error, the predicted risk prediction level, and a historical risk prediction level into a target loss function to obtain a loss value; a second calculation module, configured to calculate the gradient of the parameters of the candidate model based on the loss value and the target loss function, to calculate the product of the gradient and a preset learning rate, and to calculate the difference between the parameters of the candidate model of the last iteration and the product to obtain the parameters of the candidate model of the current iteration.

[0128] Optionally, in the business risk management apparatus provided by the embodiment of the present application, the apparatus further comprises: a fifth determination unit configured to determine a preset query vector, wherein the preset query vector is used to represent the business risk index; a first calculation unit configured to calculate, for each hidden feature, a feature similarity between the hidden feature and the preset query vector to obtain a similarity weight; an encoding unit configured to obtain a position encoding based on a position of each hidden feature in a time step in the hidden feature vector, and determine a position weight based on the position encoding of each hidden feature; and a second calculation unit configured to calculate, for each hidden feature, a sum of the similarity weight and the position weight to obtain an attention weight of the hidden feature.

[0129] The business risk management apparatus comprises a processor and a memory, the above-mentioned acquisition unit 201, input unit 202 and first determination unit 203 and the like are stored in the memory as program units, and the corresponding functions are realized by the processor executing the above-mentioned program units stored in the memory.

[0130] The processor comprises a core, and the core retrieves the corresponding program units from the memory. The core can be one or more, and the business risk management efficiency can be improved by adjusting the core parameters.

[0131] The memory can comprise a non-permanent memory in a computer readable medium, random access memory (RAM) and / or non-volatile memory such as read-only memory (ROM) or flash memory (flash RAM), and the memory comprises at least one memory chip.

[0132] The embodiment of the present application provides a computer readable storage medium, which stores a program, and the program is executed by a processor to realize the business risk management method.

[0133] The embodiment of the present application provides a processor, which is used to run a program, and the program is executed to realize the business risk management method.

[0134] Figure 3 FIG. 1 is a schematic diagram of an electronic device according to the embodiment of the present application. As shown in FIG. 1, the electronic device comprises a processor 100 and a memory 200. Figure 3As shown, the electronic device 301 comprises a processor, a memory, and a program stored on the memory and executable on the processor, and the processor implements the following steps when executing the program: obtaining business data, and performing encoding processing on the business data to obtain a business vector; inputting the business vector into a target model to obtain a risk prediction level, wherein the target model comprises an input layer, a hidden layer, an attention mechanism layer, and an output layer, the input layer is used to extract a state vector from the business vector, the hidden layer is used to extract a hidden feature from the state vector, the attention mechanism layer is used to calculate an attention weight, and determine a target feature through the hidden feature and the attention weight, the attention weight is determined by a similarity between a business risk index and the hidden feature and a position encoding of the hidden feature, and the output layer outputs the risk prediction level based on the target feature; and determining a target risk management strategy through the risk prediction level, wherein the target risk management strategy comprises at least one of the following: a risk warning notification, a suspension of business operation, and supplementary audit materials. The device in the present application can be a server, a PC, a PAD, a mobile phone, etc.

[0135] The present application also provides a computer program product adapted to execute a program that initializes the following method steps when executed on a data processing device: obtaining business data, and performing encoding processing on the business data to obtain a business vector; inputting the business vector into a target model to obtain a risk prediction level, wherein the target model comprises an input layer, a hidden layer, an attention mechanism layer, and an output layer, the input layer is used to extract a state vector from the business vector, the hidden layer is used to extract a hidden feature from the state vector, the attention mechanism layer is used to calculate an attention weight, and determine a target feature through the hidden feature and the attention weight, the attention weight is determined by a similarity between a business risk index and the hidden feature and a position encoding of the hidden feature, and the output layer outputs the risk prediction level based on the target feature; and determining a target risk management strategy through the risk prediction level, wherein the target risk management strategy comprises at least one of the following: a risk warning notification, a suspension of business operation, and supplementary audit materials.

[0136] Those skilled in the art should understand that the embodiments of the present application can be provided as a method, a system, or a computer program product. Therefore, the present application can take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware aspects. Moreover, the present application can take the form of a computer program product implemented on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROMs, optical storage, etc.) containing computer-usable program code.

[0137] The computer program instructions can also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer-implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the functions specified in the flowchart block or blocks. Figure 1 one or more flow or blocks Figure 1 means for functionally implementing the steps listed in the flowchart block or blocks.

[0138] These computer program instructions can also be stored in a computer readable memory that can direct a computer or other programmable data processing apparatus to function in a particular manner, such that the instructions stored in the computer readable memory produce an article of manufacture including instructions which implement the function specified in the flowchart block or blocks. Figure 1 one or more flow or blocks Figure 1 means for functionally implementing the steps listed in the flowchart block or blocks.

[0139] The computer program instructions can also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer-implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the functions specified in the flowchart block or blocks. Figure 1 one or more flow or blocks Figure 1 means for functionally implementing the steps listed in the flowchart block or blocks.

[0140] In one typical configuration, the computing device includes one or more processors (CPU's), input / output interfaces, network interfaces, and memory.

[0141] The memory can include non-persistent memory and / or persistent memory, such as flash memory, read-only memory (ROM), and / or volatile or non-volatile random access memory (RAM), among others. The memory is an example of computer readable media.

[0142] Computer-readable media includes permanent and non-permanent, movable and non-movable media that can be implemented by any method or technology to store information. The information can be computer-readable instructions, data structures, program modules or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, compact disc read-only memory (CD-ROM), digital versatile disc (DVD) or other optical storage, magnetic cassette, magnetic tape disk storage or other magnetic storage devices, or any other non-transmission medium that can be used to store information accessible to a computing device. According to the definition herein, computer-readable media does not include transitory media such as modulated data signals and carriers.

[0143] It should also be noted that the terms "comprising", "containing", or any other variant thereof are intended to cover non-exclusive inclusions, such that a process, method, article or apparatus that comprises a list of elements does not only include those elements, but also other elements not explicitly listed or inherent to such process, method, article or apparatus. Without more limitations, the element defined by the phrase "comprising a" does not exclude the presence of additional identical elements in the process, method, article or apparatus that includes the element.

[0144] Those skilled in the art will appreciate that embodiments of the present application can be provided as a method, system or computer program product. Accordingly, the present application can take the form of an entirely hardware embodiment, an entirely software embodiment or an embodiment combining software and hardware aspects. Furthermore, the present application can take the form of a computer program product implemented on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROMs, optical storage devices, etc.) containing computer usable program code.

[0145] The above merely provides embodiments of the present application and is not intended to limit the present application. Various modifications and changes can be made to the present application by those skilled in the art. Any modification, equivalent replacement, improvement, etc. made within the spirit and principles of the present application shall be included in the scope of the claims of the present application.

Claims

1. A business risk management method characterized by, The method comprises: obtaining business data and encoding the business data to obtain a business vector; inputting the business vector into a target model to obtain a risk prediction level, wherein the target model comprises an input layer, a hidden layer, an attention mechanism layer and an output layer, the input layer is used to extract a state vector from the business vector, the hidden layer is used to extract a hidden feature from the state vector, the attention mechanism layer is used to calculate an attention weight and determine a target feature through the hidden feature and the attention weight, the attention weight is determined by the similarity between a business risk indicator and the hidden feature and the position encoding of the hidden feature, and the output layer outputs the risk prediction level based on the target feature; determining a target risk management strategy through the risk prediction level, wherein the target risk management strategy comprises at least one of the following: risk warning notification, suspension of business operation and supplementary audit material.

2. The method of claim 1, wherein, The target model is obtained in the following manner: extracting a plurality of historical business data from a business system, encoding each historical business data to obtain a plurality of historical business vectors; determining a pre-labeled historical risk prediction level of each historical business vector, determining each historical business vector and the historical risk prediction level of the historical business vector as a training sample to obtain a training sample set; determining the number of nodes of the input layer based on the number of feature dimensions of all training samples in the training sample set, and determining the number of nodes of the output layer based on the number of preset risk prediction level types; training the parameters of the input layer, the hidden layer, the attention mechanism layer and the output layer through the training sample set to obtain the target model.

3. The method of claim 1, wherein, The hidden layer comprises an input gate, a forget gate and an output gate, and the hidden feature is extracted in the following manner: determining the feature of each time step contained in the state vector, inputting the feature into the forget gate in the order of time steps to obtain a feature vector after the forget feature; inputting the features in the feature vector after the forget feature into the input gate in the order of time steps to obtain a screened candidate feature vector; inputting the features in the screened candidate feature vector into the output gate in the order of time steps to obtain the hidden feature.

4. The method of claim 2, wherein, Training the parameters of the input layer, the hidden layer, the attention mechanism layer and the output layer through the training sample set to obtain the target model comprises: screening a related feature related to business risk from the features of the business data, constructing a state space based on the related feature, and constructing an action space based on a preset risk management strategy, wherein the risk management strategy comprises at least one of the following: risk warning notification, suspension of business operation and supplementary audit material; training parameters of the input layer, the hidden layer, the attention mechanism layer and the output layer based on the state space, the action space and the training sample set multiple times to obtain a candidate model after each iteration training, wherein each candidate model corresponds to the parameters of the input layer, the hidden layer, the attention mechanism layer and the output layer after one iteration training; determining a target loss function, calculating a loss value of each candidate model based on the target loss function, and determining the candidate model corresponding to the minimum loss value as a target model.

5. The method of claim 4, wherein, training parameters of the input layer, the hidden layer, the attention mechanism layer and the output layer based on the state space, the action space and the training sample set multiple times includes: for each iteration training, inputting a training sample into the candidate model to obtain a predicted risk prediction level; inputting the predicted risk prediction level into a policy network, the policy network determining a risk management strategy to be executed from the action space, inputting the risk management strategy into a value network to obtain a reward value; updating parameters of the candidate model based on the reward value, the risk management strategy to be executed and a state vector of the state space to obtain an iteration trained candidate model, wherein the parameters of the candidate model include parameters of the input layer, the hidden layer, the attention mechanism layer and the output layer.

6. The method of claim 5, wherein, updating parameters of the candidate model based on the reward value, the risk management strategy to be executed and a state vector of the state space includes: calculating a time difference error based on the reward value, the risk management strategy to be executed and the state vector of the state space; inputting the time difference error, the predicted risk prediction level and a historical risk prediction level into the target loss function to obtain a loss value; calculating a gradient of the parameters of the candidate model based on the loss value and the target loss function, calculating a product of the gradient and a preset learning rate, and calculating a difference value between parameters of a candidate model of a previous iteration and the product to obtain parameters of the candidate model of the current iteration.

7. The method of claim 1, wherein, The attention weight is calculated in the following way: determining a preset query vector, wherein the preset query vector is used to represent a business risk indicator; for each hidden feature, calculating a feature similarity between the hidden feature and the preset query vector to obtain a similarity weight; encoding the position of each hidden feature in the time step of the hidden feature vector to obtain a position encoding, and determining a position weight based on the position encoding of each hidden feature; for each hidden feature, calculating the sum of the similarity weight and the position weight to obtain the attention weight of the hidden feature.

8. A business risk management apparatus characterized by comprising: includes: an acquisition unit configured to acquire business data, encode the business data to obtain a business vector, and input the business vector into the input layer of the candidate model. An input unit is configured to input the business vector into a target model to obtain a risk prediction level, wherein the target model comprises an input layer, a hidden layer, an attention mechanism layer, and an output layer, the input layer is configured to extract a state vector from the business vector, the hidden layer is configured to extract a hidden feature from the state vector, the attention mechanism layer is configured to calculate an attention weight and determine a target feature by the hidden feature and the attention weight, the attention weight is determined by a similarity between a business risk indicator and the hidden feature and a position encoding of the hidden feature, and the output layer is configured to output the risk prediction level based on the target feature; A first determination unit is configured to determine a target risk management strategy by the risk prediction level, wherein the target risk management strategy comprises at least one of the following: a risk warning notification, a suspension of business operation, and supplementary audit materials.

9. A computer program product comprising a computer program, characterized in that, The computer program, when executed by a processor, implements the business risk management method of any one of claims 1 to 7.

10. An electronic device, comprising: One or more processors and a memory are included, and the memory is configured to store one or more programs, wherein when the one or more programs are executed by the one or more processors, the one or more processors are caused to implement the business risk management method of any one of claims 1 to 7.

Images