Deep learning-based credit risk assessment method and apparatus, and electronic device

Abstract

The invention discloses a credit risk assessment method and device based on deep learning and electronic equipment. The method comprises the following steps: acquiring target data of a target object; obtaining customer portrait data of the target object according to the target data; determining risk classification models respectively corresponding to the feature information of the plurality of risk dimensions; for the feature information of each risk dimension, performing risk assessment on the target object by adopting a risk classification model corresponding to the feature information of each risk dimension to obtain an initial assessment result, and summarizing the initial assessment results of the plurality of risk dimensions; and determining a final risk assessment result of the target object according to the summarized result. The technical problem that the risk assessment result is not accurate enough due to the fact that customer risk prevention and control are carried out through artificial experience in the prior art is solved.

Description

Technical Field

[0001] The present application relates to the technical field of credit risk assessment, and in particular, to a credit risk assessment method, device and electronic device based on deep learning. Background Art

[0002] Credit risk assessment is the main mechanism for financial institutions to evaluate the credit status of loan applicants. Related technologies use manual experience to prevent and control customer risks, which makes risk assessment susceptible to the subjective judgment of the assessor, and the accuracy of the assessment results is low. In addition, it is difficult for credit risk assessment to conduct in-depth exploration of customer characteristics.

[0003] To address the above-mentioned problems, no effective solution has been proposed yet. Summary of the invention

[0004] The embodiments of the present application provide a credit risk assessment method, device and electronic device based on deep learning, so as to at least solve the technical problem that the related technology uses manual experience to prevent and control customer risks, resulting in inaccurate risk assessment results.

[0005] According to one aspect of an embodiment of the present application, a deep learning-based credit risk assessment method is provided, comprising: obtaining target data of a target object, wherein the target data includes first data obtained from an external device through multiple service interfaces and second data stored by the target object in multiple business systems of a financial institution; obtaining customer portrait data of the target object based on the target data, wherein the customer portrait data is used to reflect characteristic information of the target object in multiple risk dimensions; determining risk classification models corresponding to the characteristic information of multiple risk dimensions, wherein the risk classification models are obtained by training with historical customer portrait data of the financial institution; for the characteristic information of each risk dimension, performing risk assessment on the target object using the risk classification model corresponding to the characteristic information of each risk dimension to obtain an initial assessment result, and summarizing the initial assessment results of multiple risk dimensions; determining a final risk assessment result of the target object based on the summarized result.

[0006] In some embodiments of the present application, customer portrait data of a target object is obtained based on target data, including: generating graph data corresponding to the target data, wherein the nodes in the graph data include feature nodes of the target object and nodes of other objects associated with the target object, and the edges in the graph data include associated edges between feature nodes and associated edges between the target object and the other objects; inputting the graph data into a graph neural network model for feature extraction to obtain feature vectors corresponding to multiple features of the target object; and determining the customer portrait data of the target object corresponding to the feature vector.

[0007] In some embodiments of the present application, graph data corresponding to target data is generated, including: determining an atomic label set corresponding to the target data, wherein each atomic label in the atomic label set corresponds to a class of features of the target object; combining the atomic labels according to the correlation between the atomic labels in the atomic label set to obtain a derived label set; and determining graph data corresponding to the atomic label set and the derived label set, wherein the nodes in the graph data include atomic labels and derived labels.

[0008] In some embodiments of the present application, risk classification models corresponding to characteristic information of multiple risk dimensions are determined, including: determining target characteristic information corresponding to multiple risk dimensions from customer portrait data; for each risk dimension, determining the business type corresponding to the target characteristic information, wherein the business type is used to reflect the application scenario corresponding to risk assessment in the business process of the financial institution; and determining the risk classification model corresponding to the characteristic information of each risk dimension based on the business type.

[0009] In some embodiments of the present application, after obtaining the evaluation results, the method also includes: obtaining positioning information in the portrait results of the target object, wherein the positioning information is used to reflect the virtual position of the target object in the business promotion grid division within the financial institution, wherein the business promotion grid is used to represent the smallest management unit of the service area of ​​the financial institution; determining a target business promotion grid corresponding to the positioning information from multiple business promotion grids, wherein each business promotion grid corresponds to a business promotion strategy; and determining a business promotion plan corresponding to the target object based on the business promotion strategy of the target business promotion grid and the evaluation results.

[0010] In some embodiments of the present application, multiple business promotion grids are determined in the following manner: obtaining the service area corresponding to the financial institution, wherein the service area includes a geographical location area corresponding to the service coverage of the financial institution; determining customer information within the service area from multiple systems of the financial institution; dividing the service area into multiple sub-areas based on the customer information, wherein each sub-area corresponds to a customer life cycle stage; determining at least one business promotion grid corresponding to each sub-area, and a business promotion strategy corresponding to the at least one business promotion grid.

[0011] In some embodiments of the present application, after obtaining the target data of the target object, the method also includes: generating a transaction behavior pattern corresponding to the target object based on the target data before a preset time point, wherein the transaction behavior pattern includes transaction features of multiple dimensions; determining transaction thresholds corresponding to the transaction features of multiple dimensions, respectively, wherein the transaction threshold is used to quantitatively represent the normal range of the transaction behavior of the target object; comparing the target data after the preset time point with the transaction threshold to obtain a comparison result; and generating an alarm message when the comparison result meets a preset condition.

[0012] According to another aspect of the embodiment of the present application, a deep learning-based credit risk assessment device is also provided, including: an acquisition module, used to acquire target data of a target object, wherein the target data includes first data acquired from an external device through multiple service interfaces and second data stored by the target object in multiple business systems of a financial institution; a portrait module, used to acquire customer portrait data of the target object based on the target data, wherein the customer portrait data is used to reflect the characteristic information of the target object in multiple risk dimensions; a determination module, used to determine the risk classification models corresponding to the characteristic information of multiple risk dimensions, wherein the risk classification models are obtained by training the historical customer portrait data of the financial institution; an assessment module, used to perform risk assessment on the target object according to the characteristic information of each risk dimension, using the risk classification model corresponding to the characteristic information of each risk dimension, to obtain an initial assessment result, and summarize the initial assessment results of multiple risk dimensions; a summary module, used to determine the final risk assessment result of the target object based on the summary result.

[0013] According to another aspect of the embodiments of the present application, there is also provided an electronic device, including: a memory and a processor, the memory being used to store program instructions; the processor being connected to the memory and being used to execute and implement the above-mentioned credit risk assessment method based on deep learning.

[0014] According to another aspect of an embodiment of the present application, a non-volatile storage medium is also provided, which includes a stored computer program, wherein the device where the non-volatile storage medium is located executes the above-mentioned credit risk assessment method based on deep learning by running the computer program.

[0015] According to another aspect of the embodiments of the present application, a computer program product is also provided, including computer instructions, which, when executed by a processor, implement the above-mentioned credit risk assessment method based on deep learning.

[0016] In the embodiment of the present application, by integrating external data and internal business system data to construct a customer portrait, and then using a risk classification model to conduct a multi-dimensional risk assessment of the customer, the purpose of accurately locating the customer's credit risk is achieved, thereby achieving the technical effect of improving the accuracy and comprehensiveness of the credit risk assessment of financial institutions, and further solving the technical problem that the related technology uses manual experience to prevent and control customer risks, resulting in inaccurate risk assessment results. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] The drawings described herein are used to provide a further understanding of the present application and constitute a part of the present application. The illustrative embodiments of the present application and their descriptions are used to explain the present application and do not constitute an improper limitation on the present application. In the drawings:

[0018] Figure 1 It is a hardware structure block diagram of a computer terminal of a credit risk assessment method based on deep learning according to an embodiment of the present application;

[0019] Figure 2 is a flow chart of a credit risk assessment method based on deep learning according to an embodiment of the present application;

[0020] Figure 3 It is a structural diagram of a credit risk assessment device based on deep learning according to an embodiment of the present application. DETAILED DESCRIPTION

[0021] In order to enable those skilled in the art to better understand the solution of the present application, the technical solution in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work should fall within the scope of protection of the present application.

[0022] It should be noted that the terms "first", "second", etc. in the specification and claims of the present application and the above-mentioned drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence. It should be understood that the data used in this way can be interchangeable where appropriate, so that the embodiments of the present application described herein can be implemented in an order other than those illustrated or described herein. In addition, the terms "including" and "having" and any of their variations are intended to cover non-exclusive inclusions, for example, a process, method, system, product or device comprising a series of steps or units is not necessarily limited to those steps or units clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0023] In order to better understand the embodiments of the present application, the technical terms involved in the embodiments of the present application are explained as follows:

[0024] Customer Profiling: refers to a detailed description of customer characteristics, behaviors, needs and preferences constructed through the collection and analysis of customer data.

[0025] Risk Classification Model (RCM): In the field of financial risk management, a model established through statistical analysis or machine learning algorithms based on a series of financial and non-financial indicators is used to assess and classify the risks of financial assets.

[0026] Marketing Grid: A business promotion management strategy that divides the market or customer base into multiple grids based on geographic location, industry characteristics, customer attributes, etc. Each grid sets specific business promotion goals and strategies to achieve accurate delivery and efficient management of business promotion resources.

[0027] Credit risk assessment is a long-standing problem faced by financial institutions. It directly affects the asset quality and profitability of financial institutions and is also an issue that financial institutions need to focus on in their credit business. The credit risk assessment method used by relevant technologies is mainly carried out by risk assessment models constructed by credit risk management experts based on their own experience. However, there are the following problems. First, credit risk assessment often only assesses one or several risk dimensions, and cannot assess the overall risk of customers in multiple risk dimensions; second, credit risk assessment is usually based on the customer's corporate financial statements. Due to the limitations of financial statement data, it cannot fully reflect the customer's credit status; and credit risk assessment relies on the experience of credit risk management experts, and it is difficult to avoid the subjective factors of credit risk management experts, which affects the accuracy of credit risk assessment.

[0028] In order to solve the above technical problems, the embodiments of the present application provide corresponding solutions, which are described in detail below.

[0029] The credit risk assessment method embodiment based on deep learning provided in the embodiments of the present application can be executed in a mobile terminal, a computer terminal or a similar computing device. Figure 1 FIG. 1 shows a hardware structure block diagram of a computer terminal for implementing a credit risk assessment method based on deep learning. Figure 1 As shown, the computer terminal 10 may include one or more (102a, 102b, ..., 102n are used to illustrate) processors (the processor may include but is not limited to a processing device such as a microprocessor MCU or a programmable logic device FPGA), a memory 104 for storing data, and a transmission module 106 for communication functions connected via a wired and / or wireless network. In addition, it may also include: a display, a keyboard, a cursor control device, an input / output interface (I / O interface), a universal serial bus (USB) port (which may be included as one of the ports of the I / O interface), a network interface, and a BUS bus. Those skilled in the art will understand that Figure 1 The structure shown is only for illustration and does not limit the structure of the above electronic device. Figure 1 More or fewer components as shown, or with Figure 1 Different configurations shown.

[0030] It should be noted that the one or more processors and / or other data processing circuits described above may generally be referred to herein as "data processing circuits". The data processing circuits may be embodied in whole or in part as software, hardware, firmware, or any other combination thereof. In addition, the data processing circuit may be a single independent processing module, or may be incorporated in whole or in part into any of the other components in the computer terminal 10. As described in the embodiments of the present application, the data processing circuit acts as a processor control (e.g., selection of a variable resistor terminal path connected to an interface).

[0031] The memory 104 can be used to store software programs and modules of application software, such as the program instructions / data storage device corresponding to the credit risk assessment method based on deep learning in the embodiment of the present application. The processor executes various functional applications and data processing by running the software programs and modules stored in the memory 104, that is, realizing the above-mentioned credit risk assessment method based on deep learning. The memory 104 may include a high-speed random access memory, and may also include a non-volatile memory, such as one or more magnetic storage devices, flash memory, or other non-volatile solid-state memory. In some examples, the memory 104 may further include a memory remotely arranged relative to the processor, and these remote memories may be connected to the computer terminal 10 via a network. Examples of the above-mentioned network include, but are not limited to, the Internet, an intranet, a local area network, a mobile communication network, and combinations thereof.

[0032] The transmission module 106 is used to receive or send data via a network. The specific example of the above network may include a wireless network provided by a communication provider of the computer terminal 10. In one example, the transmission module 106 includes a network adapter (Network Interface Controller, NIC), which can be connected to other network devices through a base station so as to communicate with the Internet. In one example, the transmission module 106 can be a radio frequency (RF) module, which is used to communicate with the Internet wirelessly.

[0033] The display may be, for example, a touch screen liquid crystal display (LCD) that enables a user to interact with a user interface of the computer terminal 10 .

[0034] It should be noted that, in some optional embodiments, the above Figure 1 The computer terminal shown may include hardware elements (including circuits), software elements (including computer code stored on a computer-readable medium), or a combination of hardware elements and software elements. It should be noted that Figure 1 This is merely one example of a particular embodiment and is intended to illustrate the types of components that may be present in the computer terminal described above.

[0035] In the above operating environment, an embodiment of the present application provides an embodiment of a credit risk assessment method based on deep learning. It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer executable instructions, and although the logical order is shown in the flowchart, in some cases, the steps shown or described can be executed in an order different from that shown here.

[0036] Figure 2 is a flowchart of a credit risk assessment method based on deep learning according to an embodiment of the present application, such as Figure 2 As shown, the method comprises the following steps:

[0037] Step S202, obtaining target data of the target object, wherein the target data includes first data obtained from an external device through multiple service interfaces and second data of the target object stored in multiple business systems of the financial institution.

[0038] In the above step S202, the target data refers to a data set related to the target object, used to describe the characteristics, behavior and risk status of the target object, and includes internal and external information of the target object, such as information of the target object collected from different data sources (such as credit system, core system, third-party data service, etc.). The first data refers to data obtained from an external device or a third-party service interface. The multiple service interfaces include but are not limited to service interfaces that provide geographic location information, market data, credit information, and industry information, which are used to supplement the deficiencies of internal data of financial institutions, provide a more comprehensive external perspective, and enhance the accuracy of customer portraits. The second data refers to data stored in the internal business system of a financial institution, such as basic information, transaction records, credit history, etc. of customers. For example, the loan history and repayment record of the target customer can be extracted from the credit system, the basic account information and deposit status of the customer can be obtained from the core system, and the financial statement data of the target customer can be collected from the financial system.

[0039] In some embodiments of the present application, after obtaining the target data of the target object, the following process can also be performed: generating a transaction behavior pattern corresponding to the target object based on the target data before a preset time point, wherein the transaction behavior pattern includes transaction features of multiple dimensions; determining transaction thresholds corresponding to the transaction features of multiple dimensions, respectively, wherein the transaction threshold is used to quantitatively represent the normal range of the transaction behavior of the target object; comparing the target data after the preset time point with the transaction threshold to obtain a comparison result; and generating an alarm message when the comparison result meets the preset conditions.

[0040] The preset time point refers to the time cutoff point of the historical data based on which the system generates the transaction behavior pattern and sets the transaction threshold, which is used to divide the historical data and the real-time data, and ensure that the transaction behavior pattern and the transaction threshold are established based on a stable, past data set to accurately reflect the normal transaction behavior of the target object. In some embodiments of the present application, it is possible to analyze whether the transaction behavior of the target object is periodic, such as whether there is a regular transaction peak at a specific time every week, month or year. For example, it can be achieved by a time series analysis method, such as a periodicity detection algorithm, to determine the seasonal fluctuation characteristics of the transaction behavior; at the same time, the stability of the historical data can also be evaluated to ensure that the data set used to establish the transaction behavior pattern and the transaction threshold is reliable. For example, a statistical method (such as calculating the variance, coefficient of variation, etc. of the data set) can be used to evaluate the degree of data fluctuation, or a specific anomaly detection algorithm (such as Isolation Forest) can be used to identify outliers and abnormal transactions in the data set; based on the results of the business cycle and data stability evaluation, the preset time point can be dynamically adjusted to determine an optimal historical data window length. The adjustment of window length can be based on business rules (such as transaction data of the most recent year) or optimized through machine learning algorithms (such as using an automatic model selection algorithm to automatically determine the window size based on the characteristics of the data). For example, for target objects with obvious seasonal fluctuations, the preset time point can be set at the end of the most recent complete business cycle to ensure that the model learns a complete cycle of data.

[0041] For example, assuming that the target object is an account of a company, analysis finds that its transaction behavior has obvious monthly periodicity, and the stability of the data set is good in the past year. Based on this information, the preset time point can be set to the end of the most recent month, and the data of the most recent month can be used as the basis for the transaction behavior pattern and transaction threshold. By monitoring real-time transaction data, if there is a large difference from the transaction behavior predicted by the historical model (such as a significant increase in the transaction amount), the data window can be automatically adjusted through the real-time data impact analysis algorithm, such as selecting the transaction data of the last 6 months to retrain the model to adapt to the changes in the target object's transaction behavior.

[0042] The transaction behavior pattern refers to the transaction characteristics and behavior patterns of the target object in multiple dimensions, which are analyzed based on the historical transaction data of the target object before the preset time point. The transaction characteristics of multiple dimensions refer to different types of characteristics extracted from the transaction data of the target object, such as transaction frequency, transaction amount, transaction time, transaction location, transaction counterparty, etc. In some embodiments of the present application, the transaction data can be sorted into a time series in chronological order, and transaction characteristics of multiple dimensions can be extracted, such as transaction amount, transaction frequency, transaction time (specific to hours or minutes), transaction location code, transaction type (such as transfer, consumption, cash withdrawal), etc. These characteristics constitute multiple variables of time series data. For transaction data, a long short-term memory network (LSTM) can be used to train the pre-processed transaction time series data. During the training process, the model will learn the time series pattern of the normal transaction behavior of the target object, that is, the change law of transaction characteristics in different dimensions over time. After training, the LSTM model can be used to identify the transaction behavior pattern of the target object. That is, given a piece of transaction data, the model can predict the next transaction behavior, or determine whether the current transaction data conforms to the learned pattern. If the value of a certain dimension of the transaction data (such as transaction amount, transaction frequency) is too different from the value predicted by the model, the transaction behavior can be considered abnormal.

[0043] The transaction threshold refers to the quantitative range of normal transaction behavior of the target object defined in multiple dimensions, and transactions beyond this range are considered abnormal. In some embodiments of the present application, based on the transaction behavior pattern of the target object, the system sets the upper and lower limits of the normal transaction range in each dimension, and the transaction threshold can be set by calculating the statistics of historical data (such as the mean plus or minus the multiple of the standard deviation) or using anomaly detection algorithms (such as Isolation Forest, Local Outlier Factor, etc.).

[0044] The transaction data of the target object after the preset time point can be compared with the pre-set transaction threshold, and the difference or distance between the transaction data and the transaction threshold can be calculated in each dimension to determine whether the transaction is abnormal. For example, the transaction data of the target object can be monitored in real time, and the amount, time, location and other dimensions of each new transaction can be compared with the transaction threshold. If the transaction amount exceeds the set threshold, it will be regarded as an abnormal transaction. When the comparison result of the transaction data and the transaction threshold meets the preset conditions (such as the number of times or amplitude exceeding the threshold reaches the pre-set threshold), the system will automatically trigger the alarm process, generate alarm information and send it to the corresponding risk management personnel.

[0045] Step S204, obtaining customer portrait data of the target object based on the target data, wherein the customer portrait data is used to reflect characteristic information of the target object in multiple risk dimensions.

[0046] In the above step S204, the customer profile data is obtained by analyzing the target data to comprehensively reflect the characteristic information of the target object in multiple risk dimensions, including basic information, transaction behavior, credit status, financial status, preference behavior, risk level, etc. Multiple risk dimensions refer to the multiple factors and indicators considered when assessing customer risk, including but not limited to credit risk, market risk, operational risk, liquidity risk, etc.

[0047] In order to capture the complex relationship between the target object and other objects in its social network, and thus construct a customer portrait more accurately, the customer portrait data can be determined in the following way: generate graph data corresponding to the target data, wherein the nodes in the graph data include feature nodes of the target object and nodes of other objects associated with the target object, and the edges in the graph data include associated edges between feature nodes and associated edges between the target object and the other objects; input the graph data into a graph neural network model for feature extraction to obtain feature vectors corresponding to multiple features of the target object; and determine the customer portrait data of the target object corresponding to the feature vector.

[0048] The nodes of other objects are entity nodes related to the target object in the graph data, such as affiliated companies, counterparties, other accounts in the credit record, etc. By analyzing the relationship between the target object and these nodes, the behavior patterns and potential risks of the target object can be revealed, enriching the content and depth of the customer portrait data.

[0049] In some embodiments of the present application, feature information of the target object and relationship information between the target object and other entities can be extracted from the target data to construct nodes and edges. For example, the transaction record of the target object can be converted into a feature node and an associated edge with the transaction counterparty, and the associated account information in the credit record of the target object can be converted into nodes of other objects and associated edges with the target object, that is, the above-mentioned associated edges can represent the business relationship between the target object and its transaction counterparty (such as the transaction counterparty). Taking corporate customers as an example, the financial data of the enterprise itself can be used as a feature node, and the upstream and downstream enterprises in the supply chain can be used as nodes of other objects, and the associated edges between nodes can be established through information such as transaction records and contractual relationships.

[0050] The Graph Neural Network Model (GNN) can capture the local and global features of nodes from the graph structure through a multi-layer information aggregation and propagation mechanism to generate a feature vector of the target object. Among them, the local features focus on the information of the target object and its directly associated nodes (such as the counterparty, affiliated enterprises, etc.). For example, local features can include the attributes of the target object itself, such as transaction frequency, transaction amount, credit score, financial indicators, etc.; the attributes of the nodes directly connected to the target object, such as the credit score of the counterparty, the financial status of the affiliated enterprise, etc.; the strength of the relationship between the target object and its directly associated nodes, such as the size of the transaction amount, the level of equity ratio, etc. The global features focus on the position and role of the target object in the entire financial network graph, as well as the macro-structure information of the entire graph. For example, global features can include the connectivity of the target object in the entire network, such as how many other entities it is directly or indirectly associated with; the network community structure to which the target object belongs, that is, its "social circle" or "business circle" in the graph; statistical information of the entire graph, such as the average transaction amount, average financial indicators, etc., which reflect the general risk level of the entire market or industry.

[0051] Local and global features are generated through multiple layers of message passing and aggregation operations. In each layer of propagation, a node receives information (messages) from its direct neighbor nodes, and then fuses this information with the node's own features through aggregation operations (such as averaging, summing, maximum, etc.) to generate a new feature representation. As the depth of the network increases, the information received by the node will come from more distant neighbor nodes, and thus gradually merge into the global features. For example, the first layer of propagation may only capture the information of the direct transaction counterparty, while by the third layer of propagation, the relationship between the target object and the second and third level affiliated enterprises, as well as the additional information brought by these enterprises themselves and the associations between them, may have been taken into account.

[0052] Specifically, first, local feature vectors are initialized for each node in the graph. These vectors can be based on the initial attributes of the node (such as age, gender) or pre-trained word vectors (if the node attributes are text); in each layer of propagation, the node receives information from its direct neighbors, which can be feature vectors of neighbor nodes, the strength of the relationship between the neighbor nodes and itself (such as transaction amount, social connection strength), etc.; after receiving the message, the node will combine its own features and the messages of the neighbor nodes to update the features, for example, through a defined message passing function (such as linear transformation, nonlinear activation function) to reflect the interaction between the node and its neighbors; after the node features are updated, an aggregation operation can be performed to fuse Local features, such as average, sum, maximum, and attention mechanisms, are used to integrate the information of neighboring nodes and generate feature representations of new nodes in this layer. After aggregation, the feature representation of the node not only contains information about itself and its direct neighbors, but also incorporates information from more distant nodes, gradually forming more comprehensive local features. After multi-layer propagation and aggregation operations, each node will obtain a final representation that combines multiple layers of local features. In order to generate global features, global aggregation operations can be performed on the feature representations of nodes in the entire graph, such as global average, global sum, global maximum, etc., to fuse the feature representations of all nodes into a feature vector representing the entire graph structure, namely, the global feature.

[0053] Taking enterprise customer graph data as an example, the GCN (Graph Convolutional Network) model can be used to integrate the financial data characteristics of the enterprise, the characteristics of upstream and downstream enterprises in the supply chain, and the relationship characteristics between them through graph convolution operations, and finally generate feature vectors of enterprise customers in dimensions such as credit risk and market risk.

[0054] In order to more carefully characterize the characteristics of the target object and improve the depth and breadth of risk assessment, the graph data corresponding to the target data can be generated through the following steps. Specifically: determine the atomic label set corresponding to the target data, wherein each atomic label in the atomic label set corresponds to a type of feature of the target object; combine the atomic labels according to the correlation between the atomic labels in the atomic label set to obtain a derived label set; determine the graph data corresponding to the atomic label set and the derived label set, wherein the nodes in the graph data include atomic labels and derived labels.

[0055] Atomic tags are the smallest unit tags used to describe the basic attributes of the target object. Each atomic tag corresponds to a type of feature of the target object. For example, atomic tags such as "gender", "age", and "occupation" can be extracted from the basic information of customers, and atomic tags such as "transaction amount", "transaction frequency", and "transaction time" can be extracted from transaction records. The atomic tag set can include not only tags that directly describe the characteristics of the target object, but also feature information of other objects related to the target object extracted from the target data. This information can also be converted into nodes in the graph data to more comprehensively reflect the position, relationship, and dynamics of the target object in the financial network, thereby generating richer and more accurate graph data.

[0056] In some embodiments of the present application, information about the counterparty, associated companies, and third-party guarantors related to the loan customer can be extracted from the target data, and this information can be converted into atomic tags, and these atomic tags can be converted into nodes in the graph data, and then the associated edges between the target object and these nodes can be established to represent the business relationship between them. For example, for a target object, the atomic tag set can include atomic tags that directly describe the characteristics of the loan customer, such as "age", "gender", "income level", "occupation", "credit score", etc., and can also include atomic tags that describe the characteristics of other objects that have a direct business relationship with the loan customer, such as "counterparty credit score", which is used to represent the credit score of other individuals or companies that have a transaction relationship with the loan customer, "associated company financial status", which is used to represent the financial health of companies that have equity or other relationships with the loan customer, and "third-party guarantor credit history", which is used to represent the credit record of a third-party individual who provides a guarantee for the loan customer.

[0057] A derived label is a composite label obtained by combining multiple atomic labels by analyzing the correlation between labels in the atomic label set. It can reflect the comprehensive characteristics of the target object in different dimensions. For example, the "age" atomic label can be combined with the "income level" atomic label to generate a derived label of "high-income young group"; the "transaction amount" atomic label can be combined with the "transaction frequency" atomic label to generate a derived label of "frequent large transactions".

[0058] In some embodiments of the present application, the correlation between atomic tags in an atomic tag set can be determined by the following steps: converting non-numeric atomic tags into numerical representations; selecting features related to risk assessment of the target object or establishment of a customer profile from the atomic tag set, and performing cluster analysis on the standardized atomic tag data using a clustering algorithm (such as K-means, DBSCAN, hierarchical clustering, etc.) to identify a sample set with similar features; based on the clustering analysis results, identifying atomic tag combinations with similar features. For example, if a group of samples are found to have similar features on atomic tags such as "transaction frequency", "transaction amount" and "customer age" in the clustering analysis, these atomic tags can be combined into a derived tag to reflect the transaction behavior pattern and age characteristics of the target object; based on the atomic tag combinations identified in the clustering results, a derived tag set is constructed.

[0059] After obtaining the atomic labels and derived labels, the atomic label sets and derived label sets can be converted into nodes in the graph data. In addition, other entities related to the target object (which can also be regarded as part of the atomic labels), such as counterparties, affiliated companies, co-borrowers, etc., should also be added to the graph as nodes, and then the association edges between the nodes are constructed. The association edges represent the relationship between the nodes, including the association between feature nodes and the association between feature nodes and entity nodes. For example:

[0060] (1) Associative edges between feature nodes: Based on the association between atomic labels identified by statistical methods or machine learning algorithms, and the combinatorial logic of derived labels, associative edges between feature nodes are constructed. For example, if "age" and "income level" derive the "age group income" label, then in the graph, associative edges can be established between the "age" node, the "income level" node, and the "age group income" node to represent the logical derivative relationship between them.

[0061] (2) Associative edges between entity nodes and feature nodes: For each entity node, an associative edge is constructed between the entity node and the feature node based on the characteristics of the entity in the target data. For example, for the target loan customer, feature nodes such as "customer age" and "customer income" can be constructed, and associative edges can be established with the "loan customer" entity node to represent the specific values ​​of the loan customer on these characteristics.

[0062] At the same time, you can also define the association edges between the target object and related entities, and you can build association edges between the target object and other entities in the graph based on business data, such as transaction records, equity relationships, etc. For example, for the transaction relationship between the target loan customer and the counterparty, you can establish a "transaction" association edge between the loan customer node and the counterparty node, and the weight of the edge can represent the amount or frequency of the transaction.

[0063] In some embodiments of the present application, after the preliminary graph data is constructed, the graph data can also be integrated and optimized to ensure the quality and structural rationality of the graph data. Specifically, the graph data from different sources can be integrated to ensure the integrity and consistency of the graph data. For example, if there are multiple data sources for a certain feature of the target object, data fusion is required to determine the final feature value. The graph data can also be optimized, including removing redundant edges, adjusting the weights of nodes and edges, processing isolated nodes, etc., to improve the analysis efficiency and accuracy of the graph data.

[0064] Step S206, determining the risk classification models corresponding to the characteristic information of multiple risk dimensions, wherein the risk classification models are obtained by training the historical customer portrait data of the financial institution.

[0065] In the above step S206, the risk dimensions may include any of credit risk, operational risk, market risk, liquidity risk, reputation risk, etc., which can be determined according to the specific needs of the financial institution. Each risk dimension is associated with a set of specific feature information in the customer portrait data, which is used to evaluate the risk level under this dimension.

[0066] The characteristic information of multiple risk dimensions refers to specific indicators used to describe the customer's status or behavior. These indicators can be atomic tags or derived tags, covering multiple aspects such as the customer's personal attributes, transaction behavior, financial status, credit record, etc.

[0067] The risk classification model is used to classify customers into different risk levels. The model is trained based on the historical customer profile data of financial institutions. By learning from the risk performance of past customers, the risk level of new customers is predicted. For example, it can include logistic regression, support vector machine, random forest, deep learning model, etc.

[0068] For different business types, in order to adopt a more appropriate model for risk assessment to improve the pertinence and effectiveness of the assessment, in some embodiments of the present application, the risk classification models corresponding to the characteristic information of multiple risk dimensions can be determined in the following ways. Specifically: determine the target characteristic information corresponding to the multiple risk dimensions from the customer portrait data; for each risk dimension, determine the business type corresponding to the target characteristic information, wherein the business type is used to reflect the application scenario of risk assessment in the business process of the financial institution; determine the risk classification model corresponding to the characteristic information of each risk dimension based on the business type.

[0069] Target feature information refers to key indicators or attributes in customer portrait data that are directly related to a specific risk dimension, and is used to reflect the risk characteristics under that risk dimension. In some embodiments of the present application, features associated with a specific risk dimension can be extracted from customer portrait data through feature engineering. For example, for the credit risk dimension, features such as customer age, income level, credit score, and historical repayment record may be involved. On this basis, risk factor analysis can also be performed based on historical customer portrait data to identify key features that have a significant impact on risk. For example, statistical methods such as factor analysis and principal component analysis (PCA) can be used, or machine learning algorithms such as decision trees and random forests can be used to analyze customer portrait data to determine the degree of influence of feature information on the risk dimension, thereby identifying target feature information.

[0070] Business type refers to the specific application scenarios corresponding to risk assessment in the business process of a financial institution, such as loan approval, account monitoring, transaction review, etc. Different business types have their own specific risk assessment requirements, so they need to match the corresponding risk classification models. In some embodiments of the present application, machine learning models (such as decision trees, logistic regression, neural networks) can be used to predict the relationship between feature information and business types. Model training is based on historical data of financial institutions, which contains business type tags and customer feature information. The trained model can predict the business types that may be involved based on customer feature information, thereby determining in which business scenarios to perform risk assessment.

[0071] Step S208, for the characteristic information of each risk dimension, use the risk classification model corresponding to the characteristic information of each risk dimension to perform risk assessment on the target object, obtain an initial assessment result, and summarize the initial assessment results of multiple risk dimensions.

[0072] In the above step S208, the initial assessment result is a risk level prediction obtained by using the corresponding risk classification model for the characteristic information of each risk dimension. By predicting the risk level under each risk dimension through the risk classification model, a comprehensive risk assessment can be performed on the target object, and finally a set of initial assessment results containing multiple risk dimensions is formed.

[0073] Step S210, determining the final risk assessment result of the target object based on the summary result.

[0074] In the above step S210, the final risk assessment result is the overall risk level or score of the target object determined through comprehensive analysis and decision-making based on the summary results. For example, the final risk assessment result can be determined using a weighted average method, or a risk matrix can be constructed based on business strategies and risk management objectives. The matrix maps the score intervals of different risk dimensions to risk levels (such as low, medium, and high), maps the initial assessment results of the target object in each risk dimension to the risk matrix, and determines the final risk assessment result based on decision rules (such as cross-referencing risk levels of multiple dimensions).

[0075] By combining the risk assessment results with the business promotion grid, accurate business promotion can be achieved, and the business promotion effect and customer satisfaction can be improved. Specifically: the positioning information in the portrait result of the target object is obtained, wherein the positioning information is used to reflect the virtual position of the target object in the business promotion grid division within the financial institution, wherein the business promotion grid is used to represent the smallest management unit of the service area of ​​the financial institution; the target business promotion grid corresponding to the positioning information is determined from multiple business promotion grids, wherein each business promotion grid corresponds to a business promotion strategy; based on the business promotion strategy and evaluation results of the target business promotion grid, the business promotion plan corresponding to the target object is determined.

[0076] The business promotion grid is the smallest management unit of a financial institution's service area. It subdivides the service area into multiple grids. Each grid can be divided according to geographical location, customer characteristics, industry characteristics, etc., in order to implement targeted business promotion strategies.

[0077] Positioning information refers to the virtual location information of a customer in the business promotion grid of a financial institution, which can be determined based on the customer's basic information, transaction records, geographic location and other data, and is used to identify the customer's location in the business promotion grid, thereby matching the corresponding business promotion strategy. In some embodiments of the present application, the geographic location information of the target object can be used in combination with the geographic location boundary of the business promotion grid to determine the business promotion grid to which the target object belongs.

[0078] In some embodiments of the present application, a business promotion strategy can be formulated based on the customer life cycle stage and geographic location information, specifically: obtaining the service area corresponding to the financial institution, wherein the service area includes the geographic location area corresponding to the service coverage of the financial institution; determining customer information within the service area from multiple systems of the financial institution; dividing the service area into multiple sub-areas based on the customer information, wherein each sub-area corresponds to a customer life cycle stage; determining at least one business promotion grid corresponding to each sub-area, and a business promotion strategy corresponding to the at least one business promotion grid.

[0079] The customer life cycle stage refers to the development stage of the relationship between the customer and the financial institution, which may include potential customers, new customers, active customers, mature customers, declining customers or lost customers, etc. In some embodiments of the present application, customer segmentation analysis can be performed based on customer information, and customers in the service area can be grouped according to the life cycle stage. For example, using machine learning algorithms such as cluster analysis or decision trees, customers are divided into different groups based on customer characteristics (such as age, income, transaction frequency) and geographic location information, and each group corresponds to a sub-area. In addition, the service area can also be geographically analyzed in combination with the customer's geographic location information, and divided into multiple sub-areas, each of which represents a customer group with similar life cycle stages and geographic characteristics.

[0080] Through the above steps S202 to S210, by integrating external data and internal business system data, a customer portrait is constructed, and then a risk classification model is used to conduct a multi-dimensional risk assessment of the customer, thereby achieving the purpose of accurately locating the customer's credit risk, thereby achieving the technical effect of improving the accuracy and comprehensiveness of the credit risk assessment of financial institutions, and further solving the technical problem that the related technology uses manual experience to prevent and control customer risks, resulting in inaccurate risk assessment results.

[0081] Figure 3 is a structural diagram of a credit risk assessment device based on deep learning according to an embodiment of the present application, such as Figure 3 As shown, the device comprises:

[0082] An acquisition module 302 is used to acquire target data of a target object, wherein the target data includes first data acquired from an external device through multiple service interfaces and second data stored by the target object in multiple business systems of a financial institution;

[0083] A portrait module 304 is used to obtain customer portrait data of the target object based on the target data, wherein the customer portrait data is used to reflect characteristic information of the target object in multiple risk dimensions;

[0084] A determination module 306 is used to determine risk classification models corresponding to characteristic information of multiple risk dimensions, wherein the risk classification models are obtained by training with historical customer profile data of financial institutions;

[0085] An evaluation module 308 is used to perform risk evaluation on the target object according to the characteristic information of each risk dimension by using a risk classification model corresponding to the characteristic information of each risk dimension, obtain an initial evaluation result, and summarize the initial evaluation results of multiple risk dimensions;

[0086] The summarization module 310 is used to determine the final risk assessment result of the target object according to the summary result.

[0087] It should be noted that Figure 3 The credit risk assessment device based on deep learning is used to perform Figure 2 The credit risk assessment method based on deep learning is shown, so Figure 2 The relevant explanations in the deep learning-based credit risk assessment method in Figure 3 The deep learning-based credit risk assessment device shown will not be described in detail here.

[0088] An embodiment of the present application also provides an electronic device, which includes a memory and a processor, wherein the memory is used to store program instructions; the processor is connected to the memory and is used to execute the steps of implementing the deep learning-based credit risk assessment method in each embodiment of the present application.

[0089] For example, the processor performs the following functions by executing program instructions stored in the memory:

[0090] Obtain target data of a target object, wherein the target data includes first data obtained from an external device through multiple service interfaces and second data stored by the target object in multiple business systems of a financial institution; obtain customer portrait data of the target object based on the target data, wherein the customer portrait data is used to reflect characteristic information of the target object in multiple risk dimensions; determine risk classification models corresponding to the characteristic information of multiple risk dimensions, wherein the risk classification models are obtained by training with historical customer portrait data of the financial institution; for the characteristic information of each risk dimension, use the risk classification model corresponding to the characteristic information of each risk dimension to perform risk assessment on the target object, obtain initial assessment results, and summarize the initial assessment results of multiple risk dimensions; determine a final risk assessment result of the target object based on the summarized results.

[0091] An embodiment of the present application also provides a non-volatile storage medium, which includes a stored computer program, wherein the device where the non-volatile storage medium is located executes the steps of the deep learning-based credit risk assessment method in each embodiment of the present application by running the computer program.

[0092] An embodiment of the present application also provides a computer program product, including computer instructions, which, when executed by a processor, implement the steps of the deep learning-based credit risk assessment method in each embodiment of the present application.

[0093] The embodiments of the present application also provide a computer program, which, when executed by a processor, implements the steps of the deep learning-based credit risk assessment method in each embodiment of the present application.

[0094] The serial numbers of the above-mentioned embodiments of the present application are for description only and do not represent the advantages or disadvantages of the embodiments.

[0095] In the above embodiments of the present application, the description of each embodiment has its own emphasis. For parts that are not described in detail in a certain embodiment, please refer to the relevant description of other embodiments.

[0096] In the several embodiments provided in this application, it should be understood that the disclosed technical content can be implemented in other ways. Among them, the device embodiments described above are only schematic. For example, the division of the units can be a logical function division. There may be other division methods in actual implementation. For example, multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some interfaces, indirect coupling or communication connection of units or modules, which can be electrical or other forms.

[0097] The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed on multiple units. Some or all of the units may be selected according to actual needs to achieve the purpose of the present embodiment.

[0098] In addition, each functional unit in each embodiment of the present application may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit. The above-mentioned integrated unit may be implemented in the form of hardware or in the form of software functional units.

[0099] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present application, or the part that contributes to the prior art, or all or part of the technical solution can be embodied in the form of a software product. The computer software product is stored in a storage medium, including several instructions for a computer device (which can be a personal computer, server or network device, etc.) to execute all or part of the steps of the method described in each embodiment of the present application. The aforementioned storage medium includes: U disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), mobile hard disk, disk or optical disk, etc., which can store program code.

[0100] The above is only a preferred implementation of the present application. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles of the present application. These improvements and modifications should also be regarded as the scope of protection of the present application.

Claims

1. A credit risk assessment method based on deep learning, characterized in that: include: Acquire target data of a target object, wherein the target data includes first data acquired from an external device through a plurality of service interfaces and second data of the target object stored in a plurality of business systems of a financial institution; Acquire customer portrait data of the target object according to the target data, wherein the customer portrait data is used to reflect characteristic information of the target object in multiple risk dimensions; Determine risk classification models corresponding to the characteristic information of the multiple risk dimensions, wherein the risk classification models are obtained by training the historical customer profile data of the financial institution; For each risk dimension, a risk classification model corresponding to the risk dimension is used to perform risk assessment on the target object to obtain an initial assessment result, and the initial assessment results of the multiple risk dimensions are summarized; The final risk assessment result of the target object is determined based on the summary results.

2. The method according to claim 1, characterized in that Acquiring customer portrait data of the target object according to the target data includes: Generate graph data corresponding to the target data, wherein the nodes in the graph data include characteristic nodes of the target object and nodes of other objects associated with the target object, and the edges in the graph data include associated edges between the characteristic nodes and associated edges between the target object and the other objects; Inputting the graph data into a graph neural network model for feature extraction to obtain feature vectors corresponding to multiple features of the target object; Determine customer portrait data of the target object corresponding to the feature vector.

3. The method according to claim 2, characterized in that Generating atlas data corresponding to the target data, including: Determine an atomic tag set corresponding to the target data, wherein each atomic tag in the atomic tag set corresponds to a type of feature of the target object; According to the correlation between the atomic labels in the atomic label set, the atomic labels are combined to obtain a derived label set; The graph data corresponding to the atomic label set and the derived label set is determined, wherein nodes in the graph data include atomic labels and derived labels.

4. The method according to claim 1, characterized in that: Determining the risk classification models corresponding to the characteristic information of the multiple risk dimensions respectively includes: Determine multiple target feature information corresponding to multiple risk dimensions respectively from the customer portrait data; For each risk dimension, determining business types corresponding to the plurality of target feature information, wherein the business types are used to reflect application scenarios corresponding to risk assessment in the business process of the financial institution; A risk classification model corresponding to the characteristic information of each risk dimension is determined according to the business type.

5. The method according to claim 1, characterized in that After obtaining the evaluation result, the method further includes: Acquire positioning information in the portrait result of the target object, wherein the positioning information is used to reflect the virtual position of the target object in the business promotion grid division within the financial institution, wherein the business promotion grid is used to represent the smallest management unit of the service area of ​​the financial institution; Determining a target business promotion grid corresponding to the positioning information from a plurality of business promotion grids, wherein each business promotion grid corresponds to a business promotion strategy; A business promotion plan corresponding to the target object is determined according to the business promotion strategy of the target business promotion grid and the evaluation result.

6. The method according to claim 5, characterized in that The multiple business promotion grids are determined in the following manner: Acquire a service area corresponding to the financial institution, wherein the service area includes a geographical location area corresponding to the service coverage of the financial institution; determining customer information within the service area from a plurality of systems of the financial institution; Dividing the service area into a plurality of sub-areas according to the customer information, wherein each sub-area corresponds to a customer life cycle stage; At least one business promotion grid corresponding to each sub-region and a business promotion strategy corresponding to the at least one business promotion grid are determined.

7. The method according to claim 1, characterized in that After acquiring the target data of the target object, the method further includes: Generate a transaction behavior pattern corresponding to the target object according to the target data before a preset time point, wherein the transaction behavior pattern includes transaction features in multiple dimensions; Determining transaction thresholds corresponding to the transaction features of the multiple dimensions, respectively, wherein the transaction thresholds are used to quantitatively represent a normal range of the transaction behavior of the target object; Comparing the target data after the preset time point with the transaction threshold to obtain a comparison result; When the comparison result meets the preset condition, an alarm message is generated.

8. A credit risk assessment device based on deep learning, characterized in that: include: An acquisition module, used for acquiring target data of a target object, wherein the target data includes first data acquired from an external device through a plurality of service interfaces and second data of the target object stored in a plurality of business systems of a financial institution; A portrait module, used to obtain customer portrait data of the target object based on the target data, wherein the customer portrait data is used to reflect characteristic information of the target object in multiple risk dimensions; A determination module, used to determine the risk classification models corresponding to the characteristic information of the multiple risk dimensions, wherein the risk classification models are obtained by training the historical customer profile data of the financial institution; An assessment module, configured to perform risk assessment on the target object according to the feature information of each risk dimension by using a risk classification model corresponding to the feature information of each risk dimension, obtain an initial assessment result, and summarize the initial assessment results of the multiple risk dimensions; The summary module is used to determine the final risk assessment result of the target object based on the summary result.

9. An electronic device, characterized in that: include: A memory and a processor, wherein the memory is used to store program instructions; the processor is connected to the memory and is used to execute the deep learning-based credit risk assessment method described in any one of claims 1 to 7.

10. A non-volatile storage medium, characterized in that: The non-volatile storage medium includes a stored computer program, wherein the device where the non-volatile storage medium is located executes the deep learning-based credit risk assessment method described in any one of claims 1 to 7 by running the computer program.

11. A computer program product comprising computer instructions, characterized in that: When the computer instructions are executed by a processor, the deep learning-based credit risk assessment method described in any one of claims 1 to 7 is implemented.

Images